Intelligent fresh-keeping management method for fresh rice dumpling leaves based on warehouse data information

Abstract

The application discloses a fresh rice dumpling leaf product intelligent preservation management method based on warehouse data information and relates to the technical field of intelligent regulation and control of agricultural products in warehouses, which comprises the following steps: collecting temperature time series data, humidity time series data, oxygen concentration, carbon dioxide concentration and ethylene concentration time series data and product image information in a fresh rice dumpling leaf warehouse unit in real time; extracting leaf color distribution, leaf vein browning index and leaf dehydration area proportion to form a quality characteristic vector through visual quality analysis; adopting an improved LSTM neural network, dynamically adjusting a forgetting gate mechanism according to the nonlinear relationship between gas concentration and leaf vein browning rate, fusing environmental data and the quality characteristic vector, predicting quality deterioration trajectories and remaining shelf life, and generating temperature and humidity regulation, modified atmosphere gas replacement and a warehouse-out priority sequence. The method can be matched with the nonlinear change characteristics of fresh rice dumpling leaves, improve the quality prediction accuracy, realize the dynamic matching of preservation strategies and product states, and improve the fine degree of warehouse preservation control.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for agricultural product storage, specifically a method for intelligent preservation management of fresh zongzi leaves based on storage data information. Background Technology

[0002] Current methods for storing and preserving fresh bamboo leaves for zongzi (sticky rice dumplings) mostly employ constant temperature and humidity control and conventional controlled atmosphere storage. These methods only collect environmental data such as temperature, humidity, oxygen, carbon dioxide, and ethylene concentrations, relying on experience or traditional models to determine the preservation status. Some solutions only use images for simple visual recognition. Environmental data and visual inspection operate independently and are not linked to a unified analytical system.

[0003] The standard LSTM model uses fixed weights for the forgetting gate, which cannot match the nonlinear changes in gas concentration and leaf vein browning and leaf dehydration during the storage of fresh bamboo leaves. Environmental data and product quality characteristics cannot be learned collaboratively, resulting in significant biases in the prediction of quality degradation trajectories and low accuracy in determining the remaining shelf life. Preservation strategies are mostly based on manually preset fixed parameters; temperature and humidity regulation, controlled atmosphere replacement, and outbound sorting cannot be dynamically adapted to real-time quality changes.

[0004] Since it is impossible to achieve deep integration of environmental time-series data and visual quality characteristics, and the fixed forgetting gate is difficult to match the quality deterioration law driven by gas concentration, it is necessary to optimize the model mechanism and integrate multi-dimensional data to accurately predict the quality deterioration trajectory and remaining shelf life, and form a dynamically matched storage and preservation control scheme. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] Therefore, this invention proposes an intelligent preservation management method for fresh zongzi leaves based on warehousing data information, including:

[0007] Real-time data collection of preservation environment information within the storage unit of fresh zongzi leaves. The preservation environment data information includes time-series data of temperature, humidity, gas concentration, and product image information, wherein the gas concentration includes oxygen concentration, carbon dioxide concentration, and ethylene concentration.

[0008] Visual quality analysis is performed on the product image information to generate a real-time quality feature vector of fresh zongzi leaves. The real-time quality feature vector includes leaf color distribution, leaf vein browning index and leaf dehydration area ratio.

[0009] An improved LSTM neural network model is invoked to fuse and learn the preservation environment data information and the real-time quality feature vector to predict the quality deterioration trajectory and remaining shelf life of fresh zongzi leaf products. The improved LSTM neural network model dynamically adjusts its forgetting gate mechanism based on the nonlinear relationship between the gas concentration time series data and the leaf browning rate.

[0010] Based on the quality degradation trajectory and remaining shelf life, a preservation strategy is generated for the storage unit. The preservation strategy includes a temperature setting curve, humidity adjustment instructions, modified atmosphere gas replacement scheme, and outbound priority sequence.

[0011] Furthermore, visual quality analysis is performed on the product image information to generate a real-time quality feature vector for the fresh bamboo leaf product, including:

[0012] The collected product image information is preprocessed, including background segmentation, image enhancement and size normalization, to segment out a clear target area of ​​fresh bamboo leaves;

[0013] Color features are extracted from the preprocessed image of the target area of ​​fresh bamboo leaves, including the mean and variance in multiple color spaces, and the histogram distribution of each color component is statistically analyzed to calculate the color distribution of the leaves.

[0014] The vein enhancement process is performed on the target area image of the fresh zongzi leaf to segment the vein structure. The contrast between the average gray value of the vein area and the average gray value of the non-vein area is calculated. Combined with the proportion of brown pixels in the vein area, the browning index of the vein is calculated.

[0015] By analyzing the grayscale gradient changes of edge pixels in the target area of ​​fresh zongzi leaves, the curled or withered areas on the leaf edges are detected, and the ratio of the area of ​​the curled or withered areas to the total area of ​​the leaf is calculated to obtain the dehydrated area ratio of the leaf.

[0016] The leaf color distribution, leaf vein browning index, and leaf dehydration area ratio are normalized and vectorized, and combined to form the real-time quality feature vector.

[0017] Furthermore, the preservation environment data and the real-time quality feature vector are fused and learned to predict the quality deterioration trajectory and remaining shelf life of fresh zongzi leaves, including:

[0018] The time-series data of temperature, humidity, and gas concentration that are aligned in time are concatenated with the real-time quality feature vector to form a multi-dimensional fused time series.

[0019] The multi-dimensional fused time series is divided into multiple consecutive input segments according to a fixed time window;

[0020] Each input segment is sequentially fed into the improved LSTM neural network model, and the improved LSTM neural network model sequentially outputs the quality feature prediction vector corresponding to each input segment.

[0021] The true quality feature vector of a historical time period is compared with the predicted quality feature vector of the corresponding time period, the prediction error is calculated, and the prediction error is used to backpropagate and train the parameters of the improved LSTM neural network model.

[0022] During the prediction phase, the latest input fragment is fed into the trained model, and the model scrolls forward to predict the quality feature prediction vectors for multiple future time steps.

[0023] By connecting historical real values ​​with future predicted values, a complete quality degradation trajectory is formed.

[0024] Based on the time point at which the quality deterioration trajectory reaches the preset quality deterioration threshold, the time span from the current moment to the time point is calculated as the remaining shelf life.

[0025] Furthermore, the improved LSTM neural network model dynamically adjusts its forgetting gate mechanism based on the nonlinear relationship between the gas concentration time-series data and the leaf browning rate, including:

[0026] In the improved LSTM neural network model, a concentration influence weight is assigned to the ethylene concentration and carbon dioxide concentration in the gas concentration time series data, respectively;

[0027] In the forgetting gate structure of the model, the forgetting gate activation calculation at each time step is decomposed into a basic forgetting part and a concentration effect part;

[0028] The basic forgetting part is calculated based on the hidden state of the previous moment and the currently input temperature time series data and humidity time series data;

[0029] The concentration influence component generates a concentration adjustment factor by multiplying the ethylene concentration value and carbon dioxide concentration value at the current time step by their corresponding concentration influence weights, summing them, and then passing the sum through a nonlinear transformation function.

[0030] The concentration adjustment factor is weighted and fused with the basic forgetting component to form the final forgetting gate output of the time step;

[0031] This mechanism allows the model to dynamically adjust the degree of forgetting of historical quality states when there are abnormal changes in ethylene or carbon dioxide concentrations, thus responding more sensitively to sudden changes in browning rate caused by changes in the gas environment.

[0032] Furthermore, based on the quality degradation trajectory and remaining shelf life, a refined preservation strategy is generated for the storage unit, including:

[0033] Analyze the quality degradation trajectory to identify the key time points and rates of quality decline;

[0034] Based on the key time points and the rate of decline, the intensity of the physiological and biochemical reactions to be suppressed can be deduced, thereby deriving the target temperature range and the target humidity range.

[0035] Based on the derived target temperature range and target humidity range, and combined with the actual temperature and humidity of the current storage unit, the time-segmented temperature setting curve and humidity adjustment command are generated.

[0036] Based on the trend of the gas concentration time series data, especially the cumulative trend of ethylene concentration, the gas replacement volume and replacement frequency required to delay quality deterioration are calculated, and the modified atmosphere gas replacement scheme is generated.

[0037] Based on the remaining shelf life, the fresh bamboo leaf products of different batches or locations in the storage unit are sorted, and products with shorter remaining shelf life are assigned higher outbound priority, thus generating the outbound priority sequence.

[0038] The temperature setting curve, humidity adjustment command, modified atmosphere gas replacement scheme, and outbound priority sequence are integrated into a unified refined preservation strategy command set.

[0039] Furthermore, the generation of the time-segmented temperature setting curve and the humidity adjustment command includes:

[0040] Extract the predicted rate of change of leaf dehydration area ratio and the rate of change of leaf vein browning index for multiple consecutive time periods in the future from the quality deterioration trajectory.

[0041] Based on the correlation between the rate of change of the leaf dehydration area ratio and the intensity of transpiration, a transpiration inhibition model was established, with the target environmental humidity as the main control variable.

[0042] Based on the correlation between the rate of change of leaf vein browning index and polyphenol oxidase activity, an enzyme activity inhibition model was established, with the target ambient temperature as the main regulating variable.

[0043] By jointly solving the transpiration inhibition model and the enzyme activity inhibition model, the combination of target temperature and target humidity that reduces the predicted quality degradation rate to below the preset rate is obtained.

[0044] Multiple target temperatures and target humidity combinations corresponding to consecutive future time periods are converted into temperature setpoint sequences and humidity setpoint sequences that can be executed by the control device, thus forming the temperature setpoint curve and the humidity adjustment command.

[0045] Furthermore, the browning index of the leaf veins is calculated by combining the proportion of brown pixels in the leaf vein region, including:

[0046] Determine the color threshold range used to define brown pixels. The color threshold range is defined in the RGB color space as R component being greater than a first preset threshold, G component being less than a second preset threshold, and B component being less than a fourth preset threshold, or in the HSV color space as hue being within a preset brown hue range.

[0047] For the leaf vein region image after leaf vein enhancement and segmentation processing, determine whether the color value of each pixel falls within the color threshold range of the brown pixel, and count the total number of pixels that are judged to be brown.

[0048] The ratio of the total number of brown pixels to the total number of pixels in the leaf vein region image is calculated as the proportion of brown pixels in the leaf vein region.

[0049] The average gray value of the leaf vein region image and the average gray value of the non-leaf vein region image are obtained, and the ratio or difference between the average gray value of the leaf vein region and the average gray value of the non-leaf vein region is calculated as the contrast of the leaf vein region.

[0050] The brown pixel ratio and the contrast of the leaf vein region are weighted and summed, with the brown pixel ratio assigned a higher weight coefficient. The weighted sum is then normalized and mapped to a preset index range to obtain the final leaf vein browning index.

[0051] Furthermore, the method also includes:

[0052] The storage equipment is driven to execute the preservation strategy, and the prediction results of the improved LSTM neural network model are continuously verified and the parameters are iteratively updated based on the subsequently collected preservation environment data and real-time quality feature vectors.

[0053] The drive storage equipment executes the preservation strategy, including:

[0054] The preservation strategy instruction set is parsed into specific equipment control instructions, which include refrigeration unit start / stop instructions, humidifier power adjustment instructions, controlled atmosphere storage gas valve opening / closing instructions, and conveyor line scheduling instructions.

[0055] The start / stop command of the refrigeration unit and the power adjustment command of the humidifier are sent to the environmental control system to control the temperature and humidity in the storage unit to be adjusted according to the temperature setting curve and the humidity adjustment command.

[0056] The control atmosphere storage gas valve opening and closing commands are sent to the gas control system to control the operation of the nitrogen generator, carbon dioxide removal machine and ethylene removal machine, and to execute the control atmosphere gas replacement scheme.

[0057] The conveyor line scheduling instruction is sent to the warehousing and logistics execution system. Based on the outbound priority sequence, the handling equipment is planned and scheduled to reserve outbound channels for high-priority products or move them to the pre-outbound area.

[0058] Furthermore, the step of performing rolling verification and parameter iterative updates on the prediction results of the improved LSTM neural network model based on subsequently collected preservation environment data and real-time quality feature vectors includes:

[0059] After implementing the preservation strategy, new preservation environment data and new product image information are collected again at the preset feedback cycle point.

[0060] Perform the same visual quality analysis on the new product image information to generate a new real-time quality feature vector.

[0061] The preservation environment data and real-time quality feature vector from the previous time are reassembled with the new preservation environment data and new real-time quality feature vector to form a verification dataset.

[0062] The validation dataset is input into the improved LSTM neural network model currently in use to obtain simulated predictions of the quality feature vectors for the current and past time points.

[0063] Calculate the error loss between the simulated predicted value and the true value in the validation dataset;

[0064] When the error loss exceeds a preset threshold, the improved LSTM neural network model is incrementally trained using the validation dataset to update its model parameters and generate a model version with iteratively updated parameters.

[0065] Further, the incremental training of the improved LSTM neural network model using the validation dataset includes:

[0066] The weight parameters of most network layers, except the output layer, are fixed in the improved LSTM neural network model.

[0067] Using the aforementioned validation dataset, the gradient calculation and update are performed primarily on the weight parameters of the model output layer and the concentration influence weights via the backpropagation algorithm.

[0068] An optimizer with a small learning rate is used to perform parameter updates to prevent catastrophic forgetting of previously learned knowledge;

[0069] After the update is completed, the generalization performance of the model with iteratively updated parameters is evaluated using an independent test dataset. The new model is used to replace the original model only if the generalization performance is not lower than that of the original model.

[0070] Compared with the prior art, the beneficial effects of the present invention are:

[0071] Based on the nonlinear correlation between time-series data of oxygen, carbon dioxide, and ethylene concentrations and leaf vein browning rate, the weights of the LSTM forgetting gate are dynamically adjusted. The forgetting gate can adapt to feature selection rules in real time according to changes in gas concentration, accurately extracting effective information related to quality deterioration from time-series data and weakening the weight influence of redundant data segments. The model's feature fitting ability for the nonlinear quality deterioration process of fresh zongzi leaves is enhanced. The time-series feature extraction closely matches the real changes in browning and dehydration driven by gas concentration, reducing the feature fitting bias caused by the traditional fixed forgetting gate, strengthening the accuracy of capturing key nodes and trends of quality deterioration, and improving the reliability of predicting quality deterioration trajectory and remaining shelf life.

[0072] By synchronously fusing and learning time-series environmental data on temperature, humidity, and gas concentration with visual feature vectors such as leaf color distribution, leaf vein browning index, and leaf dehydration area ratio, a complete feature system is formed, integrating external environmental parameters with the product's intrinsic quality characteristics. This overcomes the incomplete representation caused by single data dimensions. Multimodal feature fusion can fully reflect the linkage between the storage environment and product status, refine the continuous changes in the quality deterioration trajectory, and improve the accuracy and continuity of prediction results. The temperature setpoint curve, humidity adjustment command, modified atmosphere gas replacement scheme, and outbound priority sequence generated based on the fused prediction results can be matched with the real-time quality status of fresh bamboo leaves. Storage control parameters are dynamically adjusted according to the quality deterioration trajectory, and preservation control actions are synchronized with the product deterioration rhythm, achieving refined adaptation of the storage environment and preservation strategy. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating the steps of the intelligent preservation management method for fresh bamboo leaf products based on warehouse data information as described in this invention.

[0074] Figure 2 A flowchart for generating real-time quality feature vectors for visual quality analysis;

[0075] Figure 3 A flowchart for dynamically adjusting the forget gate mechanism in an improved LSTM neural network model. Detailed Implementation

[0076] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] See Figure 1 This invention provides an intelligent preservation management method for fresh bamboo leaves based on warehousing data information. The specific method includes:

[0078] Real-time data collection of the preservation environment within the fresh zongzi leaf storage unit is performed. This data includes time-series temperature, humidity, and gas concentration data, as well as product image information. The gas concentrations include oxygen, carbon dioxide, and ethylene concentrations. Visual quality analysis is conducted on the product image information to generate a real-time quality feature vector for the fresh zongzi leaves. This vector includes leaf color distribution, leaf vein browning index, and leaf dehydration area ratio. An improved LSTM neural network model is then used to fuse and learn the preservation environment data and the real-time quality feature vector to predict the quality degradation trajectory and remaining shelf life of the fresh zongzi leaves. The improved LSTM neural network model dynamically adjusts its forgetting gate mechanism based on the nonlinear relationship between the gas concentration time-series data and the leaf browning rate. Based on the quality degradation trajectory and remaining shelf life, a preservation strategy is generated for the storage unit. This strategy includes a temperature setpoint curve, humidity adjustment instructions, a controlled atmosphere gas replacement scheme, and an outbound priority sequence.

[0079] In one embodiment of the present invention, see [reference] Figure 2The collected product image information is preprocessed, including background segmentation, image enhancement, and size normalization, to segment a clear target area of ​​fresh bamboo leaves. Color features are extracted from the preprocessed image of the target area of ​​fresh bamboo leaves, including the mean and variance in multiple color spaces, and the histogram distribution of each color component is calculated to obtain the leaf color distribution. Vein enhancement processing is performed on the image of the target area of ​​fresh bamboo leaves to segment the vein structure, and the contrast between the average gray value of the vein area and the average gray value of the non-vein area is calculated. A color threshold range for defining brown pixels is determined. In the RGB color space, the R component is greater than a first preset threshold, the G component is less than a second preset threshold, and the B component is less than a fourth preset threshold; or in the HSV color space, the hue is within a preset brown hue range. For the image of the vein area after vein enhancement and segmentation, each pixel is judged to see if its color value falls within the color threshold range of the brown pixels, and the total number of pixels judged as brown is counted. The ratio of the total number of brown pixels to the total number of pixels in the leaf vein region image is calculated as the brown pixel proportion in the leaf vein region. The average gray value of the leaf vein region image and the average gray value of the non-leaf vein region image are obtained, and the ratio or difference between the average gray value of the leaf vein region and the average gray value of the non-leaf vein region is calculated as the leaf vein region contrast. The brown pixel proportion and the leaf vein region contrast are weighted and summed, with the brown pixel proportion assigned a higher weight coefficient. The weighted sum is then normalized and mapped to a preset index range to obtain the final leaf vein browning index. By analyzing the edge pixel gray-scale gradient changes in the target area of ​​fresh bamboo leaves, curled or dried areas at the leaf edges are detected, and the ratio of the area of ​​the curled or dried areas to the total leaf area is calculated to obtain the leaf dehydration area proportion. The leaf color distribution, leaf vein browning index, and leaf dehydration area proportion are normalized and vectorized, and combined to form the real-time quality feature vector.

[0080] In practice, preprocessing is performed on the collected images of fresh bamboo leaves. The preprocessing steps include background segmentation using a color feature-based region growing method to remove background interference from the storage environment, image enhancement using histogram equalization to improve the recognizability of leaf texture, and scaling the image to a fixed pixel size for size normalization, thereby segmenting a clear target area for the fresh bamboo leaves. In the specific implementation, color features are extracted from the preprocessed image of the target area of ​​the fresh bamboo leaves. The mean and variance of each channel are calculated in RGB, HSV, and Lab color spaces, and the histogram distributions of the R, G, and B components are statistically analyzed. The histograms are then smoothed to eliminate noise. The above-mentioned mean, variance, and histogram distribution statistics are combined to form the leaf color distribution.

[0081] In the specific implementation, Gaussian difference filtering is applied to the target area image of fresh bamboo leaves for vein enhancement processing. Threshold segmentation is used to separate the vein structure. The difference between the average gray value of the vein area and the average gray value of the non-vein area is calculated as the contrast of the vein area. A color threshold range for defining brown pixels is defined. In the RGB color space, the range where the R component is greater than 150, the G component is less than 100, and the B component is less than 80 is set as the brown range. In the HSV color space, the range where the hue is between 15° and 30° is set as the brown tone range. For the vein area image after vein enhancement and segmentation processing, the RGB value and HSV value of each pixel are judged to see if they fall within the above-mentioned color threshold range of brown pixels. The total number of pixels judged as brown is counted. The ratio of the total number of brown pixels to the total number of pixels in the vein area image is calculated as the proportion of brown pixels in the vein area.

[0082] In practice, the proportion of brown pixels and the contrast of the leaf vein area are substituted into the following formula for weighted summation:

[0083]

[0084] in: Indicates the leaf vein browning index. This indicates the percentage of brown pixels in the leaf vein area. Indicates the contrast of the leaf vein area. and These are the weighting coefficients for the proportion of brown pixels and the contrast of the leaf vein area, respectively, and they satisfy... The weighted summation result is linearly mapped to the exponential range of [0,10] to obtain the final leaf vein browning index.

[0085] In practical implementation, the Sobel operator is used to analyze the grayscale gradient changes of edge pixels in the target area image of fresh zongzi leaves. A gradient amplitude threshold is set to detect curled or withered areas at the leaf edges. The area of ​​the closed region enclosed by edge pixels with gradient amplitudes exceeding the threshold is counted, and the ratio of this region area to the total leaf area is calculated to obtain the dehydrated area ratio of the leaf. In practical implementation, the leaf color distribution, leaf vein browning index, and leaf dehydrated area ratio are normalized to their maximum and minimum values ​​so that each feature value falls within the [0,1] interval. The normalized feature values ​​are then concatenated in sequence into a one-dimensional array to form the real-time quality feature vector of the fresh zongzi leaf product.

[0086] In one embodiment of the present invention, the time-aligned temperature time-series data, humidity time-series data, and gas concentration time-series data are concatenated with the real-time quality feature vector to form a multi-dimensional fused time series. The multi-dimensional fused time series is divided into multiple consecutive input segments according to a fixed time window. Each input segment is sequentially input into the improved LSTM neural network model. See also... Figure 3 In the improved LSTM neural network model, a concentration influence weight is assigned to the ethylene concentration and carbon dioxide concentration in the gas concentration time series data. In the model's forget gate structure, the forget gate activation calculation at each time step is decomposed into a basic forgetting part and a concentration influence part. The basic forgetting part is calculated based on the hidden state of the previous time step and the current input temperature and humidity time series data. The concentration influence part generates a concentration adjustment factor by multiplying the current ethylene concentration value and carbon dioxide concentration value by their corresponding concentration influence weights, summing them, and then applying a nonlinear transformation function. The concentration adjustment factor is weighted and fused with the basic forgetting part as the final forget gate output for the time step. This mechanism allows the model to dynamically adjust the degree of forgetting of historical quality states when ethylene or carbon dioxide concentrations change abnormally. The improved LSTM neural network model sequentially outputs the quality feature prediction vector corresponding to each input segment. The true quality feature vector of the historical time period is compared with the quality feature prediction vector of the corresponding time period to calculate the prediction error, and the prediction error is used to backpropagate and train the parameters of the improved LSTM neural network model. During the prediction phase, the latest input segment is fed into the trained model, which then advances to predict quality feature vectors for multiple future time steps. The historical true values ​​and future predicted values ​​are connected to form a complete quality degradation trajectory. Based on the time point at which the quality degradation trajectory reaches a preset quality degradation threshold, the time span from the current moment to that threshold is calculated and used as the remaining shelf life.

[0087] In practice, the time-series data of temperature, humidity, and gas concentration collected by the sensors are time-aligned with the real-time quality feature vector obtained from image analysis. The multi-dimensional data are then concatenated column by column using the same timestamp as a reference to form a multi-dimensional fusion time series that covers both environment and quality. A fixed time window length of 24 hours is set, with each sampling point being 5 minutes. The continuous multi-dimensional fusion time series is divided into 288 time-step input segments, each containing all environmental monitoring values ​​and quality feature vectors within that time period.

[0088] In the specific implementation, an improved LSTM neural network model is constructed. The input layer dimension of the model is consistent with the feature dimension of the multi-dimensional fusion time series, and the number of hidden layer units is set to 128. In the forget gate design of the improved LSTM neural network model, concentration influence weights are assigned to ethylene concentration and carbon dioxide concentration respectively, with the weight of ethylene concentration set to 0.7 and the weight of carbon dioxide concentration set to 0.3. In the forget gate calculation at each time step, the basic forget part is based on the hidden state of the previous time step and the temperature and humidity time series data of the current time step, and the initial forget value is calculated by using the sigmoid activation function. The concentration influence part takes the ethylene concentration value and carbon dioxide concentration value of the current time step, multiplies them by the corresponding concentration influence weights respectively, sums them, and then performs a nonlinear transformation through the tanh function to generate a concentration adjustment factor in the range of [-1,1].

[0089] In practice, the concentration adjustment factor and the basal forgetting component are weighted and fused as follows:

[0090]

[0091] in: The final forget gate output represents the current time step. Represents the basic forgetting part. Represents the concentration adjustment factor. The balance coefficient is set to a fixed value of 0.6. This fusion process causes the final forget gate output to change dynamically with the fluctuation of gas concentration. When the concentration of ethylene or carbon dioxide increases significantly, the concentration adjustment factor increases, prompting the model to forget the historical state more quickly in response to the sudden change in browning rate.

[0092] In practice, the predefined input segments are sequentially fed into the improved LSTM neural network model. At each time step, the model outputs a corresponding quality feature prediction vector, including the predicted leaf color distribution, leaf vein browning index, and leaf dehydration area ratio. During the training phase, historical data from the past 30 days is selected as the training set. The quality feature prediction vector output by the model is compared with the actual quality feature vectors of the same time period. The mean squared error is used as the loss function, and backpropagation training is performed through the Adam optimizer to iteratively update the weight parameters of the improved LSTM neural network model.

[0093] In the specific implementation, during the prediction phase, the latest input segment from the 24 hours prior to the current moment is input into the trained improved LSTM neural network model. The model outputs the predicted value every 5 minutes for the next 72 hours in a rolling prediction manner. The historical true quality feature vectors of the past 72 hours are connected with the predicted values ​​for the next 72 hours along the time axis to form a complete quality deterioration trajectory spanning 144 hours. The leaf vein browning index is set to reach 6.0 as the quality deterioration threshold. The time point when the leaf vein browning index first exceeds the threshold is found in the quality deterioration trajectory, and the time from the current moment to that time point is calculated as the remaining shelf life.

[0094] In some embodiments, the fixed time window length can be adjusted to 12 hours or 48 hours according to actual storage needs, and the time step can also be set to 1 minute or 10 minutes to adapt to different data collection frequencies. In other embodiments, the concentration influence weights of ethylene concentration and carbon dioxide concentration can be optimized on different datasets through grid search, and the balance coefficient α can be selected between 0.5 and 0.8. Optionally, the division of multi-dimensional fusion time series can also adopt a sliding window method with a window step of 1 time step to increase the number of training samples. Optionally, in addition to the leaf vein browning index, the quality deterioration threshold can also be combined with the proportion of leaf dehydration area exceeding 20% ​​as a comprehensive judgment condition. It is understood that the number of hidden layer units of the improved LSTM neural network model can be expanded to 256 or 512 according to the data scale to adapt to more complex nonlinear relationships. It is understood that the calculation of the remaining shelf life can adopt a multi-feature weighted scoring method in addition to a single threshold judgment, and the shelf life end is determined when the comprehensive score is lower than the preset qualified line.

[0095] In one embodiment of the present invention, the quality degradation trajectory is analyzed to identify key time points and rates of quality decline. Based on these key time points and rates of decline, the required intensity of the physiological and biochemical reactions to be suppressed is deduced, thereby deriving the target temperature and humidity ranges. The predicted rate of change of leaf dehydrated area ratio and the rate of change of leaf vein browning index over multiple consecutive time periods are extracted from the quality degradation trajectory. Based on the correlation between the rate of change of leaf dehydrated area ratio and transpiration intensity, a transpiration inhibition model is established, with target environmental humidity as the primary control variable. Based on the correlation between the rate of change of leaf vein browning index and polyphenol oxidase activity, an enzyme activity inhibition model is established, with target environmental temperature as the primary control variable. The transpiration inhibition model and the enzyme activity inhibition model are jointly solved to calculate a combination of target temperature and target humidity that reduces the predicted quality degradation rate to below a preset rate. Multiple target temperature and target humidity combinations corresponding to consecutive future time periods are converted into temperature setpoint sequences and humidity setpoint sequences executable by the control device, respectively, forming the temperature setpoint curve and the humidity adjustment command. Based on the derived target temperature and humidity ranges, and combined with the actual temperature and humidity of the current storage unit, time-segmented temperature setting curves and humidity adjustment commands are generated. According to the trend of the gas concentration time-series data, especially the cumulative trend of ethylene concentration, the required gas replacement volume and frequency to delay quality deterioration are calculated, generating the controlled atmosphere gas replacement scheme. Based on the remaining shelf life, fresh bamboo leaf products from different batches or locations within the storage unit are sorted, with products having shorter remaining shelf lives assigned higher outbound priority, generating the outbound priority sequence. The temperature setting curves, humidity adjustment commands, controlled atmosphere gas replacement scheme, and outbound priority sequence are integrated into a unified refined preservation strategy command set.

[0096] In practice, the quality degradation trajectory output by the improved LSTM neural network model is analyzed to identify the time point when the growth slope of the leaf vein browning index exceeds 0.05 per hour as the key time node for quality decline. The time and decline rate corresponding to this node are recorded. Based on the key time node and decline rate, the physiological and biochemical reaction intensity required to inhibit polyphenol oxidase activity and transpiration is deduced, and the target temperature range is derived as [2℃, 5℃], and the target humidity range is derived as [85%, 90%].

[0097] In practical implementation, the predicted rate of change of leaf dehydrated area ratio and the rate of change of leaf vein browning index over four consecutive six-hour periods are extracted from the quality deterioration trajectory. Based on the positive correlation between the rate of change of leaf dehydrated area ratio and transpiration intensity, a transpiration inhibition model is established with target environmental humidity as the sole regulating variable, expressed as follows: ,in This represents the decrease in transpiration intensity. Humidity influence coefficient The difference between the target ambient humidity and the baseline is used to establish an enzyme activity inhibition model with the target ambient temperature as the sole regulating variable, based on the positive correlation between the leaf vein browning index change rate and polyphenol oxidase activity. This model is expressed as:

[0098]

[0099] in: This represents the amount of enzyme activity attenuation. This is the temperature influence coefficient. This represents the difference between the target ambient temperature and the baseline.

[0100] In practice, the transpiration inhibition model and the enzyme activity inhibition model are solved together. With the constraint that the predicted quality deterioration rate drops to below 0.01 per hour, the target temperature and humidity combinations for the next four time periods are calculated. The calculation results are converted into a sequence of temperature setpoints and humidity setpoints that can be executed by the control equipment, forming temperature setpoint curves and humidity adjustment commands. Based on the derived target temperature range and target humidity range, combined with the current measured temperature of 4.2℃ and humidity of 82% in the storage unit, time-segmented temperature setpoint curves and humidity adjustment commands are generated.

[0101] In practice, based on the trend of a daily cumulative increase of 0.3 ppm in ethylene concentration in the gas concentration time series data, the gas replacement volume required to maintain the ethylene concentration below 1 ppm is calculated to be 10% of the storage capacity, with a replacement frequency of once every eight hours. A controlled atmosphere gas replacement scheme is generated. Based on the remaining shelf life, the five batches of fresh zongzi leaf products in the storage unit are sorted by the remaining days. The two batches with the shorter remaining shelf life are assigned the highest outbound priority, generating an outbound priority sequence. The temperature setting curve, humidity adjustment instructions, controlled atmosphere gas replacement scheme, and outbound priority sequence are integrated into a unified and refined preservation strategy instruction set, as shown in Table 1.

[0102] Table 1: Segmented Target Temperature and Humidity Settings for the Next 24 Hours

[0103] 0-6 4.5 88 6-12 4.0 89 12-18 3.5 87 18-24 4.0 86

[0104] In some embodiments, the identification of critical time points of quality degradation can use the mutation of the second derivative of the leaf vein browning index as a criterion, replacing the simple slope threshold judgment; in some embodiments, the derivation of the target temperature range and target humidity range can incorporate a light intensity correction factor, suitable for special storage units equipped with supplemental lighting systems. Optionally, the joint solution of the transpiration inhibition model and the enzyme activity inhibition model can use a particle swarm optimization algorithm to find the global optimum; optionally, the generation of the outbound priority sequence can take into account the product batch size factor, assigning appropriate priority to large-volume products. It is understood that the calculation of gas replacement volume and replacement frequency can be dynamically adjusted based on real-time ethylene concentration readings, rather than relying on a fixed trend; it is understood that the integrated format of the refined preservation strategy instruction set can be JSON or XML structure, which is convenient for the warehouse management system to parse directly.

[0105] In one embodiment of the present invention, the preservation strategy instruction set is parsed into specific equipment control instructions, including refrigeration unit start / stop instructions, humidifier power adjustment instructions, controlled atmosphere storage gas valve opening / closing instructions, and conveyor line scheduling instructions. The refrigeration unit start / stop instructions and the humidifier power adjustment instructions are sent to the environmental control system to control the temperature and humidity within the storage unit according to the temperature setting curve and humidity adjustment instructions. The controlled atmosphere storage gas valve opening / closing instructions are sent to the gas control system to control the operation of the nitrogen generator, carbon dioxide removal machine, and ethylene removal machine, executing the controlled atmosphere gas replacement scheme. The conveyor line scheduling instructions are sent to the warehousing logistics execution system to plan and schedule handling equipment to reserve outbound channels for high-priority products or move them to the pre-outbound area based on the outbound priority sequence. After executing the preservation strategy, at a preset feedback cycle point, new preservation environment data and new product image information are collected again. The same visual quality analysis is performed on the new product image information to generate a new real-time quality feature vector. The preservation environment data and real-time quality feature vectors from previous times are concatenated with the new preservation environment data and real-time quality feature vectors to form a validation dataset. This validation dataset is then input into the improved LSTM neural network model to obtain simulated predictions of the quality feature vectors for the current and past time points. The error loss between the simulated predictions and the actual values ​​in the validation dataset is calculated. When the error loss exceeds a preset threshold, the improved LSTM neural network model is incrementally trained using the validation dataset to update its parameters, generating a model version with iteratively updated parameters.

[0106] In practical implementation, the refined preservation strategy instruction set is parsed into specific equipment control instructions through a protocol conversion module. These include: refrigeration unit start / stop instructions containing compressor operating frequency setpoints; humidifier power adjustment instructions containing atomization percentage parameters; controlled atmosphere gas valve opening / closing instructions containing the opening duration of nitrogen injection valves and the start interval of the carbon dioxide removal machine; and conveyor line scheduling instructions containing pallet numbers and target outlet channel identifiers. The refrigeration unit start / stop instructions and humidifier power adjustment instructions are then sent to the environmental control system via the Modbus protocol, driving the refrigeration unit at 7... The system operates at 5Hz until the temperature reaches 4.5℃ of the set curve, then controls the humidifier to operate at 60% power to stabilize the humidity at 88%. The system sends the opening and closing commands of the controlled atmosphere storage gas valves to the gas control system via the CAN bus, triggering the nitrogen generator to operate continuously for 120 seconds, starting the carbon dioxide removal machine to run for 180 seconds, and simultaneously activating the ethylene removal machine for cyclic operation. The system also transmits the conveyor line scheduling command to the warehousing and logistics execution system, instructing the stacker crane to move the pallet with batch number A23 to the No. 3 pre-shipment area, reserving a fast outbound channel for the batch with the shortest remaining shelf life.

[0107] In specific implementation, after six hours of implementing the preservation strategy, a preset feedback cycle data acquisition task is triggered. New time-series temperature data, new time-series humidity data, and new time-series gas concentration data are collected again in the storage unit through a distributed sensor array. Simultaneously, a new batch of product image information is captured using a high-definition industrial camera. The newly acquired product image information is processed using the same visual quality analysis process as in the previous embodiment to generate a new real-time quality feature vector containing the color distribution of new leaves, the browning index of new leaf veins, and the proportion of dehydrated area of ​​new leaves. The original preservation environment data information and the original real-time quality feature vector from the twelve hours prior to the current moment are aligned and spliced ​​with the newly acquired preservation environment data information and the new real-time quality feature vector in chronological order to construct a verification dataset containing 216 time points.

[0108] In practice, the validation dataset is input into the currently deployed improved LSTM neural network model to obtain the model's simulated predictions of the quality feature vectors for the past six hours and the current moment. The error loss between the simulated predictions and the actual quality feature vectors in the validation dataset is calculated using the following formula:

[0109]

[0110] in: This represents the error loss value. This indicates the number of samples in the validation dataset. Indicates the first in the validation dataset The true quality feature vector of each sample The model represents the first The simulated predicted value for each sample; when the error loss value When the value exceeds the preset threshold of 0.02, the labeling model needs to update its parameters iteratively, as shown in Table 2.

[0111] Table 2: Validation Dataset Structure Table

[0112] 2023-06-0814:00 4.52 83 0.95 3.21 11.34 2023-06-0816:00 4.61 81 1.03 3.45 13.78 2023-06-0818:00 4.58 84 1.22 3.67 15.92

[0113] In some embodiments, the transmission of equipment control commands can use the EtherCAT industrial Ethernet protocol instead of the Modbus protocol to meet higher real-time control requirements. In some embodiments, the duration of the feedback cycle can be adjusted to four or eight hours according to the seasonal characteristics of warehousing to adapt to different temperature fluctuation frequencies. Optionally, the construction of the validation dataset can include historical data from the previous 24 hours to expand the time-series coverage. Optionally, the threshold for judging error loss can be dynamically configured according to product grade, with a stricter threshold of 0.015 used for high-quality products. It is understood that the acquisition of simulated prediction values ​​can skip the model training stage and directly call the inference interface to reduce computational resource consumption. It is understood that the triggering condition for parameter iteration updates can be supplemented with the judgment logic of two consecutive exceedances of the threshold to avoid erroneous updates caused by a single fluctuation.

[0114] In one embodiment of the present invention, the weight parameters of most network layers in the improved LSTM neural network model, except for the output layer, are fixed. Using the validation dataset, gradient calculations and updates are performed primarily on the weight parameters of the model's output layer and the concentration-influence weights using the backpropagation algorithm. An optimizer with a small learning rate is used to perform the parameter updates. After the updates are completed, an independent test dataset is used to evaluate the generalization performance of the model after iterative parameter updates. The new model replaces the original model only if the generalization performance is not lower than that of the original model.

[0115] In practice, when the error loss exceeds a preset threshold and triggers the model update process, the weight parameters of all convolutional layers, recurrent layers, and fully connected layers in the improved LSTM neural network model, except for the output layer, are frozen. The internal parameters of the basic feature extraction structure and the dynamic forgetting mechanism are locked unchanged, and the output layer weights and the weights affecting ethylene concentration and carbon dioxide concentration are retained as updateable objects. Using 216 samples in the validation dataset, the gradients of the output layer weights and the weights affecting concentration are calculated through the backpropagation algorithm. The gradient calculation covers the cumulative error of all samples to avoid parameter oscillations caused by single-sample fluctuations.

[0116] In practice, the Adam optimizer is used to perform parameter updates. A small learning rate of 0.0001 is set, momentum parameters β1 and β2 are set to 0.9 and 0.999 respectively, and the weight decay coefficient is set to 1e-5. The update step size is jointly adjusted by the first and second moment estimates of the gradient to constrain the magnitude of parameter changes. The update formula is as follows:

[0117]

[0118] in: This indicates the values ​​of the output layer weights or concentration influence weights before the update. This represents the updated value. Indicates the learning rate. This represents the first moment deviation correction term of the gradient. The second-order moment deviation correction term represents the gradient. This is to prevent the addition of a very small constant to prevent the denominator from being zero; the update process iterates 50 times, each time traversing all samples in the validation dataset, gradually fine-tuning the output layer weights and concentration influence weights, avoiding significant modifications that could damage the stability of the original model.

[0119] In practice, after the parameters are updated, the generalization performance of the updated model is evaluated using an independent test dataset. The independent test dataset contains 720 samples that were not used for training and validation in the past three months, covering temperature fluctuation patterns and humidity changes in different seasons. The mean squared error of the updated model's prediction on the independent test dataset is calculated and compared with the mean squared error of the original model on the same dataset. If the mean squared error of the updated model is not greater than that of the original model, the updated model version is deployed to the warehouse management system, replacing the original improved LSTM neural network model. If the mean squared error of the updated model is higher than that of the original model, the update results are discarded, and the original model continues to run.

[0120] In some embodiments, the updatable weight range can be extended to the last hidden layer of the improved LSTM neural network model, allowing adjustment of the fine structure of the feature representation. In some embodiments, the value of the small learning rate can be dynamically adjusted based on the sample size of the validation dataset, with a smaller learning rate used to enhance stability when the sample size is large. Optionally, the selection of the independent test dataset can exclude anomalous data under extreme weather conditions, containing only typical working condition samples under normal warehousing environments. Optionally, the mean absolute error can be added as an auxiliary metric for evaluating generalization performance, allowing model replacement when the mean squared error is the same but the mean absolute error is lower. It is understood that to prevent weight drift caused by multiple incremental updates, an upper limit can be set for model version iterations, requiring complete retraining after reaching the upper limit. It is understood that the update cycle of the independent test dataset can be synchronized with quarterly climate changes, with new data added at the end of each quarter to ensure the timeliness of the evaluation results.

[0121] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. 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 be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for intelligent preservation management of fresh bamboo leaves for making zongzi (sticky rice dumplings) based on warehouse data information, characterized in that: The method includes: Real-time data collection of preservation environment information within the storage unit of fresh zongzi leaves. The preservation environment data information includes time-series data of temperature, humidity, gas concentration, and product image information, wherein the gas concentration includes oxygen concentration, carbon dioxide concentration, and ethylene concentration. Visual quality analysis is performed on the product image information to generate a real-time quality feature vector of fresh zongzi leaves. The real-time quality feature vector includes leaf color distribution, leaf vein browning index and leaf dehydration area ratio. An improved LSTM neural network model is invoked to fuse and learn the preservation environment data information and the real-time quality feature vector to predict the quality deterioration trajectory and remaining shelf life of fresh zongzi leaf products. The improved LSTM neural network model dynamically adjusts its forgetting gate mechanism based on the nonlinear relationship between the gas concentration time series data and the leaf browning rate. Based on the quality degradation trajectory and remaining shelf life, a preservation strategy is generated for the storage unit. The preservation strategy includes a temperature setting curve, humidity adjustment instructions, modified atmosphere gas replacement scheme, and outbound priority sequence.

2. The intelligent preservation management method for fresh zongzi leaves based on warehousing data information according to claim 1, characterized in that, Visual quality analysis is performed on the product image information to generate a real-time quality feature vector for the fresh bamboo leaf product, including: The collected product image information is preprocessed, including background segmentation, image enhancement and size normalization, to segment out a clear target area of ​​fresh bamboo leaves; Color features are extracted from the preprocessed image of the target area of ​​fresh bamboo leaves, including the mean and variance in multiple color spaces, and the histogram distribution of each color component is statistically analyzed to calculate the color distribution of the leaves. The vein enhancement process is performed on the target area image of the fresh zongzi leaf to segment the vein structure. The contrast between the average gray value of the vein area and the average gray value of the non-vein area is calculated. Combined with the proportion of brown pixels in the vein area, the browning index of the vein is calculated. By analyzing the grayscale gradient changes of edge pixels in the target area of ​​fresh zongzi leaves, the curled or withered areas on the leaf edges are detected, and the ratio of the area of ​​the curled or withered areas to the total area of ​​the leaf is calculated to obtain the dehydrated area ratio of the leaf. The leaf color distribution, leaf vein browning index, and leaf dehydration area ratio are normalized and vectorized, and combined to form the real-time quality feature vector.

3. The intelligent preservation management method for fresh zongzi leaves based on warehousing data information according to claim 1, characterized in that, The preservation environment data and the real-time quality feature vector are fused and processed to predict the quality deterioration trajectory and remaining shelf life of fresh bamboo leaf products, including: The time-series data of temperature, humidity, and gas concentration that are aligned in time are concatenated with the real-time quality feature vector to form a multi-dimensional fused time series. The multi-dimensional fused time series is divided into multiple consecutive input segments according to a fixed time window; Each input segment is sequentially fed into the improved LSTM neural network model, and the improved LSTM neural network model sequentially outputs the quality feature prediction vector corresponding to each input segment. The true quality feature vector of a historical time period is compared with the predicted quality feature vector of the corresponding time period, the prediction error is calculated, and the prediction error is used to backpropagate and train the parameters of the improved LSTM neural network model. During the prediction phase, the latest input fragment is fed into the trained model, and the model scrolls forward to predict the quality feature prediction vectors for multiple future time steps. By connecting historical real values ​​with future predicted values, a complete quality degradation trajectory is formed. Based on the time point at which the quality deterioration trajectory reaches the preset quality deterioration threshold, the time span from the current moment to the time point is calculated as the remaining shelf life.

4. The intelligent preservation management method for fresh zongzi leaves based on warehousing data information according to claim 3, characterized in that, The improved LSTM neural network model dynamically adjusts its forgetting gate mechanism based on the nonlinear relationship between the gas concentration time-series data and the leaf browning rate, including: In the improved LSTM neural network model, a concentration influence weight is assigned to the ethylene concentration and carbon dioxide concentration in the gas concentration time series data, respectively; In the forgetting gate structure of the model, the forgetting gate activation calculation at each time step is decomposed into a basic forgetting part and a concentration effect part; The basic forgetting part is calculated based on the hidden state of the previous moment and the currently input temperature time series data and humidity time series data; The concentration influence component generates a concentration adjustment factor by multiplying the ethylene concentration value and carbon dioxide concentration value at the current time step by their corresponding concentration influence weights, summing them, and then passing the sum through a nonlinear transformation function. The concentration adjustment factor is weighted and fused with the basic forgetting component to form the final forgetting gate output of the time step; This mechanism allows the model to dynamically adjust the degree of forgetting of historical quality states when there are abnormal changes in ethylene or carbon dioxide concentrations, thus responding more sensitively to sudden changes in browning rate caused by changes in the gas environment.

5. The intelligent preservation management method for fresh bamboo leaves based on warehousing data information according to claim 1, characterized in that, Based on the quality degradation trajectory and remaining shelf life, a refined preservation strategy is generated for the storage unit, including: Analyze the quality degradation trajectory to identify the key time points and rates of quality decline; Based on the key time points and the rate of decline, the intensity of the physiological and biochemical reactions to be suppressed can be deduced, thereby deriving the target temperature range and the target humidity range. Based on the derived target temperature range and target humidity range, and combined with the actual temperature and humidity of the current storage unit, the time-segmented temperature setting curve and humidity adjustment command are generated. Based on the trend of the gas concentration time series data, especially the cumulative trend of ethylene concentration, the gas replacement volume and replacement frequency required to delay quality deterioration are calculated, and the modified atmosphere gas replacement scheme is generated. Based on the remaining shelf life, the fresh bamboo leaf products of different batches or locations in the storage unit are sorted, and products with shorter remaining shelf life are assigned higher outbound priority, thus generating the outbound priority sequence. The temperature setting curve, humidity adjustment command, modified atmosphere gas replacement scheme, and outbound priority sequence are integrated into a unified refined preservation strategy command set.

6. The intelligent preservation management method for fresh zongzi leaves based on warehousing data information according to claim 5, characterized in that, The generation of the time-segmented temperature setting curve and the humidity adjustment command includes: Extract the predicted rate of change of leaf dehydration area ratio and the rate of change of leaf vein browning index for multiple consecutive time periods in the future from the quality deterioration trajectory. Based on the correlation between the rate of change of the leaf dehydration area ratio and the intensity of transpiration, a transpiration inhibition model was established, with the target environmental humidity as the main control variable. Based on the correlation between the rate of change of leaf vein browning index and polyphenol oxidase activity, an enzyme activity inhibition model was established, with the target ambient temperature as the main regulating variable. By jointly solving the transpiration inhibition model and the enzyme activity inhibition model, the combination of target temperature and target humidity that reduces the predicted quality degradation rate to below the preset rate is obtained. Multiple target temperatures and target humidity combinations corresponding to consecutive future time periods are converted into temperature setpoint sequences and humidity setpoint sequences that can be executed by the control device, thus forming the temperature setpoint curve and the humidity adjustment command.

7. The intelligent preservation management method for fresh zongzi leaves based on warehousing data information according to claim 2, characterized in that, The leaf vein browning index is calculated by combining the proportion of brown pixels in the leaf vein region, including: Determine the color threshold range used to define brown pixels. The color threshold range is defined in the RGB color space as R component being greater than a first preset threshold, G component being less than a second preset threshold, and B component being less than a fourth preset threshold, or in the HSV color space as hue being within a preset brown hue range. For the leaf vein region image after leaf vein enhancement and segmentation processing, determine whether the color value of each pixel falls within the color threshold range of the brown pixel, and count the total number of pixels that are judged to be brown. The ratio of the total number of brown pixels to the total number of pixels in the leaf vein region image is calculated as the proportion of brown pixels in the leaf vein region. The average gray value of the leaf vein region image and the average gray value of the non-leaf vein region image are obtained, and the ratio or difference between the average gray value of the leaf vein region and the average gray value of the non-leaf vein region is calculated as the contrast of the leaf vein region. The brown pixel ratio and the contrast of the leaf vein region are weighted and summed, with the brown pixel ratio assigned a higher weight coefficient. The weighted sum is then normalized and mapped to a preset index range to obtain the final leaf vein browning index.

8. The intelligent preservation management method for fresh zongzi leaves based on warehousing data information according to claim 1, characterized in that, The method further includes: The storage equipment is driven to execute the preservation strategy, and the prediction results of the improved LSTM neural network model are continuously verified and the parameters are iteratively updated based on the subsequently collected preservation environment data and real-time quality feature vectors. The drive storage equipment executes the preservation strategy, including: The preservation strategy instruction set is parsed into specific equipment control instructions, which include refrigeration unit start / stop instructions, humidifier power adjustment instructions, controlled atmosphere storage gas valve opening / closing instructions, and conveyor line scheduling instructions. The start / stop command of the refrigeration unit and the power adjustment command of the humidifier are sent to the environmental control system to control the temperature and humidity in the storage unit to be adjusted according to the temperature setting curve and the humidity adjustment command. The control atmosphere storage gas valve opening and closing commands are sent to the gas control system to control the operation of the nitrogen generator, carbon dioxide removal machine and ethylene removal machine, and to execute the control atmosphere gas replacement scheme. The conveyor line scheduling instruction is sent to the warehousing and logistics execution system. Based on the outbound priority sequence, the handling equipment is planned and scheduled to reserve outbound channels for high-priority products or move them to the pre-outbound area.

9. The intelligent preservation management method for fresh bamboo leaves based on warehousing data information according to claim 8, characterized in that, The step of performing rolling verification and parameter iterative updates on the prediction results of the improved LSTM neural network model based on subsequently collected preservation environment data and real-time quality feature vectors includes: After implementing the preservation strategy, new preservation environment data and new product image information are collected again at the preset feedback cycle point. Perform the same visual quality analysis on the new product image information to generate a new real-time quality feature vector. The preservation environment data and real-time quality feature vector from the previous time are reassembled with the new preservation environment data and new real-time quality feature vector to form a verification dataset. The validation dataset is input into the improved LSTM neural network model currently in use to obtain simulated predictions of the quality feature vectors for the current and past time points. Calculate the error loss between the simulated predicted value and the true value in the validation dataset; When the error loss exceeds a preset threshold, the improved LSTM neural network model is incrementally trained using the validation dataset to update its model parameters and generate a model version with iteratively updated parameters.

10. The intelligent preservation management method for fresh zongzi leaves based on warehousing data information according to claim 9, characterized in that, The incremental training of the improved LSTM neural network model using the validation dataset includes: The weight parameters of most network layers, except the output layer, are fixed in the improved LSTM neural network model. Using the aforementioned validation dataset, the gradient calculation and update are performed primarily on the weight parameters of the model output layer and the concentration influence weights via the backpropagation algorithm. An optimizer with a small learning rate is used to perform parameter updates to prevent catastrophic forgetting of previously learned knowledge; After the update is completed, the generalization performance of the model with iteratively updated parameters is evaluated using an independent test dataset. The new model is used to replace the original model only if the generalization performance is not lower than that of the original model.

Images