A method for predicting the quality of low-temperature preserved fish meat

Abstract

The application discloses a kind of prediction methods for the quality of low-temperature fresh-keeping of segmented fish meat, belong to aquatic product quality management technical field.The method includes: collecting the quality data of segmented fish meat under constant temperature and fluctuating temperature storage conditions, determine the total number of colonies, volatile base nitrogen content as core prediction index;Microbial growth kinetics model and volatile base nitrogen change kinetics model are constructed;Optimized BPNN prediction model and optimized RBFNN prediction model are constructed by fusing physical constraints;A stacked ensemble learning framework is constructed to input the prediction results of each base learner into the meta-learner and output the final prediction value of the quality of low-temperature fresh-keeping of segmented fish meat.The application can accurately adapt to the cold chain fluctuating temperature scene, realize real-time monitoring and intelligent early warning of quality, effectively overcome the problems of traditional detection lag, inaccurate prediction and untimely warning, and provide efficient, stable and reliable technical support for intelligent quality control of segmented fish meat during low-temperature storage and transportation.

Description

Technical Field

[0001] This invention relates to a method for predicting the quality of cut fish meat during low-temperature preservation, belonging to the field of aquatic product quality management technology. Background Technology

[0002] Animal-based aquatic products occupy an important position in the food consumption market due to their tender meat and rich nutrition. Fish meat, as the mainstream processing method, is susceptible to quality deterioration due to its high moisture and protein content during cold chain transportation, storage, and sales. This directly shortens shelf life and reduces distribution efficiency, becoming a key bottleneck in the development of the aquatic product industry.

[0003] Current methods for monitoring and predicting the quality of segmented fish meat primarily rely on sensory evaluation, routine physicochemical testing, and single kinetic models. Sensory evaluation is highly subjective and lacks stability; physicochemical testing is cumbersome and has significant time lag, making real-time online monitoring impossible; single kinetic models can only describe quality changes at specific temperatures, making it difficult to adapt to the actual fluctuating temperatures in cold chain environments, thus limiting prediction accuracy. Furthermore, existing data-driven models are often used independently, failing to fully integrate mechanistic knowledge, resulting in weak generalization ability and large errors, which cannot meet the demands of modern cold chain logistics for accurate quality prediction and intelligent early warning systems.

[0004] In summary, existing technologies are insufficient to achieve dynamic quantitative assessment and timely early warning of the quality of segmented fish meat throughout the entire storage and transportation process. There is an urgent need for a high-precision method for predicting the quality of segmented fish meat at low temperatures, which is adaptable to variable temperature environments, integrates the advantages of multiple models, and addresses industry pain points such as lagging quality monitoring, inaccurate predictions, and untimely early warnings, thereby providing technical support for intelligent management and control of cold chain logistics. Summary of the Invention

[0005] To address the problems of delayed, inaccurate, and untimely monitoring and early warning of the low-temperature quality of cut fish, this invention provides a method for predicting the low-temperature preservation quality of cut fish. The technical solution includes the following steps: (1) Collect quality data of cut fish meat under constant temperature and fluctuating temperature storage conditions, and determine the total number of colonies and volatile basic nitrogen content as the core predictive indicators. (2) Construct a kinetic model of microbial growth and a kinetic model of changes in volatile basic nitrogen; (3) Construct an optimized BPNN prediction model and an optimized RBFNN prediction model that integrate physical constraints; (4) Construct stacked ensemble learning frameworks with kinetic model, optimized BPNN model, and optimized RBFNN model as base learners and CART tree as meta learner, respectively, to predict the total number of colonies and the content of volatile basic nitrogen; (5) Input the prediction results of each base learner into the meta learner and fuse the output to obtain the final predicted value of the low-temperature preservation quality of segmented fish meat, and set the warning threshold and shelf life prediction based on the prediction results.

[0006] Optionally, the construction of the microbial growth kinetics model includes: Constructing a primary model of microbial growth:

[0007] in, t Storage time; N ( t )for t The number of colonies at any given time; N 0 represents the initial colony count; N max The maximum number of colonies; μ max This represents the maximum specific growth rate. λ The lag period; Constructing a two-dimensional model of microbial growth:

[0008]

[0009] in, , These are the equation coefficients; T Storage temperature (°C); T min This indicates the minimum growth temperature threshold.

[0010] Optionally, the microbial growth kinetics model is further decomposed into multiple isothermal processes, represented as follows: when t =d t At 1 o'clock,

[0011] when t =d t 1+d t At 2 o'clock,

[0012] when t =d t 1+d t 2+……+ d t i hour,

[0013] Where, d t iThis represents a short time interval assuming a constant temperature. N ( t i () indicates the number of colonies; This indicates the maximum specific growth rate.

[0014] Optionally, the kinetic model for the change in volatile basic nitrogen is expressed as follows: when t =d t At 1 o'clock,

[0015] when t = d t 1+d t At 2 o'clock,

[0016] when t =d t 1+d t 2+……+ d t i hour

[0017] in, B ( t i () represents the volatile basic nitrogen content; K is the rate of change of volatile basic nitrogen. t Storage time; The rate of change of volatile basic nitrogen is expressed as follows:

[0018] in, k 0 is the frequency factor, and R is the gas constant. E a It is the activation energy.

[0019] Optionally, the shelf life prediction expression is as follows:

[0020] In the formula: Shelf life; B These are the limits for volatile basic nitrogen at the end of the shelf life; B 0 represents the initial volatile basic nitrogen content.

[0021] Optionally, both the optimized BPNN model and the optimized RBFNN model are trained using a fusion loss function, which is: L total = α·L data + β·L physics+ γ·L reg Where L data Let L be the data fitting loss term, where L is the loss term. physics For the physical constraint loss term, L reg α is the regularization term, and β and γ are the weight parameters.

[0022] Optionally, step (5) includes: establishing a quality early warning model for segmented fish meat, wherein the early warning model is obtained by differentiating the temperature dynamics model of volatile basic nitrogen content.

[0023] Optionally, the optimized BPNN model employs Adam adaptive learning rate optimization during the training process, and determines the optimal number of hidden layer neurons through Bayesian optimization.

[0024] Optionally, the optimized RBFNN model uses K-means clustering to determine the radial basis function centers and adjusts the radial basis function spread parameter through Bayesian optimization.

[0025] The beneficial effects of this invention are: This invention employs a deep coupling of a kinetic mechanism model and an optimized neural network, along with stacked ensemble learning for collaborative decision-making. It constructs quality change mechanism constraints using a microbial growth kinetic model and a volatile basic nitrogen change kinetic model, providing a theoretical basis for prediction and avoiding overfitting due to the lack of mechanistic support in the data model. Furthermore, it enhances nonlinear fitting and anti-interference capabilities by introducing BPNN and RBFNN models with fusion loss functions, adaptive learning rates, and Bayesian optimization, compensating for the insufficient adaptability of mechanistic models under complex temperature variations. Finally, a stacked ensemble learning framework is used to collaboratively fuse the prediction results of the two models, with a CART tree-element learner completing the optimal weight allocation and decision output, fully exploring the complementary information between different models and effectively improving prediction accuracy.

[0026] This invention can accurately adapt to fluctuating temperature scenarios in the cold chain, significantly improve prediction accuracy, realize real-time quality monitoring and intelligent early warning, effectively overcome the problems of traditional detection lag, inaccurate prediction, and untimely early warning, and provide efficient, stable and reliable technical support for intelligent quality control of the entire process of low-temperature storage and transportation of cut fish meat. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1This is a microbial fitting curve of the Belehradek equation in Embodiment 2 of the present invention.

[0029] Figure 2 This is a BPNN fitting curve diagram with the total number of colonies as the indicator in Embodiment 2 of the present invention.

[0030] Figure 3 This is the RBFNN fitting curve diagram with the total number of colonies as the indicator in Embodiment 2 of the present invention.

[0031] Figure 4 This is a BPNN fitting curve diagram using volatile basic nitrogen as an indicator in Example 2 of the present invention.

[0032] Figure 5 This is the RBFNN fitting curve diagram with volatile basic nitrogen content as the indicator in Example 2 of the present invention.

[0033] Figure 6 This is a structural diagram of the fish meat segmentation and low-temperature preservation quality prediction model of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0035] Example 1: This embodiment simulates temperature fluctuations that may occur during actual cold chain transportation by designing a fluctuating temperature experiment. The specific temperatures were set as follows: 4℃ for 2 days → 9℃ for 2 days (sampling point a) → 14℃ for 1 day → 9℃ for 2 days → 5℃ for 2 days (sampling point b). During the fluctuating temperature storage experiment, the total bacterial count and volatile basic nitrogen content were measured at regular intervals.

[0036] A modified Gompertz model was used to fit the data, obtaining the relationship between total colony count and storage time at different temperatures, i.e., a first-order model of microbial growth. Data processing and nonlinear regression were performed on the modified Gompertz model to obtain the corresponding fitting curves and growth kinetic parameters, including the maximum specific growth rate (…). μ max ), delay period ( λ ) and maximum colony count ( N max The modified Gompertz model equations are expressed as follows:

[0037] In this formula: t For time (d); N ( t )for tColony count at time (lg(CFU / g)); N 0 represents the initial colony count (lg(CFU / g)); N max The maximum colony count (lg(CFU / g)); μ max The maximum specific growth rate (d –1 ); λ The delay period is d.

[0038] The Belehradek equation was used as a second-order model to characterize the quantitative relationship between ambient temperature and the maximum specific growth rate and growth lag time of microorganisms, expressed as:

[0039]

[0040] In the formula, , These are the equation coefficients; T Storage temperature (°C); T min This indicates the minimum growth temperature threshold.

[0041] The modified Gompertz model equations are further decomposed into multiple isothermal processes, as follows: when t =d t At 1 o'clock,

[0042] when t =d t 1+d t At 2 o'clock,

[0043] when t =d t 1+d t 2+……+ d t i hour

[0044] In the formula: d t i (i=1,2,3…) represents a short time interval (d) assuming constant temperature; N ( t i )for( t =d t 1+d t 2+……+d t iThe colony count at time (lg(CFU / g)); the effect of temperature on colony growth can be represented by the square root model, i.e. ; ; d t i-1 The specific growth rate at that time i- Indicates the first i The isothermal stage lag period (d). t i Indicates the first i The duration of each isothermal phase N i-1 d t i-1 The number of colonies at that time.

[0045] Based on storage experiments at different temperatures, the Arrhenius equation was used to analyze the variation of volatile basic nitrogen content, expressed as:

[0046] In the formula: k 0 represents the frequency factor; T R is the absolute temperature (K); R is the gas constant, 8.3144 J / (mol·K); E a , where is the activation energy (J / mol).

[0047] The equation for predicting shelf life is expressed as:

[0048] In the formula: SL 挥发性盐基氮 Shelf life (d); B The limit value (mg / 100g) for volatile basic nitrogen at the end of the shelf life. B 0 represents the initial volatile basic nitrogen content (mg / 100g).

[0049] Based on the characteristics of unsteady temperature and referring to the construction principles of microbial dynamic growth models, a temperature-dependent kinetic model for volatile basic nitrogen content is established: when t =d t At 1 o'clock,

[0050] when t = d t 1+d t At 2 o'clock,

[0051] whent =d t 1+d t 2+……+ d t i hour

[0052] In the formula: d t i (i=1,2,3…) represents a short time interval (d) assuming constant temperature; B ( t i )for( t =d t 1+d t 2+……+d t i The volatile basic nitrogen content (mg / 100g) at that time; K is the rate constant of change of volatile basic nitrogen. t Storage time (d); K i Indicates the first i The rate constant of change of volatile basic nitrogen during each isothermal phase. t i Indicates the first i Duration of each isothermal phase B i-1 express t =d t i-1 The content of volatile basic nitrogen at that time (mg / 100g).

[0053] A BPNN prediction model was constructed using the `newff` function in Matlab. All experimental data obtained from storage conditions at 0, 5, 10, and 15°C were used for model training and validation. In the traditional BPNN model training, the `trainlm` function of the Levenberg-Marquardt algorithm was used as the training function, `learngdm` as the weight learning function, `tansig` as the hidden layer transfer function, and `purelin` as the output layer transfer function. An adaptive learning rate method (Adam) was introduced during training, enabling the network to dynamically adjust the parameter update step size based on historical gradient information. This effectively avoids gradient oscillations and accelerates convergence under noisy low-temperature storage data conditions. Simultaneously, a Bayesian optimization method was used to perform a global search for the number of hidden layer neurons to automatically obtain the optimal configuration. Through joint optimization of the hidden layer structure, training strategy, and hyperparameters, a high-precision BPNN model suitable for predicting the quality of fish meat stored at low temperatures was finally constructed.

[0054] An RBFNN prediction model was constructed using Matlab based on the radial basis function network construction method, achieving near-zero error fitting on the training set. The training samples were the same as those for the BPNN model. To improve model performance, K-means clustering was first used to cluster the training samples, automatically determining the centers of the radial basis functions to reduce the number of hidden layer neurons and lower network complexity. Subsequently, a Bayesian optimization method was introduced to intelligently search and adjust the spread parameter of the radial basis functions. Bayesian optimization constructs a Gaussian process surrogate model and combines it with the acquisition function to select the optimal candidate points, finding the spread parameter that enables the model to achieve optimal prediction performance with fewer experiments.

[0055] To enhance the model's generalization ability and ensure that the prediction results conform to the dynamic laws, the dynamic prediction model is introduced as prior knowledge into the training process, and a fusion loss function is constructed: L total = α·L data + β·L physics + γ·L reg Where L data For the data fitting term; where L physics For the physical constraint loss term; where L reg The regularization term, a penalty weight sum-of-squares regularization term, is added to suppress extreme weights, resulting in a smoother model output and better generalization. α, β, and γ are adjustable weight parameters used to balance data fitting accuracy and physical consistency. Through this optimization process, the BPNN / RBFNN model can simultaneously utilize the nonlinear fitting capabilities driven by data and the mechanistic constraints of the dynamic model, effectively improving prediction accuracy and robustness.

[0056] The data loss term is calculated using the following formula:

[0057] in These are the model's predicted values. is the experimentally determined value, and n is the sample size.

[0058] The physical constraint loss term is calculated using the following formula:

[0059] in, , These are the predicted values ​​from the RBFNN model and the dynamic model, respectively, where n is the number of samples.

[0060] The prediction residuals of the BPNN / RBFNN neural network model are iteratively fitted by gradient boosting, thereby performing a secondary correction on the prediction results.

[0061] To establish a quality early warning model for fish meat segmentation, the differential of the temperature-dependent kinetic model of volatile basic nitrogen content was obtained, yielding the following expression:

[0062] This formula represents the slope of the first-order reaction kinetic curve. When the slope reaches a certain value, the curve rises rapidly. This value can be used as an early warning threshold for the volatile basic nitrogen model. By monitoring the rate of change of volatile basic nitrogen content, the quality changes of cut fish meat can be predicted in a timely manner, providing a scientific basis for the storage and preservation of cut fish meat.

[0063] A stacked ensemble learning framework is constructed, which combines the prediction results of multiple base learners through a single meta-learner, thereby integrating a heterogeneous model and improving overall prediction performance. Specifically, the base learners used in this invention include a dynamic model (a modified Gompertz model / Arrhenius equation) and two optimized neural network models (BPNN / RBFNN):

[0064] The stacked ensemble learning framework "Meta-learner" uses the CART tree algorithm to fuse the prediction results of the first layer. It is constructed using a binary recursive partitioning approach. At each node, it traverses all input variables and their candidate split points, selecting the one that minimizes the mean squared error (MSE) of the samples within the node. Splitting stops when the node depth reaches a set maximum depth. The input vector Z(t) is partitioned and mapped using the CART tree to obtain the final prediction result.

[0065] To evaluate the predictive performance and accuracy of the prediction model, several performance metrics are used to assess its predictive effectiveness, including the coefficient of fit. R 2 The errors include relative error (RE), mean bias error (MBE), mean absolute percentage error (MAPE), and root mean square error (RMSE).

[0066] Example 2: This embodiment models the preservation of different parts of the catfish, using total bacterial count and volatile basic nitrogen content as core indicators to predict and verify the shelf life of the catfish. The specific implementation plan is as follows.

[0067] (I) Construction of a dynamic model for total colony count 1. Pre-treatment of fish meat cutting: Samples were sealed in polyethylene bags and placed in a constant temperature and humidity incubator under different storage temperatures (0, 4, 5, 10, 15°C). During storage, samples were taken periodically at preset time points (every 2 days for storage at 0, 4, and 5°C, and every 12 hours for storage at 10 and 15°C) to determine the total bacterial count and volatile basic nitrogen content. Data measured at 0, 5, 10, and 15°C were used as the training set for establishing a prediction model, while data measured at 4°C were used as the validation set for evaluating the prediction model. Furthermore, a fluctuating temperature experiment was designed to simulate temperature fluctuations that may occur during actual cold chain transportation. The specific temperature settings were: 4°C for 2 days → 9°C for 2 days (sampling point a) → 14°C for 1 day → 9°C for 2 days → 5°C for 2 days (sampling point b). In the fluctuating temperature storage experiment, the total bacterial count and volatile basic nitrogen content were measured at regular intervals.

[0068] 2. Testing Method: The content of volatile basic nitrogen was determined according to the method in GB5009.228-2016 "National Food Safety Standard - Determination of Volatile Basic Nitrogen in Food"; the total bacterial count was determined according to the method in "National Food Safety Standard - Microbiological Examination of Food - Determination of Total Bacterial Count".

[0069] 3. The modified Gompertz equation was used to perform nonlinear fitting on the change of total bacterial count over time in fish meat products stored at temperatures ranging from 0 to 15°C. The fitting parameters are shown in Table 1. Table 1 Microbial growth parameters

[0070] The coefficients of determination for each fit R 2 A value greater than 0.9 indicates that the model can effectively describe the growth dynamics of microorganisms in different parts of catfish under different storage temperatures. Furthermore, changes in storage temperature significantly affect the lag phase of microbial growth. λ and maximum specific growth rate μ max As storage temperature increases, microbial growth... λ Significantly shortened, while μ max and N max Then it increases significantly.

[0071] Based on the first-level model, the Belehradek equation (square root model) is used as the second-level model. Figure 1 This demonstrates the effect of temperature on temperature within the range of 0~15°C. and The linear relationship is presented. The Belehradek equation shows good fitting results for the microbial growth parameters during the storage of fish meat products at different temperatures, with a coefficient of determination. R 2 All values ​​were greater than 0.87, indicating that the established secondary model can accurately describe the microbial activity under different temperature conditions. μ max and λ The changing pattern.

[0072] Table 1 N max Substituting into the second-order equation, we can obtain the kinetic model for total bacterial count at temperatures ranging from 0 to 15°C, namely: head:

[0073] Beef belly:

[0074] belly:

[0075] back:

[0076] tail:

[0077] (II) Construction of a Neural Network Model for Total Colony Count like Figure 2 As shown in Table 2, the BPNN model was trained and modeled using the newff function in MATLAB, and optimized based on the constructed fusion loss function and gradient boosting method.

[0078] Table 2 shows the prediction performance of the BPNN training set using total bacterial count as an indicator.

[0079] Fit coefficients for different parts R 2 A value greater than 0.98 indicates a good model fit; furthermore, an MBE within ±0.05 indicates a small systematic deviation between the model's predicted and actual values; a MAPE less than 0.05 indicates high prediction accuracy; and an RMSE less than 0.4 further validates the model's predictive ability. These results demonstrate that the optimized BPNN model can effectively predict changes in total bacterial count in fish fillets during storage, providing a reliable tool for real-time monitoring and prediction of their quality.

[0080] like Figure 3As shown, the RBFNN model was trained and modeled using the `newrbe` function in MATLAB. Optimization was performed based on the constructed fusion loss function and gradient boosting method. Total bacterial count data obtained under storage conditions of 0, 5, 10, and 15°C were used as the training set for training the RBFNN model. The established RBFNN model was used to predict the change in total bacterial count of fish meat products during storage. The accuracy of the model is shown in Table 3.

[0081] Table 3. Prediction performance of RBFNN training set based on total colony count.

[0082] Different parts R 2 A value greater than 0.95 indicates a good model fit. With an MBE within ±0.01, MAPE less than 0.07, and RMSE less than 0.6, the optimized RBFNN model can effectively predict changes in total bacterial count in fish fillets during storage.

[0083] (III) Construction of a kinetic model for volatile basic nitrogen The changes in volatile basic nitrogen content over time in five parts of the catfish under storage temperatures ranging from 0 to 15°C were nonlinearly fitted using first-order kinetic equations. The volatile basic nitrogen content exhibited an exponential increasing trend with prolonged storage time, and the rate of increase varied significantly under different temperature conditions. The fitted parameters are shown in Table 4. Table 4. Kinetic model parameters for changes in volatile basic nitrogen content of fish meat products during storage under different temperature conditions.

[0084] The coefficients of determination for each fit R 2 All values ​​are greater than 0.9, indicating that the first-order kinetic equation effectively describes the dynamics of the increase in volatile basic nitrogen content in fish meat products at different storage temperatures. The rate constant of change in volatile basic nitrogen content increases with increasing storage temperature. K Significant increase.

[0085] The reaction order and calculated reaction constant obtained from the established Arrhenius equation are shown in Table 5, and the fitting coefficients are... R 2 A value greater than 0.95 indicates that the Arrhenius equation has a good fitting effect.

[0086] Table 5. Relationship between temperature and rate of change parameters

[0087] Substituting the equations in the table above into the second-order equations, we obtain the growth kinetic model for volatile basic nitrogen at fluctuating temperatures, namely: head:

[0088] Beef belly:

[0089] belly:

[0090] back:

[0091] tail:

[0092] (iv) Construction of a neural network model for volatile basic nitrogen The BPNN model is trained and modeled using MATLAB's `newff` function, such as... Figure 4 As shown in Table 6, the training set results are optimized based on the constructed fusion loss function and gradient boosting method.

[0093] Table 6. Prediction performance of the BPNN training set using volatile basic nitrogen content as an indicator.

[0094] Fit coefficients R 2 A value greater than 0.85 indicates a good model fit. Furthermore, with MBE within ±0.9, MAPE less than 0.3, and RMSE less than 4.5, the systematic bias between the model's predictions and measured values ​​is small. These results demonstrate that the optimized BPNN can effectively predict changes in volatile basic nitrogen content in fish fillets.

[0095] The RBFNN model was trained and modeled using the `newrbe` function in MATLAB, and optimized based on the constructed fusion loss function and gradient boosting method. A three-dimensional structure of the RBFNN prediction model for fish meat segmentation storage indicators was established, as shown below. Figure 5 As shown in Table 7, the accuracy of the established model is as follows.

[0096] Table 7. Prediction performance of the RBFNN training set using volatile basic nitrogen content as an indicator.

[0097] Fit coefficients for different parts R 2The values ​​are greater than 0.8, MBE is close to 0, MAPE is less than 0.5, and RMSE is less than 6.5, indicating that the optimized RBFNN model can predict the changes in volatile basic nitrogen content of fish meat products during storage.

[0098] (v) Validate and evaluate the prediction model. The predicted values ​​of total bacterial count in fish meat products were compared with the measured values ​​using the kinetic model. The results are shown in Table 8. In the early stage of storage (within 2 days), the prediction effect was better, with a relative error within ±15%.

[0099] Table 8. Predicted and measured total bacterial count values ​​of fish meat products during storage.

[0100] The predicted values ​​of volatile basic nitrogen content in fish meat segments were compared with the measured values ​​by the kinetic model. The results are shown in Table 9. Except for day 0, the relative errors of different parts were basically within ±15%, indicating that the model can accurately predict the volatile basic nitrogen content of different parts of catfish at different temperatures.

[0101] Table 9. Predicted and measured values ​​of volatile basic nitrogen content in fish meat products during storage (kinetic model).

[0102] The measured total bacterial count was used as the experimental value, and the error was compared with the prediction results of the BPNN model to evaluate the accuracy of the model. As shown in Table 10, except for the relative error of -13.82% on the second day of the head, the relative errors were all within ±10%. This result indicates that the BPNN model can accurately predict the change in total bacterial count of fish meat products under storage conditions of 4°C in most cases.

[0103] Table 10. Predicted and Measured Total Bacterial Count of Fish Meat Products During Storage (BPNN Model)

[0104] The measured total bacterial count was used as the experimental value, and the error was compared with the prediction results of the RBPNN model to evaluate the accuracy of the model. As shown in Table 11, except for the relative errors of the tail on days 0-6 and the back on day 2, the relative errors of the others were all within ±20%. This result indicates that the RBFNN model can accurately predict the changes in the total bacterial count of fish fillets stored at 4°C in most cases.

[0105] Table 11. Predicted and Measured Total Bacterial Count of Fish Meat Products During Storage (RBPNN Model)

[0106] The measured volatile basic nitrogen values ​​were compared with the prediction results of the BPNN model to evaluate the accuracy of the model. As shown in Table 12, the relative error was generally within ±15%. This result indicates that the BPNN model can accurately predict the changes in volatile basic nitrogen content in different parts of catfish under 4°C storage conditions in most cases, providing reliable support for real-time monitoring of catfish quality.

[0107] Table 12. Predicted and measured values ​​of volatile basic nitrogen during storage of fish meat products (BPNN model)

[0108] The measured volatile basic nitrogen values ​​were used as experimental values, and their errors were compared with the prediction results of the RBPNN model to evaluate the accuracy of the model. As shown in Table 13, except for days 0-2, the relative errors of each part were basically within ±20%.

[0109] Table 13. Predicted and measured values ​​of volatile basic nitrogen during storage of fish meat products (RBPNN model)

[0110] The measured total bacterial count was used as the experimental value, and the error was compared with the final prediction result under the stacked ensemble learning framework to evaluate the accuracy of the model. As shown in Table 14, the stacked ensemble learning method has high prediction accuracy, with the total bacterial count prediction error all below 10%, which is better than independent neural network models and kinetic models. In particular, for the 2-4 day storage stage, where the prediction effect of single models is relatively poor, the prediction accuracy of the stacked ensemble learning method has been significantly improved. This result shows that the stacked ensemble learning model combines the advantages of single models and can accurately predict the change of total bacterial count in fish meat products under 4°C storage conditions in most cases.

[0111] Table 14. Predicted and Measured Total Bacterial Count of Fish Meat Products During Storage (Stacked Ensemble Learning)

[0112] The measured volatile basic nitrogen values ​​were compared with the final prediction results under the stacked ensemble learning framework to evaluate the accuracy of the model. As shown in Table 15, the stacked ensemble learning method exhibits high prediction accuracy, with volatile basic nitrogen prediction errors all below 15%, outperforming independent neural network models and kinetic models. Particularly for the 0-2 day storage stage, where single models perform relatively poorly, the stacked ensemble learning method demonstrates better performance. This result indicates that the stacked ensemble learning model combines the advantages of single models and can accurately predict changes in volatile basic nitrogen in fish fillets stored at 4°C in most cases.

[0113] Table 15. Predicted and Measured Values ​​of Volatile Basic Nitrogen During Storage of Fish Meat Products (Stacked Ensemble Learning)

[0114] (vi) Establishment of prediction and early warning models Establishing a fish meat quality early warning time point model is crucial for ensuring food safety and reducing waste. By establishing early warning time points, timely measures can be taken to avoid economic losses caused by the spoilage of fish meat products. By substituting the kinetic model parameters of the change in volatile basic nitrogen content during the storage of fish meat products into the slope of the first-order reaction kinetic curve, the instantaneous change rate of volatile basic nitrogen or total bacterial count over time can be obtained. This can reflect the dynamic changes in the spoilage of catfish during storage, determine when the rate of change of volatile basic nitrogen content reaches a certain critical value, and thus set an early warning threshold based on the requirements of GB 2733-2015 National Food Safety Standard for Fresh and Frozen Aquatic Animal Products, as shown in Table 16.

[0115] Table 16 Warning Thresholds for Fish Meat Cuts under Different Temperature Conditions

[0116] (vii) Construction and evaluation of shelf life prediction models The shelf life of different parts of the catfish was obtained using a first-order chemical reaction kinetic model of volatile basic nitrogen, as shown in the following formula.

[0117]

[0118] In the formula: SL is the shelf life (days); B is the limit value of volatile basic nitrogen at the end of the shelf life (mg / 100g); B0 is the initial volatile basic nitrogen content (mg / 100g); T is the absolute temperature (K); R is the gas constant, 8.3144 J / (mol·K).

[0119] The predictive model was validated and evaluated using shelf-life measurements taken at 0–15°C. According to international standards, fish are considered spoiled and their shelf life terminated when the volatile basic nitrogen value exceeds 20 mg / 100g. The shelf-life was calculated by substituting 20 mg / 100g into the formula above and compared with the measured values; the results are shown in Table 17.

[0120] Table 17. Predicted and measured values ​​of the shelf life of catfish stored at different temperatures.

[0121] The relative errors for different parts at different temperatures were all within ±20% (except for the error in the head under 0°C storage conditions), indicating that the prediction model has good accuracy. That is, the shelf life prediction model can predict the shelf life of different parts of catfish under 0~15°C storage conditions.

[0122] This embodiment uses catfish fillet as an example to verify the low-temperature preservation quality prediction method. A microbial growth kinetic model and a volatile basic nitrogen change kinetic model are constructed, which can accurately describe the changes in total bacterial count and volatile basic nitrogen at temperatures ranging from 0 to 15°C and fluctuating temperatures, overcoming the limitation of traditional models in adapting to actual cold chain temperature fluctuations. The BPNN and RBFNN models, optimized with adaptive learning rate, Bayesian optimization, and fusion loss function, achieve a high coefficient of determination. R² All values ​​were above 0.85, indicating small prediction bias and strong stability. Using the kinetic model, optimized BPNN, and optimized RBFNN as base learners, and fused with a CART tree-element learner for decision-making, the prediction error for total bacterial count was less than 10%, and the prediction error for volatile basic nitrogen was less than 15%, significantly improving prediction accuracy. Furthermore, an early warning threshold was established based on the derivative of the volatile basic nitrogen change kinetic model, enabling accurate early warning of quality deterioration and reliable prediction of shelf life, further improving the accuracy of quality control during the low-temperature storage and transportation of cut fish.

[0123] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the quality of segmented fish meat during low-temperature preservation, characterized in that, Includes the following steps: (1) Collect quality data of cut fish meat under constant temperature and fluctuating temperature storage conditions, and determine the total number of colonies and volatile basic nitrogen content as the core predictive indicators. (2) Construct a kinetic model of microbial growth and a kinetic model of changes in volatile basic nitrogen; (3) Construct an optimized BPNN prediction model and an optimized RBFNN prediction model that integrate physical constraints; (4) Construct stacked ensemble learning frameworks with kinetic model, optimized BPNN model, and optimized RBFNN model as base learners and CART tree as meta learner, respectively, to predict the total number of colonies and the content of volatile basic nitrogen; (5) Input the prediction results of each base learner into the meta learner and fuse the output to obtain the final predicted value of the low-temperature preservation quality of segmented fish meat, and set the warning threshold and shelf life prediction based on the prediction results.

2. The method for predicting the quality of segmented fish meat during low-temperature preservation according to claim 1, characterized in that, The construction of the microbial growth kinetics model includes: Constructing a primary model of microbial growth: in, t Storage time; N ( t )for t The number of colonies at any given time; N 0 represents the initial colony count; N max The maximum number of colonies; μ max This represents the maximum specific growth rate. λ The lag period; Constructing a two-stage model of microbial growth: in, , These are the equation coefficients; T Storage temperature (°C); T min This indicates the minimum growth temperature threshold.

3. The method for predicting the quality of segmented fish meat during low-temperature preservation according to claim 2, characterized in that, The microbial growth kinetics model is further decomposed into multiple isothermal processes, represented as follows: when t =d t At 1 o'clock, when t =d t 1+d t At 2 o'clock, when t =d t 1+d t 2+……+ d t i hour Where, d t i This represents a short time interval assuming a constant temperature. N ( t i () indicates the number of colonies; This indicates the maximum specific growth rate.

4. The method for predicting the quality of segmented fish meat during low-temperature preservation according to claim 1, characterized in that, The kinetic model for the change of volatile basic nitrogen is expressed as follows: when t =d t At 1 o'clock, when t = d t 1+d t At 2 o'clock, when t =d t 1+d t 2+……+ d t i hour in, B ( t i () represents the volatile basic nitrogen content; K is the rate of change of volatile basic nitrogen. t Storage time; The rate of change of volatile basic nitrogen is expressed as follows: in, k 0 is the frequency factor, and R is the gas constant. E a It is the activation energy.

5. The method for predicting the quality of segmented fish meat during low-temperature preservation according to claim 4, characterized in that, The expression for predicting shelf life is: In the formula: Shelf life; B These are the limits for volatile basic nitrogen at the end of the shelf life; B 0 represents the initial volatile basic nitrogen content.

6. The method for predicting the quality of segmented fish meat during low-temperature preservation according to claim 1, characterized in that, Both the optimized BPNN model and the optimized RBFNN model are trained using a fusion loss function, which is: L total = α·L data + β·L physics + γ·L reg Where L data Let L be the data fitting loss term, where L is the loss term. physics For the physical constraint loss term, L reg α is the regularization term, and β and γ are the weight parameters.

7. The method for predicting the quality of segmented fish meat during low-temperature preservation according to claim 4, characterized in that, Step (5) includes: establishing a quality early warning model for segmented fish meat, wherein the early warning model is obtained by differentiating the temperature-dependent kinetic model of volatile basic nitrogen content.

8. The method for predicting the quality of segmented fish meat during low-temperature preservation according to claim 1, characterized in that, The optimized BPNN model employs Adam adaptive learning rate optimization during the training process, and determines the optimal number of hidden layer neurons through Bayesian optimization.

9. The method for predicting the quality of segmented fish meat during low-temperature preservation according to claim 1, characterized in that, The optimized RBFNN model uses K-means clustering to determine the radial basis function centers and adjusts the radial basis function spread parameter through Bayesian optimization.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method for predicting the quality of segmented fish meat under low-temperature preservation as described in any one of claims 1 to 9.

Images