Visual early warning method for quality deterioration of non-living aquatic product supply chain

Abstract

The invention discloses a non-living aquatic product supply chain quality deterioration visual early warning method, and belongs to the field of non-living aquatic product dynamic monitoring, and the method comprises the following steps: S1, synchronously collecting the physical appearance, quality indication labels and environmental parameters of non-living aquatic products, and constructing a multi-dimensional data set; s2, extracting complementary features, and outputting weighted fusion features; s3, establishing a residual shelf life prediction model, and obtaining a residual shelf life prediction value and a confidence interval; s4, based on the residual shelf life prediction value, constructing a supply chain digital twinborn body, and mapping a physical entity state in real time; and S5, continuously optimizing parameters of the residual shelf life prediction model. By adopting the non-living aquatic product supply chain quality deterioration visual early warning method, multi-modal feature deep interaction, digital twinborn visualization and adaptive iteration are fused, a data acquisition-feature fusion-prediction early warning-model evolution closed loop is constructed, and the limitation of a traditional single-modal and static model is broken through.

Description

Technical Field

[0001] This invention relates to the field of dynamic monitoring technology for non-live aquatic products, and in particular to a visual early warning method for quality deterioration in the supply chain of non-live aquatic products. Background Technology

[0002] With the growth of global trade in non-live aquatic products, quality deterioration in the supply chain leads to approximately 30% of the annual loss of these products, while food safety incidents caused by spoilage are frequent. To address this issue, the industry has begun developing the following technologies for monitoring quality deterioration in non-live aquatic products:

[0003] I. Single-modal detection technologies: 1. Image-based appearance detection technology: This technology acquires color images of non-live aquatic products (such as salmon) and uses convolutional neural networks (CNNs) to extract color and texture features to classify freshness and predict shelf life. However, because it relies solely on appearance features and ignores the influence of environmental parameters (such as temperature and ammonia concentration) on microbial growth, the prediction accuracy is affected by lighting and shooting angle, and it cannot reflect intrinsic quality (such as histamine content). 2. Sensor-based environmental monitoring technology: This technology uses gas sensors (such as ammonia and carbon dioxide sensors), time-temperature indicators (TTIs), etc., to indirectly assess quality by detecting changes in environmental parameters. However, because sensor data has a weak correlation with the actual deterioration of non-live aquatic products, it cannot integrate appearance features; and traditional TTIs only provide qualitative results (such as "fresh / spoiled"), lacking quantitative remaining shelf life (RSL) prediction capabilities.

[0004] II. Traditional Early Warning and Visualization Technologies: 1. Simple threshold-based early warning systems: These systems set fixed thresholds (e.g., triggering an alert when ammonia concentration > 50 ppm) and then push alerts via indicator lights or SMS. They suffer from rigid early warning rules, fail to consider the coupling relationships between parameters (e.g., high temperatures accelerate ammonia release), and lack a visual interface, making it difficult for personnel at various stages of the supply chain to intuitively judge the location and severity of risks. 2. Static data visualization technologies: These technologies display historical sensor data (e.g., temperature change curves) through tables or line graphs to assist human decision-making. However, they cannot map the real-time status of entities in the supply chain and fail to reflect the dynamic correlation between "location-time-quality," leading to delayed early warnings. Summary of the Invention

[0005] The purpose of this invention is to provide a visual early warning method for quality deterioration in the supply chain of non-live aquatic products, thereby solving the above-mentioned technical problems.

[0006] To achieve the above objectives, this invention provides a method for visually predicting quality deterioration in the supply chain of non-live aquatic products, comprising the following steps:

[0007] S1. Multi-source heterogeneous data acquisition and standardization processing: Simultaneously collect data on the physical appearance, quality indicator labels, and environmental parameters of non-live aquatic products to construct a multi-dimensional dataset;

[0008] S2, Cross-modal feature fusion and key indicator extraction: Extract complementary features from physical appearance, environment and microbial indicators, strengthen key information through attention mechanism, and output weighted fusion features;

[0009] S3. Construction of a quality deterioration prediction model based on spatiotemporal series: Based on weighted fusion features, combined with long-term and short-term time series dependencies and nonlinear mapping, a remaining shelf life prediction model is established, and real-time data is input into the remaining shelf life prediction model to obtain the remaining shelf life prediction value and confidence interval.

[0010] S4. Building a digital twin-driven visualization early warning system: Based on the remaining shelf life forecast, construct a digital twin of the supply chain, map the physical entity status in real time, and realize early warning of deterioration risk through multi-dimensional visualization;

[0011] S5. Model Adaptive Iteration and Robust Optimization: Continuously optimize the parameters of the remaining shelf life prediction model based on real-time collected multidimensional data.

[0012] Preferably, step S1 specifically includes the following steps:

[0013] S11. Image acquisition and enhancement of the physical appearance of non-live aquatic products;

[0014] S111. Determine key time points in the supply chain of non-live aquatic products, and use industrial cameras to collect color images of non-live aquatic products at fixed shooting distances and light intensities.

[0015] S112. Image Preprocessing: The acquired images are cropped, brightness adjusted, and Gaussian filtered sequentially to obtain a standardized image set i = {i1, I2, ..., I...} n},I1,i2,...,I n These represent the 1st, 2nd, ..., nth acquired images within the standardized image set;

[0016] Among them, brightness adjustment is based on the CIELab color space:

[0017]

[0018] In the formula, Δa * Δb * and ΔL * These represent the adjustment amounts for red-green difference, yellow-blue difference, and brightness difference, respectively; ΔE * This represents the total color difference before and after image brightness adjustment in the CIELab color space;

[0019] S12. Collection of environmental and microbiological parameters;

[0020] S121. Deploy a smart sensor array: Deploy temperature sensors, humidity sensors, NH3 sensors, CO2 sensors, and time-temperature indicators;

[0021] S122. Samples are taken every 24 hours to test total bacterial count (CFU) and histamine concentration. The assay was performed using an enzyme-linked immunosorbent assay (ELISA), with a detection limit of 0.1 ppm.

[0022] Synchronously acquire time-temperature indicator data and calculate the cumulative temperature effect (TTIs) based on the Arrhenius equation:

[0023]

[0024] In the formula, t total Indicates the total monitoring duration; k r (T absolute (τ) represents the absolute temperature T at time τ. absolute Below, the rate constant of the deterioration reaction of non-living aquatic products; k r (T j,absolute () represents the absolute temperature T during the j-th data collection. j,absolute The rate constant of the deterioration reaction of non-living aquatic products; Δτ j T represents the absolute temperature of the j-th data collection. j,absolute The time interval; n represents the total number of data collections;

[0025] in,

[0026] k r =A p ×exp(-E a / (R g T absolute ));

[0027] In the formula, k r A represents the rate constant for the deterioration reaction of non-living aquatic products; p E represents the pre-exponential factor; a R represents the activation energy. g T represents the gas constant; absolute Indicates absolute temperature;

[0028] S123. Data Standardization: Environmental parameters are normalized using the min-max method, and a standardized parameter matrix is ​​output. Where, T environment and H H These represent ambient temperature and ambient humidity, respectively. and These represent the concentrations of ammonia, carbon dioxide, and histamine, respectively; CFU represents the total bacterial count.

[0029] Preferably, in step S111, the key time nodes include within 1 hour after fishing, every 6 hours during transportation, and every 12 hours during storage;

[0030] The fixed shooting distance is 50m and the light intensity is 5000 lux.

[0031] Preferably, step S2 specifically includes the following steps:

[0032] S21. Image depth feature extraction;

[0033] S211. Construct an improved CNN model. The improved CNN model includes convolutional layers, pooling layers, and two fully connected layers arranged sequentially. The convolutional layers consist of four 3×3 convolutional kernels with 64, 128, 256, and 512 channels respectively, using the ReLU activation function. The pooling layers use 2×2 max pooling with a stride of 2. The two fully connected layers have 512 and 256 neurons respectively, making the output image feature vector F... img =CNN(I), and F img ∈R 256 R represents the set of real numbers;

[0034] S212. Introducing supplementary color features: Calculating the red shift D of the image. R :

[0035] D R =R e / (R e +G+B);

[0036] Where R e G and B represent the pixel mean values ​​of the red, green, and blue channels of the image, respectively.

[0037] S213, Move the red color to D R With F img The concatenated enhanced image feature vector F' is obtained by concatenating the features. img ∈R 257 ;

[0038] S22. Environmental and Microbial Feature Mapping: A Kolmogorov-Arnold network is used to process the normalized parameter matrix P. Nonlinear mapping is then used to capture the coupling relationship between temperature, humidity, and microbial growth, yielding the environmental and microbial feature vector F. env =KAN(P), and F env ∈R 64 ;

[0039] S23. Cross-modal attention fusion: Construct a multi-head attention mechanism and calculate F'. img With F env Association weights:

[0040]

[0041] In the formula, Q, K, and V represent the query key, matching key, and value, respectively, and Q = F' img K = V = F env ;T represents the transpose operation;d k Denotes the feature dimension, and d k =64;

[0042] S24, Output weighted fusion feature F fusion (I,P)∈R 128 .

[0043] Preferably, step S3 specifically includes the following steps:

[0044] S31. Using histamine concentration ≥50ppm as the deterioration endpoint, determine the total shelf life (T) using chemical and biological methods. total Given the remaining shelf life (RSL), obtain the label dataset, where RSL = T. total -T current T current This indicates the time corresponding to the current freshness level;

[0045] S32. Divide the label dataset into training and test sets in a 7:3 ratio, and use time series cross-validation to avoid data leakage;

[0046] S33. Construct a remaining shelf life prediction model, which includes an input layer, two Bi LSTM layers, a KAN layer, and an output layer set in sequence.

[0047] S34. Set the training strategy: Use mean squared error as the loss function and the Adam optimizer.

[0048] S35. Input the training set into the remaining shelf life prediction model and train it using the set training strategy until convergence.

[0049] S36. Input the test set into the trained remaining shelf life prediction model, and evaluate the model using Kendall rank correlation coefficient and mean absolute error.

[0050] S37. Input the real-time collected multidimensional data into the remaining shelf life prediction model to predict the remaining shelf life:

[0051]

[0052] In the formula, Let f(·) represent the predicted remaining shelf life; f(·) represent the remaining shelf life prediction model; and CI represent the confidence interval. z α / 2 This represents the quantile corresponding to the confidence level. This represents the set of results from M predictions. Let represent the result of the i-th prediction, and σ represent the standard deviation.

[0053] Preferably, step S4 specifically includes the following steps:

[0054] S41, Supply Chain Digital Twin Modeling;

[0055] S411. Construct a virtual scene based on Unity 3D, and set it to receive multi-dimensional data and remaining shelf life prediction values ​​every 30 minutes, and update the virtual entity attributes;

[0056] S412. Visualization of Appearance Deterioration: Based on Image Feature Vector F img Simulate color changes in non-live aquatic products in a virtual environment;

[0057] S42, Multi-level early warning: Early warning levels are divided by setting early warning thresholds.

[0058] Preferably, in step S412, when At that time, non-live aquatic products were determined to be fresh;

[0059] when At that time, non-live aquatic products were determined to be less fresh;

[0060] when At that time, non-live aquatic products were judged to have deteriorated;

[0061] In step S42, when At that time, it was classified as a Level 1 warning;

[0062] when At that time, it was classified as a Level II warning;

[0063] when At that time, it was divided into three levels of early warning.

[0064] Therefore, the beneficial effects of the above-mentioned visual early warning method for quality deterioration in the supply chain of non-live aquatic products are as follows:

[0065] 1. Accurate fusion of multi-dimensional features: Breaking through the traditional reliance on single data from images / sensors, this method extracts appearance features (including red shift supplementation) through an improved CNN, maps environment-microorganism features using KAN, and then dynamically associates the "image-environment" coupling relationship through multi-head attention (such as the association between the red fading of salmon and high temperature / ammonia), making the capture of deterioration features more comprehensive and laying the foundation for accurate prediction.

[0066] 2. Digital Twin Visualization Decision Making: Construct a virtual twin of the supply chain to transform abstract RSL forecast values ​​into intuitive color coding (freshness visualization) and spatial mapping (risk node location), solving the problem of "data abstraction and decision lag" in traditional early warning systems, and supporting rapid response by personnel at all stages of the supply chain (such as warehouse personnel prioritizing red-alert batches).

[0067] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0068] Figure 1 This is a flowchart of a method for visually predicting quality deterioration in the supply chain of non-live aquatic products according to the present invention.

[0069] Figure 2 This is a partial image dataset of salmon collected in the simulation experiment of this invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0071] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0072] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0073] like Figure 1 As shown, a method for visually predicting quality deterioration in the supply chain of non-live aquatic products includes the following steps:

[0074] S1. Multi-source heterogeneous data acquisition and standardization processing: Simultaneously collect data on the physical appearance, quality indicator labels, and environmental parameters of non-live aquatic products to construct a multi-dimensional dataset;

[0075] Step S1 specifically includes the following steps:

[0076] S11. Image acquisition and enhancement of the physical appearance of non-live aquatic products;

[0077] S111. Determine key time points in the supply chain of non-live aquatic products, and use industrial cameras to collect color images of non-live aquatic products at fixed shooting distances and light intensities.

[0078] In step S111, key time nodes include within 1 hour after fishing, every 6 hours during transportation, and every 12 hours during storage;

[0079] The fixed shooting distance is 50m and the light intensity is 5000 lux.

[0080] S112. Image Preprocessing: The acquired images are cropped, brightness adjusted, and Gaussian filtered sequentially to obtain a standardized image set i = {i1, I2, ..., I...} n},I1,i2,...,I n These represent the 1st, 2nd, ..., nth acquired images within the standardized image set;

[0081] Among them, brightness adjustment is based on the CIELab color space:

[0082]

[0083] In the formula, Δa * Δb * and ΔL * These represent the adjustment amounts for red-green difference, yellow-blue difference, and brightness difference, respectively; ΔE * This represents the total color difference before and after image brightness adjustment in the CIELab color space;

[0084] S12. Collection of environmental and microbiological parameters;

[0085] S121. Deploy a smart sensor array: Deploy temperature sensors, humidity sensors, NH3 sensors, CO2 sensors, and time-temperature indicators;

[0086] S122. Samples are taken every 24 hours to test total bacterial count (CFU) and histamine concentration. The assay was performed using an enzyme-linked immunosorbent assay (ELISA), with a detection limit of 0.1 ppm.

[0087] Synchronously acquire time-temperature indicator data and calculate the cumulative temperature effect (TTIs) based on the Arrhenius equation:

[0088]

[0089] In the formula, t total Indicates the total monitoring duration; k r (Tabsolute (τ) represents the absolute temperature T at time τ. absolute Below, the rate constant of the deterioration reaction of non-living aquatic products; k r (T j,absolute () represents the absolute temperature T during the j-th data collection. j,absolute The rate constant of the deterioration reaction of non-living aquatic products; Δτ j T represents the absolute temperature of the j-th data collection. j,absolute The time interval; n represents the total number of data collections;

[0090] in,

[0091] k r =A p ×exp(-E a / (R g T absolute ));

[0092] In the formula, k r A represents the rate constant for the deterioration reaction of non-living aquatic products; p E represents the pre-exponential factor; a R represents the activation energy. g T represents the gas constant; absolute Indicates absolute temperature;

[0093] S123. Data Standardization: Environmental parameters are normalized using the min-max method, and a standardized parameter matrix is ​​output. Where, T environment and H H These represent ambient temperature and ambient humidity, respectively. and These represent the concentrations of ammonia, carbon dioxide, and histamine, respectively; CFU represents the total bacterial count.

[0094] S2, Cross-modal feature fusion and key indicator extraction: Extract complementary features from physical appearance, environment and microbial indicators, strengthen key information through attention mechanism, and output weighted fusion features;

[0095] Step S2 specifically includes the following steps:

[0096] S21. Image depth feature extraction;

[0097] S211. Construct an improved CNN model. The improved CNN model includes convolutional layers, pooling layers, and two fully connected layers arranged sequentially. The convolutional layers consist of four 3×3 convolutional kernels with 64, 128, 256, and 512 channels respectively, using the ReLU activation function. The pooling layers use 2×2 max pooling with a stride of 2. The two fully connected layers have 512 and 256 neurons respectively, making the output image feature vector F...img =CNN(I), and F img ∈R 256 R represents the set of real numbers;

[0098] S212. Introducing supplementary color features: Calculating the red shift D of the image. R :

[0099] D R =R e / (R e +G+B);

[0100] Where R e G and B represent the pixel mean values ​​of the red, green, and blue channels of the image, respectively.

[0101] S213, Move the red color to D R With F img The concatenated enhanced image feature vector F' is obtained by concatenating the features. img ∈R 257 ;

[0102] S22. Environmental and Microbial Feature Mapping: A Kolmogorov-Arnold network is used to process the normalized parameter matrix P. Nonlinear mapping is then used to capture the coupling relationship between temperature, humidity, and microbial growth, yielding the environmental and microbial feature vector F. env =KAN(P), and F env ∈R 64 ;

[0103] S23. Cross-modal attention fusion: Construct a multi-head attention mechanism and calculate F'. img With F env Association weights:

[0104]

[0105] In the formula, Q, K, and V represent the query key, matching key, and value, respectively, and Q = F' img K = V = F env ;T represents the transpose operation;d k Denotes the feature dimension, and d k =64;

[0106] S24, Output weighted fusion feature F fusion (I,P)∈R 128 .

[0107] S3. Construction of a quality deterioration prediction model based on spatiotemporal series: Based on weighted fusion features, combined with long-term and short-term time series dependencies and nonlinear mapping, a remaining shelf life prediction model is established, and real-time data is input into the remaining shelf life prediction model to obtain the remaining shelf life prediction value and confidence interval.

[0108] Step S3 specifically includes the following steps:

[0109] S31. Using histamine concentration ≥50ppm as the deterioration endpoint, determine the total shelf life (T) using chemical and biological methods. total Given the remaining shelf life (RSL), obtain the label dataset, where RSL = T. total -T current T current This indicates the time corresponding to the current freshness level;

[0110] S32. Divide the label dataset into training and test sets in a 7:3 ratio, and use time series cross-validation to avoid data leakage;

[0111] S33. Construct a remaining shelf life prediction model. The remaining shelf life prediction model includes an input layer, two BiLSTM layers, a KAN layer, and an output layer set in sequence.

[0112] S34. Set the training strategy: Use mean squared error as the loss function and the Adam optimizer.

[0113] S35. Input the training set into the remaining shelf life prediction model and train it using the set training strategy until convergence.

[0114] S36. Input the test set into the trained remaining shelf life prediction model, and evaluate the model using Kendall rank correlation coefficient and mean absolute error.

[0115] S37. Input the real-time collected multidimensional data into the remaining shelf life prediction model to predict the remaining shelf life:

[0116]

[0117] In the formula, Let f(·) represent the predicted remaining shelf life; f(·) represent the remaining shelf life prediction model; and CI represent the confidence interval. z α / 2 This represents the quantile corresponding to the confidence level. This represents the set of results from M predictions. Let represent the result of the i-th prediction, and σ represent the standard deviation.

[0118] S4. Building a digital twin-driven visualization early warning system: Based on the remaining shelf life forecast, construct a digital twin of the supply chain, map the physical entity status in real time, and realize early warning of deterioration risk through multi-dimensional visualization;

[0119] Step S4 specifically includes the following steps:

[0120] S41, Supply Chain Digital Twin Modeling;

[0121] S411. Construct a virtual scene based on Unity 3D, and set it to receive multi-dimensional data and remaining shelf life prediction values ​​every 30 minutes, and update the virtual entity attributes;

[0122] S412. Visualization of Appearance Deterioration: Based on Image Feature Vector F img Simulate color changes in non-live aquatic products in a virtual environment;

[0123] S42, Multi-level early warning: Early warning levels are divided by setting early warning thresholds.

[0124] In step S412, when At that time, non-live aquatic products were determined to be fresh;

[0125] when At that time, non-live aquatic products were determined to be less fresh;

[0126] when At that time, non-live aquatic products were judged to have deteriorated;

[0127] In step S42, when At that time, it was classified as a Level 1 warning;

[0128] when At that time, it was classified as a Level II warning;

[0129] when At that time, it was divided into three levels of early warning.

[0130] S5. Model Adaptive Iteration and Robust Optimization: Continuously optimize the parameters of the remaining shelf life prediction model based on real-time collected multidimensional data.

[0131] Simulation Experiment

[0132] Data preparation: such as Figure 2 As shown, salmon supply chain data was collected: 1000 sets of images (480×480, covering the entire fresh to spoilage cycle), synchronized environmental parameters (temperature, NH3, TTIs, etc.), and real RSL labels (calibrated via histamine detection). 800 sets were used as the training set, 100 sets as the validation set, and 100 sets as the test set. Comparison method 1: Single CNN image classification (without environmental fusion), outputting a qualitative result of "fresh / spoiled"; Comparison method 2: LSTM + environmental sensor (without image fusion), outputting predicted RSL values.

[0133] The model described in this invention (improved CNN+BiLSTM-KAN), comparison method 1, and comparison method 2 were trained on the training set, and the hyperparameters were adjusted to minimize the loss on the validation set. During training, the improved CNN had a learning rate of 0.001, the BiLSTM had 128 hidden layer neurons, and the KAN network had 32 nodes; in the traditional CNN+LSTM, the CNN used the standard VGG architecture, and the LSTM had a learning rate of 0.01. The three trained models were then run on the test set to obtain their respective RSL prediction values.

[0134] Table 1 Simulation Results

[0135]

[0136] As shown in Table 1, the method described in this invention, through multimodal fusion, achieves a significantly lower RSL prediction MAE (1.2h) than the comparative method 2 (2.8h), and can output quantitative RSL, solving the decision-making dilemma of "no quantitative data" in the comparative method 1. Simultaneously, this invention achieves real-time linkage between "prediction-spatial mapping-early warning" through a digital twin module, compressing the response time from minutes in traditional methods to real-time, adapting to the rapid decision-making needs of the supply chain. Furthermore, in temperature fluctuation scenarios (such as ±5℃ changes during transportation), the MAE fluctuation of the solution described in this invention is <0.2h, while the comparative method 2, due to the lack of image feature fusion, cannot capture appearance degradation compensation, resulting in MAE fluctuations exceeding 1.0h; the misjudgment rate of the comparative method 1 soars by 30%, verifying the anti-interference advantage of the multimodal fusion described in this invention.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for visually predicting and warning of quality deterioration in the supply chain of non-live aquatic products, characterized in that: Includes the following steps: S1. Multi-source heterogeneous data acquisition and standardization processing: Simultaneously collect data on the physical appearance, quality indicator labels, and environmental parameters of non-live aquatic products to construct a multi-dimensional dataset; S2, Cross-modal feature fusion and key indicator extraction: Extract complementary features from physical appearance, environment and microbial indicators, strengthen key information through attention mechanism, and output weighted fusion features; S3. Construction of a quality deterioration prediction model based on spatiotemporal series: Based on weighted fusion features, combined with long-term and short-term time series dependencies and nonlinear mapping, a remaining shelf life prediction model is established, and real-time data is input into the remaining shelf life prediction model to obtain the remaining shelf life prediction value and confidence interval. S4. Building a digital twin-driven visualization early warning system: Based on the remaining shelf life forecast, construct a digital twin of the supply chain, map the physical entity status in real time, and realize early warning of deterioration risk through multi-dimensional visualization; S5. Model Adaptive Iteration and Robust Optimization: Continuously optimize the parameters of the remaining shelf life prediction model based on real-time collected multi-dimensional data.

2. The method for visual early warning of quality deterioration in the supply chain of non-live aquatic products according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11. Image acquisition and enhancement of the physical appearance of non-live aquatic products; S111. Determine key time points in the supply chain of non-live aquatic products, and use industrial cameras to collect color images of non-live aquatic products at fixed shooting distances and light intensities. S112. Image Preprocessing: The acquired images are cropped, brightness adjusted, and Gaussian filtered sequentially to obtain a standardized image set I = {I1, I2, ..., I...} n },I1,I2,...,I n These represent the 1st, 2nd, ..., nth acquired images within the standardized image set; Among them, brightness adjustment is based on the CIELab color space: In the formula, Δa * Δb * and ΔL * These represent the adjustment amounts for red-green difference, yellow-blue difference, and brightness difference, respectively; ΔE * This represents the total color difference before and after image brightness adjustment in the CIELab color space; S12. Collection of environmental and microbiological parameters; S121. Deploy a smart sensor array: Deploy temperature sensors, humidity sensors, NH3 sensors, CO2 sensors, and time-temperature indicators; S122. Samples are taken every 24 hours to test total bacterial count (CFU) and histamine concentration. The assay was performed using an enzyme-linked immunosorbent assay (ELISA), with a detection limit of 0.1 ppm. Synchronously acquire time-temperature indicator data and calculate the cumulative temperature effect (TTIs) based on the Arrhenius equation: In the formula, t total Indicates the total monitoring duration; k r (T absolute (τ) represents the absolute temperature T at time τ. absolute Below, the rate constant of the deterioration reaction of non-living aquatic products; k r (T j,absolute () represents the absolute temperature T during the j-th data collection. j,absolute The rate constant of the deterioration reaction of non-living aquatic products; Δτ j T represents the absolute temperature of the j-th data collection. j,absolute The time interval; n represents the total number of data collections; in, k r =A p ×exp(-E a / (R g T absolute )); In the formula, k r A represents the rate constant for the deterioration reaction of non-living aquatic products; p E represents the pre-exponential factor; a R represents the activation energy. g T represents the gas constant; absolute Indicates absolute temperature; S123. Data Standardization: Environmental parameters are normalized using the min-max method, and a standardized parameter matrix is ​​output. In the formula, T environment and H H These represent ambient temperature and ambient humidity, respectively. and These represent the concentrations of ammonia, carbon dioxide, and histamine, respectively; CFU represents the total bacterial count.

3. The method for visual early warning of quality deterioration in the supply chain of non-live aquatic products according to claim 2, characterized in that: In step S111, key time nodes include within 1 hour after fishing, every 6 hours during transportation, and every 12 hours during storage; The fixed shooting distance is 50m and the light intensity is 5000 lux.

4. The method for visual early warning of quality deterioration in the supply chain of non-live aquatic products according to claim 2, characterized in that: Step S2 specifically includes the following steps: S21. Image depth feature extraction; S211. Construct an improved CNN model. The improved CNN model includes convolutional layers, pooling layers, and two fully connected layers arranged sequentially. The convolutional layers consist of four 3×3 convolutional kernels with 64, 128, 256, and 512 channels respectively, using the ReLU activation function. The pooling layers use 2×2 max pooling with a stride of 2. The two fully connected layers have 512 and 256 neurons respectively, making the output image feature vector F... img =CNN(I), and F img ∈R 256 R represents the set of real numbers; S212. Introducing supplementary color features: Calculating the red shift D of the image. R : D R =R e / (R e +G+B); In the formula, R e G and B represent the pixel mean values ​​of the red, green, and blue channels of the image, respectively. S213, Move the red color to D R With F img The concatenated enhanced image feature vector F' is obtained by concatenating the features. img ∈R 257 ; S22. Environmental and Microbial Feature Mapping: A Kolmogorov-Arnold network is used to process the normalized parameter matrix P. Nonlinear mapping is then used to capture the coupling relationship between temperature, humidity, and microbial growth, yielding the environmental and microbial feature vector F. env =KAN(P), and F env ∈R 64 ; S23. Cross-modal attention fusion: Construct a multi-head attention mechanism and calculate F'. img With F env Association weights: In the formula, Q, K, and V represent the query key, matching key, and value, respectively, and Q = F' img K = V = F env ;T represents the transpose operation;d k Denotes the feature dimension, and d k =64; S24, Output weighted fusion feature F fusion (I,P)∈R 128 .

5. The method for visual early warning of quality deterioration in the supply chain of non-live aquatic products according to claim 4, characterized in that: Step S3 specifically includes the following steps: S31. Using histamine concentration ≥50ppm as the deterioration endpoint, determine the total shelf life (T) using chemical and biological methods. total Given the remaining shelf life (RSL), obtain the label dataset, where RSL = T. total -T current , T current This indicates the time corresponding to the current freshness level; S32. Divide the label dataset into training and test sets in a 7:3 ratio, and use time series cross-validation to avoid data leakage; S33. Construct a remaining shelf life prediction model. The remaining shelf life prediction model includes an input layer, two BiLSTM layers, a KAN layer, and an output layer set in sequence. S34. Set the training strategy: Use mean squared error as the loss function and the Adam optimizer. S35. Input the training set into the remaining shelf life prediction model and train it using the set training strategy until convergence. S36. Input the test set into the trained remaining shelf life prediction model, and evaluate the model using Kendall rank correlation coefficient and mean absolute error. S37. Input the real-time collected multidimensional data into the remaining shelf life prediction model to predict the remaining shelf life: In the formula, Let f(·) represent the predicted remaining shelf life; f(·) represent the remaining shelf life prediction model; and CI represent the confidence interval. z α / 2 This represents the quantile corresponding to the confidence level. This represents the set of results from M predictions. Let represent the result of the i-th prediction, and σ represent the standard deviation.

6. The method for visual early warning of quality deterioration in the supply chain of non-live aquatic products according to claim 5, characterized in that: Step S4 specifically includes the following steps: S41, Supply Chain Digital Twin Modeling; S411. Construct a virtual scene based on Unity 3D, and set it to receive multi-dimensional data and remaining shelf life prediction values ​​every 30 minutes, and update the virtual entity attributes; S412. Visualization of Appearance Deterioration: Based on Image Feature Vector F img Simulate color changes in non-live aquatic products in a virtual environment; S42, Multi-level early warning: Early warning levels are divided by setting early warning thresholds.

7. The method for visual early warning of quality deterioration in the supply chain of non-live aquatic products according to claim 6, characterized in that: In step S412, when At that time, non-live aquatic products were determined to be fresh; when At that time, non-live aquatic products were determined to be less fresh; when At that time, non-live aquatic products were judged to have deteriorated; In step S42, when At that time, it was classified as a Level 1 warning; when At that time, it was classified as a Level II warning; when At that time, it was divided into three levels of early warning.

Images