Fresh corn circulation mildew-proof and quality-keeping method based on intelligent monitoring

Abstract

The application discloses a fresh corn circulation mildew-proof quality guarantee method based on intelligent monitoring, and particularly relates to the technical field of agricultural product circulation management. Quality data is collected through image, gas, humidity and temperature, and microorganism multi-source sensing means, and a quality degradation trend curve and a risk grade score are generated in combination with a time sequence modeling algorithm. In a high-humidity or oscillation environment, identification parameters are automatically adjusted to improve early-stage mildew identification accuracy. High-risk batches are identified through multi-dimensional cross analysis to trigger a circulation suspension response. Whole-process data is uploaded to a management platform to form traceable quality records, and whole-process dynamic monitoring and closed-loop control are realized. The application identifies implicit mildew through multi-modal sensing, predicts and evaluates shelf life, dynamically adjusts to improve identification stability, and builds a chain data traceability mechanism to realize model self-evolution, thereby improving quality early warning, risk control and continuous optimization capabilities in the whole process of fresh corn circulation, and ensuring accurate and reliable quality guarantee and mildew-proof effects.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural product circulation management, more particularly, the present application relates to a fresh corn circulation mildew-proof and quality preservation method based on intelligent monitoring. BACKGROUND

[0002] Fresh corn is a kind of high-moisture and high-activity agricultural product with fresh eating properties and nutritional value. During the transportation, storage and distribution from harvesting to terminal sales, it is prone to quality degradation and mold problems, especially in high-temperature and high-humidity environments in summer, which can easily induce the rapid reproduction of mold such as Fusarium graminearum, leading to spoilage, off-flavor, discoloration and other phenomena, seriously affecting product quality and food safety.

[0003] The existing fresh corn circulation mildew-proof and quality preservation technology mainly relies on cold chain transportation, conventional bacteriostatic treatment (such as ozone, ultraviolet, packaging modification, etc.) and artificial sampling inspection to ensure product quality. However, there are generally the following technical bottlenecks: first, the detection means is single, mainly relying on manual observation and visual appearance characteristics, which is difficult to identify early "hidden mold" risks; second, there is a lack of modeling and prediction ability for product quality change trend, which can only respond passively after the problem occurs and cannot intervene in advance; third, the quality judgment relies on manual experience and is greatly affected by subjective factors, and the response is not timely; fourth, the monitoring links are scattered, and a complete closed loop has not been formed, lacking a data-driven dynamic quality management mechanism.

[0004] Especially under the background of the rapid development of intelligent agricultural product circulation and digital supply chain, the traditional method cannot meet the intelligent needs of high-frequency circulation, stable quality and controllable quality of fresh corn. Therefore, the present application proposes a fresh corn circulation mildew-proof and quality preservation method based on intelligent monitoring to meet the needs of product quality closed-loop management in the circulation process. SUMMARY

[0005] To achieve the above purpose, the present application provides the following technical scheme:

[0006] The fresh corn circulation mildew-proof and quality preservation method based on intelligent monitoring comprises the following steps:

[0007] By configuring an image acquisition module, a gas detection module, a temperature and humidity sensing module and a microbial rapid detection device, multi-source quality data of fresh corn is synchronously collected in different circulation links, including color, skin texture, gas volatile concentration, temperature and humidity change and pathogenic bacteria metabolism index;

[0008] Using a quality trend prediction algorithm based on time series modeling, the above multi-dimensional data is sequentially trained to construct a quality degradation trend curve, automatically judge the quality degradation degree and future shelf life, and give a risk level score;

[0009] The color recognition sensitivity, gas warning threshold and microorganism detection frequency are adjusted in real time when high humidity fluctuation or shock event occurs in the storage and transportation environment, the parameter recognition model is adaptively reconstructed to improve the recognition accuracy at the initial stage of mold growth.

[0010] The multi-dimensional cross analysis is performed on the multi-modal data of each batch of fresh corn, and when the risk level predicted by the recognition model exceeds the standard warning value, the circulation stop processing signal is triggered, the potentially hidden moldy batch is preferentially removed, and the product with normal appearance but internal deterioration is prevented from entering the sales process.

[0011] The collected data, prediction results and processing records at each stage are uploaded to the centralized management platform through a data interface, and the quality traceability chain is automatically archived.

[0012] In a preferred embodiment, the image acquisition module includes a high-resolution image sensor and a near-infrared reflection unit, and the reflection image of the preset waveband is used to extract the weak color difference and local wet spot area change signal of the corn kernel skin. When there is no obvious abnormality in normal visible light detection, the spectral enhancement feature recognition is used to determine the potential moldy signs; the image data is subjected to color space conversion and region segmentation operation in the preprocessing stage, the image is converted from red-green-blue color space to brightness-saturation-hue space, and dynamic difference calculation is performed in combination with edge gradient change and color difference distribution mean, thereby improving the recognition probability of early mold point micro-cluster.

[0013] In a preferred embodiment, the gas detection module is constructed based on the principle of electronic nose, and the volatile organic compounds in the corn storage environment are recognized by multi-channel metal oxide gas sensitive elements, including aldehydes, ketones and low molecular weight fatty acid components in mold metabolites. By constructing a feature response vector under different odor intensity combinations, odor pattern recognition is achieved; at the same time, the "deterioration odor deviation sensitivity coefficient" is set, when the coefficient value is higher than the gas perception sensitive response warning threshold, the mold growth process weight adjustment of the prediction model is triggered, the recent data segment is given high risk tendency, to realize the early locking of mold growth path, and the "deterioration odor deviation sensitivity coefficient" is calculated based on any one of the three types of discrimination methods of Mahalanobis distance, Manhattan distance and Shannon entropy.

[0014] In a preferred embodiment, the microbial rapid detection device adopts a non-destructive detection mechanism based on impedance or bioluminescence method to capture the metabolic activity indicators of living colonies present on the surface of corn kernels or inside the package. Before the detection data enters the prediction model, high-frequency sample fluctuation difference fitting is first performed. When the difference value exceeds the microbial metabolic trend deviation threshold, the critical sample backtracking section is entered for historical curve comparison. The comparison adopts one of the two strategies of moving average fitting and perturbation trend regression to determine whether it is a significant deviation event. When the data in the deviation interval is in the last three consecutive periods, the prediction model raises the warning level of the batch data from low risk to medium-high risk interval.

[0015] In a preferred embodiment, the quality trend prediction algorithm adopts a multi-level time series modeling strategy, combining short-term trend window and periodic trend window for dual evolution analysis, and introducing a causal weight optimization mechanism in the core part of the model. This mechanism dynamically adjusts the data weight proportion involved in modeling based on the strength of the causal relationship between different data channels.

[0016] In a preferred embodiment, the risk level is divided into five levels when the prediction result produces the shelf life, namely stable state, fluctuation warning, initial risk, moderate risk and high risk state, each level corresponds to different warning response and disposal strategies.

[0017] In a preferred embodiment, the adaptive reconstruction process of the parameter identification model includes three parts: sensitivity dynamic adjustment module, identification error feedback module and environmental response weight mapping mechanism. After detecting the high-humidity environment trigger, the "fine gray spot variation response threshold" in the color identification model is automatically reduced, and the odor intensity change discriminant factor is simultaneously increased. When the environment fluctuation has not returned to normal state for three consecutive periods, the identification error feedback module is started, and the current model prediction boundary is adjusted based on the true determination result of the previous period.

[0018] In a preferred embodiment, the data cross-analysis adopts a vector similarity matching and time series difference fusion mechanism. Image, gas, temperature and humidity data channels are used to construct feature vectors, and an omnidirectional data cross-section map is formed for each batch of fresh corn within the time period. If the similarity index between the maps decreases and is accompanied by a prediction risk score close to the quality abnormality trigger threshold, the batch is automatically marked as a "key monitoring object". Such batches will trigger a secondary quality inspection mechanism and data refresh training before entering the warehouse or transportation process, in order to strengthen the model's ability to accurately identify boundary samples.

[0019] In a preferred embodiment, the centralized management platform displays the postharvest processing, packaging form, transportation path and quality change trend of different batches of fresh corn based on the visual chain structure, records the current input data state and output prediction result and matches with the actual quality feedback every time the prediction cycle is processed, if the prediction deviation amplitude of two consecutive periods is greater than the prediction accuracy stability threshold, the platform automatically marks the current model for retraining, and selects the latest batch of stable samples as the training base.

[0020] Technical effects and advantages of the present application:

[0021] The present application significantly improves the recognition ability of the "hidden mildew" state of fresh corn by introducing multi-modal perception means, and overcomes the technical bottleneck of traditional monitoring methods relying on appearance features and easy to miss early corruption. Image recognition combined with near-infrared band enhancement, gas sensing identification of mold metabolites, and non-destructive identification of active signals by microbial rapid detection, so that the present application can capture early quality risks through internal parameter changes when there is no obvious abnormality in appearance, realize pre-warning, and ensure food circulation safety.

[0022] The present application realizes the intelligent upgrade from "current judgment" to "trend prediction" by constructing a quality trend prediction model based on time series and combining risk level scoring and dynamic parameter adjustment mechanism. The model can accurately estimate the shelf life according to the change trend of different channel data, and dynamically adjust the recognition threshold and weight factor in abnormal environments such as high humidity and vibration, ensuring stable and accurate recognition ability under complex transportation conditions, thereby improving the reliability and adaptability of actual deployment.

[0023] The present application realizes data full-process closed-loop management in operation, and the centralized platform forms a chain record of each batch of data, prediction results and actual feedback, which is not only used for real-time display and early warning, but also supports model retraining and self-iteration optimization. When the recognition accuracy appears a downward trend, the platform automatically selects stable samples for model updating to ensure that the prediction accuracy remains stable during long-term use, thereby forming an intelligent self-evolution ability of "perception - modeling - feedback - evolution", and making the entire mildew prevention and quality preservation strategy have the technical advantage of continuous improvement. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to facilitate understanding of those skilled in the art, the present application will be further described below with reference to the accompanying drawings;

[0025] Figure 1 The principle diagram of the fresh corn circulation mildew prevention and quality preservation method based on intelligent monitoring in the present application. DETAILED DESCRIPTION

[0026] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the scope of protection of the present application.

[0027] With reference to Figure 1 The following examples are obtained:

[0028] Embodiment 1: Fresh corn circulation mildew-proof and quality preservation method based on intelligent monitoring, comprising the following steps:

[0029] By configuring an image acquisition module, a gas detection module, a humidity and temperature sensing module, and a microbial rapid detection device, multi-source quality data of fresh corn in different circulation links are synchronously acquired, including color, skin texture, gas volatile concentration, humidity and temperature change, and pathogenic bacteria metabolism index. This step belongs to the sensing stage of the present application, and the purpose is to establish a sensing network with multi-type acquisition capability to comprehensively acquire the quality state of fresh corn in each link from postharvest to sales. Specifically, it includes:

[0030] Image acquisition module: used for acquiring color, luster and texture change information of corn skin;

[0031] Gas detection module: used for monitoring odor volatile compounds produced by microbial metabolism in the storage environment;

[0032] Humidity and temperature sensing module: used for monitoring temperature and humidity fluctuations in the flow process;

[0033] Microbial rapid detection device: used for identifying possible pathogenic bacteria activity signals on the surface of corn or inside the package. Through the above configuration, a multi-dimensional data channel covering visual features, gas features, environmental parameters and biological indicators can be established to provide data support for subsequent intelligent analysis.

[0034] Using a quality trend prediction algorithm based on time series modeling, the above multi-dimensional data is sequentially trained to construct a quality degradation trend curve, automatically judge the quality deterioration degree and future shelf life, and assign a risk level score. This step is the intelligent judgment stage of the present application, and the core is to use the evolution relationship between historical data and real-time data to simulate and predict the quality change trend of fresh corn in the future period of time. The implementation includes:

[0035] The data of multiple channels such as images, gases, temperature and humidity and microorganisms are trained in a time sequence manner; a quality degradation curve is generated by using a trend modeling algorithm, reflecting the evolution path of the corn from the current state to the deteriorated state; the remaining edible period is calculated by judging the stage of the current curve; meanwhile, a five-level risk grade is assigned to the current state of the sample, corresponding to different quality states from stable to high risk, to guide the subsequent circulation management. The goal of this link is "early identification", not post-processing.

[0036] When there is a high humidity fluctuation or a shock event in the storage and transportation environment, the color recognition sensitivity, gas warning threshold and microorganism detection frequency are adjusted in real time to realize adaptive reconstruction of the parameter recognition model, so as to improve the recognition accuracy in the early stage of mildew. The problem solved by this step is that the recognition accuracy of the traditional recognition decreases under the condition of environmental abnormalities (such as sudden humidity rise and transportation vibration). Therefore, the present application proposes a parameter dynamic adjustment mechanism, which includes:

[0037] When there is a sharp rise in humidity or continuous vibration during storage or transportation, the environmental changes are monitored; on this basis, the model automatically reduces the sensitivity threshold for judging the color spot change in color recognition; at the same time, the response sensitivity to gas concentration change is improved, so that the active signs of microorganisms caused by heat and humidity can be detected earlier; the sampling frequency of microorganism detection is also increased synchronously to increase the opportunity to capture abnormal states. This step ensures that the model can still operate stably under non-ideal environment and improves the recognition ability of early features of mildew.

[0038] The multi-dimensional cross analysis is performed on the multi-modal data of each batch of fresh corn, and when the risk level predicted by the recognition model exceeds the standard warning value, the potential hidden mildew batch is preferentially removed, and the appearance normal but internal deterioration products are prevented from entering the sales process; this step is the intelligent response control stage, which is the key link for the whole method to provide intervention actions to the front end of the industry chain. The operation logic is as follows:

[0039] Each batch of corn is detected, and cross verification is performed in combination with multiple data sources to improve the judgment accuracy; if the risk level is rated as high, and exceeds the preset warning standard value; the "processing prompt" or "block signal" is actively triggered, and the batch is marked as "potential unqualified"; the management end will suspend the further circulation of the batch in the logistics, and suggest to perform artificial reinspection or cold chain interruption; at the same time, the batch will be removed from the preset sales logic path, to ensure that it does not enter the terminal market. Here, the "risk blocking mechanism" is formed, which is different from the traditional static mode relying on human experience.

[0040] The data collected in each stage, the prediction results and the processing records are uploaded to the centralized management platform through the data interface, and the quality traceability chain is automatically archived. This step is a closed-loop management link, and the purpose is to establish a quality traceability record of fresh corn from harvesting to sales through continuous archiving and data accumulation. The key point is that:

[0041] The data generated in each link (such as perception value, prediction result, risk score, processing suggestion, etc.) is recorded; the data is formed into a unified storage structure through the docking platform; the platform can visually display the circulation path and quality evolution track of each batch of products; and it also supports subsequent model optimization, quality responsibility traceability and product label intelligent assignment functions. Through this mechanism, the digitalization, standardization and transparency of quality control are truly realized.

[0042] The image acquisition module includes a high-resolution image sensor and a near-infrared reflection unit. The reflection image of the preset wave band is used to extract the weak color difference and local wet spot area change signal of the corn kernel surface. When there is no obvious abnormality in normal visible light detection, the potential mildew signs are determined through spectral enhancement feature recognition. The image data is subjected to color space conversion and region segmentation operation in the preprocessing stage, and the image is converted from the red-green-blue color space to the brightness-saturation-hue space, and the dynamic difference is calculated in combination with the edge gradient change and the color difference distribution mean, so as to improve the recognition probability of the early mildew point micro-cluster.

[0043] In this link, the high-resolution image sensor can select an industrial camera device with more than 13 million pixels, which can clearly image the details of the corn surface under different light source conditions. The near-infrared reflection unit integrates a wave band control module, so that the image acquisition is not limited to the visible light range of the human eye, but covers a specific wave band including the near-infrared segment. The commonly used reflection center wave band range can be set between 720 to 900 nanometers. This reflection range can effectively improve the imaging ability of low-contrast structures such as water stains and initial appearance of mycelium.

[0044] After the image data is collected, it enters the preprocessing stage. First, the color space conversion operation is performed to convert the original image from the traditional red-green-blue color space to the brightness-saturation-hue space. This conversion can decouple the intensity and brightness of the color, making it easier to quantitatively analyze the "small color difference" in the future. This processing step can be realized through the color space mapping function in the open source image processing tool, such as the common color space transformation algorithm in the image processing library, or a custom color mapping method with equivalent effect can be used instead.

[0045] After the color space conversion is completed, the image enters the region segmentation processing link for extracting the kernel unit region and the suspected moldy region. The region segmentation can adopt a traditional algorithm based on edge detection, such as a Sobel operator, a Canny operator and the like to extract the edge gradient change of the image; or a deep learning model can be applied, such as a semantic segmentation network based on a convolution structure to realize intelligent segmentation of the kernel contour. The present application does not limit the algorithm structure for segmentation processing, and any technical manner that can achieve accurate extraction of the kernel region can be applied instead.

[0046] In the segmented image region, the mean information of the edge gradient change and the color difference distribution is further calculated, and the pre-mold spot features possibly existing in the image are extracted through dynamic difference calculation. The dynamic difference calculation can be realized by adopting a pixel block difference accumulation and local statistical variance comparison manner, or a sliding window filter based on local gray value change can be introduced for comparison and analysis. By comparing the slight differences in color, brightness and texture of different regions, those image regions that have not yet formed visible mold spots but have potential rotting characteristics can be identified.

[0047] It is worth noting that the feature recognition process used in the image acquisition module can correspond to the visual data channel in the quality trend prediction algorithm of the present application, that is, the processed image feature value is taken as one of the inputs and is transmitted to the model structure to participate in the subsequent shelf life prediction and risk level scoring.

[0048] In addition, the present application does not limit the specific image processing algorithm or classification model to be used, and any technical path that can realize the early moldy image feature enhancement and extraction function can be realized instead, including but not limited to deep neural network, support vector machine, fuzzy logic image classifier and the like.

[0049] For further illustration, the following embodiments can be used as a reference in the present application: in the acquisition stage, a red-green-blue and near-infrared dual-channel camera is configured to acquire images of the corn skin, when there is no abnormality in the visible light band image, the reflection distribution change in the near-infrared image is used to compare the image difference graph of the same batch, and it is found that some kernel surfaces have abnormal enhanced reflection intensity regions, which are determined as local humidity abnormally high regions after processing, and are marked as potential moldy warning regions after confirmation of temperature and humidity data, and finally the image region enters the model structure as a weight feature for identification and judgment.

[0050] The gas detection module is constructed based on the principle of an electronic nose, and volatile organic compounds in a corn storage environment are identified by a multi-channel metal oxide gas sensitive element, including aldehydes, ketones and low molecular weight fatty acid components in mold metabolites. By constructing a characteristic response vector under different odor intensity combinations, odor pattern recognition is achieved. At the same time, the "rot odor offset sensitivity coefficient" is set. When the coefficient value is higher than the gas perception sensitive response warning threshold, the mold progress weight adjustment of the prediction model is triggered, the recent data segment is given a high risk tendency, and the mold path is locked in advance. The "rot odor offset sensitivity coefficient" is calculated based on any one of the three types of discrimination methods of Mahalanobis distance, Manhattan distance and Shannon entropy.

[0051] The module includes a plurality of independent gas sensitive elements in hardware structure, each element is based on a metal oxide semiconductor structure, and a redox reaction occurs at a certain working temperature to produce a resistance change, reflecting the fluctuation of a specific gas concentration. A typical sensing element can respond to n-butyl aldehyde, hexanone, acetic acid and other mold metabolism related components, with a response time of less than two seconds and a recovery time of less than five seconds, suitable for dynamic real-time acquisition in a flow process.

[0052] The original output signal of each sensing unit can be normalized to construct a set of characteristic response vectors, which contains multiple dimensions, each dimension corresponding to the output amplitude of a gas sensitive element. Through the combination structure of the vector, a unique odor pattern map can be formed. In order to simplify the implementation, the first n principal components can be extracted by principal component analysis in the specific implementation of the present application to construct the reduced dimension characteristic vector, but it is not limited to this way. Any algorithm that can achieve feature extraction and recognition accuracy, including but not limited to linear discriminant analysis, autoencoder compression network and fuzzy clustering method, can be used instead.

[0053] The above-mentioned "rot odor offset sensitivity coefficient" refers to the similarity offset degree of the gas response vector between multiple time periods, which is used to measure the abnormal change of the sample in the odor level. The present application provides three implementable but not unique calculation methods:

[0054] Mahalanobis distance calculation method: which can be used to measure the multi-dimensional statistical difference between the current gas response vector and the historical standard sample. This method considers the covariance relationship of each dimension, and is effective for non-independent features.

[0055] Manhattan distance calculation method: which measures the overall change of odor by summing the absolute value difference of each dimension, with low computational complexity and suitable for edge processing scenarios.

[0056] Shannon entropy difference comparison: by calculating the total amount of information entropy of each gas component distribution probability, and comparing the entropy difference between the current sample and the reference sample, the pattern uncertainty rising situation is found.

[0057] The present application does not limit the calculation of the odor shift sensitivity coefficient to a certain method, as long as the method used can objectively reflect the odor response difference and has stability and repeatability, it can be considered as equivalent technical means.

[0058] When the odor shift sensitivity coefficient exceeds the set gas perception sensitive response warning threshold, the subsequent processing mechanism is triggered. This threshold can be set by statistical derivation in historical moldy sample data, or it can be obtained by adjusting the false positive rate according to the actual deployment environment. The specific value can be set as a risk threshold based on the percentile of the odor response shift rate in the sample set (for example, the eighty-fifth percentile).

[0059] When the above odor shift value is determined to be a high-risk area, the mold progression weight factor in the quality trend prediction model mentioned in the present application will be dynamically adjusted, so that the model automatically increases the risk discrimination weight of the odor channel in the recent data analysis. This weight adjustment mechanism can be reflected in the input feature weighting layer of the model, such as giving a multiplicative weight amplification factor to the gas channel output result in the last two hours of data block, so that it has a higher decision-making influence in the internal calculation of the model.

[0060] The role of this mechanism is that even if the current image features or microbial indicators have not yet shown obvious abnormalities, based on the "olfactory risk change in advance" signal, it can realize "advance tendency judgment" on future trends, lock the mold risk path in time, and through the subsequent control mechanism, the relevant batches are included in the warning range.

[0061] In a specific example, the gas detection module contains five gas sensing elements with different response spectra, which respond to components such as acetaldehyde, butanone, acetic acid, etc. During storage and transportation, data sampling is performed every ten minutes, and the historical samples in the past hour are compared to construct a response vector set. The Mahalanobis distance algorithm is used to calculate the deviation between the current sample and the first five healthy samples, and the result is compared with the pre-set warning threshold. Once the shift coefficient is higher than the threshold, a "high risk" signal is immediately transmitted to the trend prediction model (i.e. the time series judgment model constructed in the present application), and the odor channel risk weighting submodule is activated, so that the model pays more attention to "olfactory abnormal trend" when making remaining shelf life judgment, so as to list the batch of corn in the high-risk pre-judgment list in advance for reference when the circulation control is used.

[0062] The microbial rapid detection device adopts a non-destructive detection mechanism based on impedance method or bioluminescence method, is used for capturing metabolic activity indicators of living bacteria colonies existing on the surface of corn kernels or inside the package, and first performs high-frequency sample fluctuation difference fitting before detection data enters a prediction model. When a difference value exceeds a microbial metabolic trend deviation threshold, a critical sample backtracking section is entered for historical curve comparison. The comparison adopts one of moving average fitting and perturbation trend regression strategies to determine whether it is a significant deviation event. When data in three consecutive periods are all in a deviation interval, the prediction model raises the warning level of the batch data from low risk to medium-high risk interval.

[0063] The detection device based on the impedance method mainly measures the change of conductivity of a solution or a surface medium between electrodes to indirectly determine the growth activity of microorganisms. Living bacteria colonies produce charge flow and ion concentration change in the metabolic process, thereby causing a specific amplitude deviation of the impedance value. This method is suitable for direct contact with corn kernels or indirect detection through a small space probe embedded in the packaging environment, and has the advantages of not damaging the packaging structure and real-time detection.

[0064] The device based on the bioluminescence method usually relies on the light signal intensity released by a specific enzyme system (such as luciferase-luciferin system) participating in the fluorescence reaction in the metabolic process of microorganisms to determine its activity state. The device captures the strength of the fluorescence signal by setting a photoelectric probe and a filtering photosensitive component, and constructs a microbial activity map in combination with parameters such as integration time and light intensity threshold.

[0065] The present application does not limit the use of any specific implementation path in impedance method or bioluminescence method, and both can be used for non-destructive and rapid acquisition of the activity state of microorganisms in the corn storage and transportation environment.

[0066] Before the detection data enters the prediction model, first perform high-frequency sample fluctuation difference fitting. When the difference value exceeds the microbial metabolic trend deviation threshold, enter the critical sample backtracking section for historical curve comparison. The comparison adopts one of moving average fitting and perturbation trend regression strategies to determine whether it is a significant deviation event.

[0067] The above-mentioned "high-frequency sample fluctuation difference fitting" refers to the change trend analysis of the microbial activity data obtained by continuous detection at a minute or hour level time interval. The purpose is to screen out data segments with sharp changes in activity value in a short time. The fluctuation difference can be completed by calculating the difference between the current data point and the mean value in the previous several periods, or calculating the mutation point of the standard deviation sequence. Specific implementation methods include sliding window variance detection, incremental gradient analysis or difference fluctuation index method, etc.

[0068] When the fluctuation amplitude exceeds the set "microbial metabolic trend deviation threshold", the critical sample analysis stage is entered. At this stage, the historical data of the current batch in the previous period is called to compare the consistency of the trend.

[0069] To improve the robustness of the judgment, the present application provides two trend comparison strategies as optional implementation modes: moving average fitting method: by weighting the average of the data points in the past several periods, a stable trend line is constructed, and the deviation from the current data is compared; perturbation trend regression method: based on the linear or nonlinear fitting of the slight disturbance in the time series, it is judged whether the current data change belongs to the normal fluctuation within the trend evolution or is an abnormal jump. The above two strategies can be used alone or jointly, or can be replaced by other algorithms with equivalent prediction and discrimination ability (such as adaptive curve regression).

[0070] When the data in the last three periods are all in the deviation range, the prediction model raises the alert level of the batch data from low risk to medium-high risk range. The prediction model mentioned here is the quality trend prediction model constructed in the present application, which predicts the evolution process of corn quality based on time series modeling architecture and outputs risk score.

[0071] The results of the microbial detection module are used as one of the model input channels, and the output values not only participate in the risk level scoring, but also have the ability to trigger the model internal risk weight switching mechanism. Specifically:

[0072] When the detection results show high amplitude deviation for three consecutive periods, it is assumed that the batch is in an abnormal biological activity state;

[0073] The model will activate the risk weight adjustment substructure to increase the participation of the microbial data channel in the multi-channel fusion prediction calculation; at the same time, in the state classifier inside the model, the current batch is directly upgraded from the low risk prediction label to the medium risk or high risk category, to drive subsequent control actions such as suspending delivery, starting refrigeration, and other responses. This mechanism ensures that the model no longer passively responds to visible abnormalities, but can intervene in the moldy trend in advance through the "hidden indicator" of microbial activity.

[0074] In a specific example, the detection device uses a miniature impedance sensing head placed in the air passage position inside the corn packaging box, and impedance change curves are collected every fifteen minutes. The threshold is set to be more than ten percentage points in the impedance value change amplitude within two consecutive periods. If it is in this deviation range for three consecutive periods, the trend mutation curve is obtained by moving average fitting, and in the quality trend prediction model in the back end, the risk weight of the microbial channel is multiplied by a doubling coefficient, and the batch is adjusted from the initial "stable state" to the "medium risk" state, and the processing suggestion is output as "cold chain reinforcement" and "sampling review".

[0075] The quality trend prediction algorithm adopts a multi-level time series modeling strategy, combines a short-term trend window and a periodic trend window for double evolution analysis, and introduces a causal weight optimization mechanism in the core part of the model. The mechanism is based on the strength of the causal relationship between different data channels to adaptively and dynamically adjust the data weight proportion involved in modeling.

[0076] In the present application, the quality trend prediction algorithm corresponds to the "quality trend prediction model" described above. Its core function is to dynamically model the quality evolution trend of fresh corn in storage, circulation, sales and other links, and output the shelf life estimate and risk level score to guide the downstream risk control and circulation strategy decision.

[0077] The multi-level modeling strategy refers to the construction of two trend evaluation windows of different time scales in the structure of the prediction algorithm:

[0078] Short-term trend window: used to capture the microscopic fluctuation trend of fresh corn quality in the next few hours or tens of minutes. This window is suitable for real-time early warning function, mainly focusing on recent sudden quality changes;

[0079] Periodic trend window: used to evaluate the medium and long-term evolution pattern of fresh corn quality over multiple collection periods, such as gradual decline trend caused by environmental fluctuations. This window can be set to a sliding evaluation period based on two days, three days or a week.

[0080] Both windows are input in the form of time series data, including image feature parameters, gas volatile component concentrations, temperature and humidity records, and microbial metabolic activity data, etc. Multiple channels, parallel modeling of trends at two scales, and fusion output at the prediction end.

[0081] Specifically, the short-term window can use a time series processing structure based on recurrent neural network or sliding regression model, while the periodic window can be modeled based on exponential weighted average, long short-term memory network, or seasonal decomposition model. The above methods are only one of the possible paths. In the present application, the causal weight optimization mechanism is used to solve the problem of varying contribution of different types of data channels to the quality evolution trend, improving the prediction flexibility and accuracy of the model in different environments or sample scenarios. The mechanism is embedded in the internal structure of the quality trend prediction model to dynamically adjust the participation proportion of different input data. Its basic logic is:

[0082] By statistically analyzing the correlation between image features, gas indicators, microbial activity and quality score, the causal relationship strength index between each channel and the target variable is extracted, and the strength is converted into the weight proportion involved in modeling. According to the current batch or environmental state, the weight configuration of each channel is updated in real time to enhance the response ability to the dominant factors.

[0083] The calculation of the strength of the causal relationship can be achieved using one of the following three methods:

[0084] Granger causality test method: test whether a channel variable is statistically "before" the quality degradation performance change;

[0085] Mutual information method: assess the relevance by calculating the information sharing degree between the input variables and the target quality parameters;

[0086] Minimum description length principle: measure the contribution of different channels to the explanatory power of the model.

[0087] The double-layer trend modeling results are fused at the output end of the model to form an integrated trend prediction result. This result includes:

[0088] A comprehensive quality degradation trend curve; the trend segment mark of the current quality; the predicted saleable period (i.e. shelf life); and the risk level score (divided into five levels from stable to high risk).

[0089] The prediction result is automatically transmitted to the "risk response control module" described in the invention for determining whether to enter the suspension of circulation processing, adjustment of packaging strategy or pre-cooling start and other measures.

[0090] The quality degradation trend curve is a curve-type data structure generated by the time series modeling strategy in the invention for continuous learning of the multi-dimensional quality data of fresh corn collected, and is used to dynamically describe the change trajectory of the quality state of fresh corn over time in each stage of storage, transportation and circulation.

[0091] The construction process of the quality degradation trend curve is as follows:

[0092] Input data source: continuous time series data from the image acquisition module, gas detection module, temperature and humidity sensing module, and microbial detection module, forming a multi-channel input matrix.

[0093] Feature representation processing: each channel data can be preprocessed by standardization, trend normalization, principal component compression, etc. to extract its representative change path.

[0094] Time series modeling: continuous training of data flow by time series prediction model (such as recurrent neural network, long short-term memory model, sliding weighted model, etc.) to form a time-quality state mapping relationship.

[0095] Trend curve generation: after concatenating and smoothing the prediction output of the model at each time node, a continuous curve that changes over time is obtained, which is called the quality degradation trend curve.

[0096] Key point extraction: According to the structural characteristics of the curve such as slope change, inflection point, critical section position, etc., the state intervals such as quality rapid decline period, stable period, critical period, etc. can be identified.

[0097] The curve is essentially a "time function image of quality score", that is, by analyzing the time sequence change of multi-channel data, a quality trend score function that decreases or fluctuates with time is derived.

[0098] Based on the quality degradation trend curve, the following two derived modules are designed at the output end of the model:

[0099] Risk level scoring module: compare the trend score corresponding to the current time point with the historical health sample distribution, and according to its position in the risk distribution interval, give different levels such as "stable state", "initial risk", "moderate risk", "high risk", etc.

[0100] Shelf life prediction module: the slope trend between the current value and the termination value of the trend curve is used to calculate when it will fall below the inedible critical value (such as the score corresponding to the third percentile of the historical sample), and thus the remaining saleable time is derived. For example, if the current trend score is eighty and the termination value is forty, and the average decline rate of the curve is five score units per hour for the last three hours, then the shelf life can be calculated to be about eight hours, which can be used as a reference for safe sale time limit.

[0101] In a specific example, the short-term trend window is set to six hours, the periodic trend window is set to a three-day sliding period, and the input data includes image spot aggregation value, volatile gas concentration integral value, air humidity fluctuation range and total amount of colony activity light signal. First, the data of the two windows are modeled using recursive regression model and exponential weighted moving average model respectively.

[0102] Then, the mutual information algorithm is called to calculate the correlation score between each channel and the shelf life label, and the modeling input weight configuration is adjusted accordingly. It is found that the causality of the gas data channel has significantly increased under the current environment, so the model automatically increases the participation weight of this channel, making the gas factor have greater influence in trend judgment. Finally, the model predicts that the remaining shelf life of the current batch of products is twenty-four hours, and the risk level is "initial risk", and the result is transmitted to the management platform for the control decision module to call and execute.

[0103] In another specific example, the image feature change rate, gas concentration curve, humidity fluctuation amplitude, and microbial metabolic index of fresh corn are collected, and continuous modeling is performed with a period of thirty minutes. The constructed quality degradation trend curve shows that the score remains between ninety and eighty-five within the first six hours, and then starts to accelerate downward after a high humidity event triggers, and drops to sixty-five at the tenth hour, with a clear inflection to rapid deterioration. It is determined that this is a "critical downward section", the current risk level is adjusted to "moderate risk", and it is predicted that if the current downward trend remains unchanged, it will fall into the unsafe score interval (such as a score below fifty) in six hours, so that the remaining shelf life is about six hours. This curve not only provides the basis for the internal early warning mechanism, but also presents it to the operating personnel through the platform graphical interface, providing auxiliary judgment basis for warehouse or shipping strategies.

[0104] While predicting the shelf life, the risk level is divided into five levels, namely stable state, fluctuation warning, initial risk, moderate risk, and high-risk deterioration state, and each level corresponds to different early warning responses and disposal strategies.

[0105] The prediction result includes a quality degradation trend curve and a trend score value at the current time point. The current score value is compared with the existing score interval threshold to determine its belonging level in the five-level division. This level is not only used for label identification, but also directly triggers different disposal logic downstream.

[0106] The definition of the five-level risk level is as follows: stable state definition: the current quality trend score is higher than the set stable threshold (for example, the score value is greater than eighty-five, and the full score is one hundred), and the trend curve has no obvious downward trend; characteristics: no odor, no appearance abnormalities, normal microbial activity; response: allow normal circulation, no additional intervention is required.

[0107] Fluctuation warning definition: the score is in the interval of eighty-five to seventy-five, or there is a slight short-term decline; characteristics: environmental fluctuations are obvious, but the quality is still within an acceptable range; response: record warning, suggest to strengthen packaging and sealing or enter short-chain sales channel.

[0108] Initial risk definition: the score is in the interval of seventy-five to sixty-five, and the trend curve has an inflection point of accelerated downward trend; characteristics: microbial metabolism starts to rise, and there is no obvious change in appearance; response: start warning prompt, suggest to sell out in advance or activate pre-cooling mechanism.

[0109] Moderate risk definition: the score is between sixty-five and fifty-five, and the downward rate of the trend curve is significant; characteristics: "offensive odor sensitivity coefficient" exceeds the warning value; response: suspend the shipment of this batch, transfer to the re-inspection process or start the active sampling mechanism.

[0110] High-risk deterioration state definition: score below fifty-five, quality trend curve has approached the endpoint value; characteristics: multi-channel signal is highly abnormal, there is visible deterioration or obvious odor; response: block the circulation authority of the batch, mark as "unsalable goods", and suggest to destroy or recycle.

[0111] The score boundaries of each risk level can be set in the following ways: experience value setting: based on the experience distribution of historical quality data, the mean percentage of stable batches is taken to set the threshold; clustering algorithm division: using unsupervised clustering methods such as k-means or Gaussian mixture model to hierarchically classify the score values; supervised learning classification model: such as support vector machine, random forest model, etc., train the existing labeled samples to obtain the division interface. The present application does not limit the use of a certain algorithm for risk interval division, as long as a stable corresponding relationship between the score segment and the risk level can be realized, it can be considered as a technically equivalent path.

[0112] The input of the risk division mechanism is the output content of the quality trend prediction model mentioned in the present application, specifically the trend score value (i.e. the current time output value of the quality score function) modeled according to the multi-modal perception data in the model. The score generation module in the prediction model can directly map the regression prediction result to the score value, or it can be obtained by normalizing the comprehensive features of the trend curve (such as the current value, curvature, slope).

[0113] For example, the score value can be calculated as follows (one of the implementable examples): divide the current time trend curve point value by the maximum value at the initial time, multiply by one hundred, form a relative quality index; take the slope change factor (trend fluctuation degree) as a weight coefficient, adjust the score result; the score is finally mapped to between zero and one hundred, and enters the risk level classification module. Through the correspondence between the score and the five-level risk standard interval, automatic determination and level labeling are realized, and each level corresponds to a different response path in the subsequent control logic.

[0114] In a specific example, a batch of fresh corn is evaluated by a real-time quality trend prediction model. The current trend score value is sixty-two, which is at the critical point between "initial risk" and "moderate risk", and the trend curve has a slope of zero point two rising to zero point five in the last three hours, indicating that the deterioration trend is accelerating. It is automatically determined that the batch enters the "moderate risk" interval, triggering the processing logic as follows: issue a suspension delivery instruction; start a microbial sampling plan and notify the operation platform to complete the operation.

[0115] The adaptive reconstruction process of the parameter identification model includes a sensitivity dynamic adjustment module, an identification error feedback module, and an environmental response weight mapping mechanism. After detecting the high-humidity environment trigger, the "fine gray scale spot variation response threshold" in the color identification model is automatically reduced, and the odor intensity change discrimination factor is simultaneously increased. When the environment fluctuation does not recover to the normal state within three consecutive periods, the identification error feedback module is started, and the real judgment result of the previous period is used to adjust the current model prediction boundary.

[0116] This model is different from the quality trend prediction model, which does not directly output the shelf life or risk score, but its identification output will be one of the important inputs of the quality trend prediction model, affecting the accuracy and stability of the latter's judgment result.

[0117] The sensitivity dynamic adjustment module is used to dynamically adjust the threshold value in the identification parameter in real time according to the external environment state. The environment detection mechanism: continuously collects temperature and humidity sensor data. When it is detected that the current humidity value continuously exceeds the predetermined "high humidity threshold" for more than thirty minutes, and the temperature simultaneously rises by more than five degrees Celsius, it is determined that a "high humidity fluctuation event" has occurred.

[0118] Dynamic adjustment response content: after the event is triggered, the key discrimination value in the color identification model for detecting image gray scale variation is adjusted, which specifically includes: automatically reducing the "fine gray scale spot variation response threshold", for example, reducing the gray scale difference judgment value from fifteen units to ten units; adjusting the sensitivity of the image noise filter to enhance the response capability of low contrast areas; at the same time, increasing the discrimination factor weight of odor intensity change, specifically amplifying the fluctuation amplitude value of the current odor response vector, making it more influential on the trend modeling output. The adjustment of the above-mentioned "response threshold" and "discrimination factor" can be automatically completed by setting a function model. For example, in the embodiment of the present application, a Sigmoid function form can be used to establish a threshold adjustment function, so that the response value adjustment is more sensitive near the critical point of environmental parameter change. Linear coefficient incremental model, step function, etc. can also be used to realize the dynamic adjustment behavior.

[0119] The environmental response weight mapping mechanism is used to automatically adjust the weight proportion of multi-channel perception data in the identification process according to the external environmental disturbance. The implementation path: a correlation mapping model between each channel data (image, gas, humidity, microorganism, etc.) and different environmental factors is preset. For example, the response weight of the odor channel to humidity rise can be defined as a linear amplification curve, and the response of the image channel to vibration can be set as a suppression function.

[0120] Response logic: When entering a high humidity or vibration fluctuation state, automatically adjust the participation weight of different channels in the recognition module according to the mapping model. For example: in a high humidity state, the odor channel recognition weight is increased by twenty percentage points; the image edge recognition weight is slightly suppressed to avoid image blur misjudgment caused by water vapor. The mapping mechanism can be realized by weight matrix setting, Bayesian network modeling or fuzzy logic control, and the above methods can be used instead, without being limited to a specific method.

[0121] Recognition error feedback module and prediction boundary adjustment mechanism: When the environmental fluctuations have not recovered to the normal state for three consecutive periods, start the recognition error feedback module, and adjust the prediction boundary of the current model based on the true judgment result of the previous period.

[0122] Trigger condition description: Define "continuous abnormal period" as the environmental parameters being in a high-risk range for three or more periods, and the recognition model judgment result deviating from the true state such as manual sampling, laboratory feedback, etc.

[0123] Boundary adjustment logic: Calculate the error deviation value between the prediction result of the previous period of the model and the actual judgment label; if the deviation value is higher than the set error tolerance rate (such as one tenth), it is judged as "recognition instability"; the model boundary adjustment includes: lowering the current recognition decision threshold, narrowing the classification boundary distance, and widening the abnormal definition trigger range.

[0124] For example, for the boundary classifier in the image recognition model, the original normal judgment boundary of eighty can be adjusted to eighty-five to reduce the possibility of missed judgment. The adjustment method can use adaptive boundary regression algorithm, boundary buffer strategy or introduce positive and negative sample weight redistribution mechanism.

[0125] Relevance to the quality trend prediction model in the present application: The recognition boundary change will affect the weight or credibility of multi-channel data in trend modeling, thereby indirectly affecting the final shelf life prediction and risk score result.

[0126] In a specific example, deployed in a long-distance transportation cold chain, encountering continuous high humidity environment, humidity from sixty-five to eighty-five, lasting more than one hour. The image recognition model originally recognizes mold based on gray spot change, with a default threshold of fifteen, which is automatically adjusted to ten; the odor recognition fluctuation sensitivity coefficient is amplified to twice the original value. The judgment accuracy rate is reduced to less than eighty-five in three consecutive periods (each period is thirty minutes), the error feedback mechanism is started, the sample is "initial mold", but the model judges "stable", the deviation is obvious. Therefore, the current recognition boundary is adjusted from eighty-five to ninety to improve the judgment sensitivity. This adjustment immediately affects the quality trend prediction model result of the next period, and the model risk score of the next batch of corn is increased from sixty-five to seventy, and the judgment level is changed to "initial risk", successfully triggering the pre-cooling measure.

[0127] Data cross analysis adopts a vector similarity matching and time series difference fusion mechanism. Feature vectors are constructed for image, gas, temperature and humidity data channels, and an omnidirectional data cross-sectional atlas is formed for each batch of fresh corn within a time period. If the similarity index between the atlases decreases and is accompanied by a predicted risk score close to the quality abnormality trigger threshold, the batch is automatically marked as a "key monitoring object". Such batches will trigger a secondary quality inspection mechanism and data refresh training before entering the warehouse or transportation process, to enhance the model's ability to accurately identify boundary samples.

[0128] Multi-channel feature vector construction: For each batch of fresh corn within a given time period (such as twenty-four hours of continuous detection records), multi-dimensional features are collected from image, gas, temperature and humidity channels. For example:

[0129] Image channel: such as color mean, gray scale distribution variance, edge texture complexity, etc.

[0130] Gas channel: such as odor response intensity, component ratio;

[0131] Temperature and humidity channel: such as temperature standard deviation, humidity fluctuation gradient, etc.

[0132] The above channel features are summarized into a high-dimensional vector at the same time point, and each dimension corresponds to a processed physical or chemical index value.

[0133] Cross-sectional atlas generation: The multi-dimensional vector sequence of the continuous period is spliced to construct the omnidirectional data cross-sectional atlas of the batch product within the period. Essentially, it is a two-dimensional matrix of time series-channel index, where the horizontal axis is time and the vertical axis is the index extracted by different data channels. This atlas is the basis for subsequent similarity matching and anomaly recognition.

[0134] Vector similarity matching and time series difference fusion mechanism: Vector similarity matching: The current batch atlas is matched with the historical healthy sample atlas in the vector space. Any of the following methods can be used to calculate the similarity: Cosine similarity: used to measure the angle between the current vector direction and the historical standard vector. The closer the value is to one, the more similar it is. Euclidean distance: used to measure the distance between the current sample and the historical sample in the feature space. Dynamic time warping: suitable for overall similarity matching of time series atlases, considering time shift problems. In this invention, it is not specified which algorithm to use, as long as similarity measurement can be achieved, and all can be replaced.

[0135] Temporal difference fusion: Based on vector comparison, further calculate the time difference trend between the current atlas and the past one period atlas, extract the change rate, direction and fluctuation intensity, etc. To determine whether the current batch appears "special deviation". The fusion mechanism reflects the potential risk changes of product quality more comprehensively by superimposing the "spatial difference" and "time drift" two-dimensional indicators.

[0136] "Key monitoring object" determination logic and operation response: When it is identified that the atlas similarity index is lower than the preset similarity threshold (such as eighty percent), and the risk score value output by the current quality trend prediction model is close to the quality abnormality trigger threshold (such as only two point score difference from moderate risk), the batch sample is automatically marked as "key monitoring object". The key monitoring object will trigger the following measures: interrupt the original storage or transportation process; enter the secondary quality inspection mechanism for manual review and sample inspection; re-enter the batch data into the quality trend prediction model of the application for data refresh training, to improve the model's judgment ability in the boundary sample area and enhance the model's identification accuracy in the "fuzzy interval".

[0137] Linkage mechanism with quality trend prediction model: The atlas and similarity calculation output do not directly generate a risk score, but are transmitted as auxiliary input factors into the quality trend prediction model defined in the application, as the basis for the following functions: adjusting the model input sample confidence parameter; starting the model internal abnormal deviation discrimination logic (such as dynamically adjusting the trend slope determination range); activating the model's retraining mechanism to increase the number of boundary samples and improve the model's generalization ability. All batches marked as "key monitoring objects" will be fed back to the model input structure through the data interface, forming a complete logical chain of "perception → identification → backflow training → identification ability improvement".

[0138] In a specific example, a batch of fresh corn changes dramatically in humidity during the transition from room temperature to cold storage. The collected image features and gas features form a multi-dimensional vector, and the cosine similarity with the standard sample is only seventy-five percent, which is lower than the threshold. At the same time, the quality trend prediction model in the application outputs a risk score value of sixty-six, which is only one point away from the "moderate risk" boundary. Immediately mark this batch as a key monitoring object: stop the transportation plan; arrange a special person for re-inspection; send the batch sample data into the training set to refresh the model boundary learning parameters and improve the subsequent identification accuracy.

[0139] The centralized management platform displays the postharvest processing, packaging form, transportation path and quality change trend of different batches of fresh corn in a graphical manner based on a visual chain structure. Each processing prediction cycle records the current input data state and output prediction result and matches them with the actual quality feedback. If the prediction deviation amplitude of two consecutive cycles is greater than the prediction accuracy stability threshold, the platform automatically marks the current model for retraining and selects the latest batch of stable samples as the training base.

[0140] The visual chain structure refers to organizing the whole process information from different batches of fresh corn in a graphical manner in a traceable data chain. The structure logic corresponds to the actual flow process of fresh corn, including the following nodes:

[0141] Postharvest processing information: operation time and parameters such as cleaning, sterilization and cooling;

[0142] Packaging form information: types of packaging materials, barrier properties, packaging pressure, etc.

[0143] Transportation path record: logistics centers passed, transportation temperature and humidity trajectory, vibration record;

[0144] Quality change trend: trend score curve, risk level determination result and predicted shelf life output generated by the quality trend prediction model defined in the present application. The above nodes are logically organized into a chain structure through time stamp, batch number and spatial position. The chain is displayed to the platform user in a graphical interface, realizing integrated tracking, analysis and management of data.

[0145] Data recording and feedback matching processing logic: record the current input data state and output prediction result and match them with the actual quality feedback after each processing prediction cycle.

[0146] When each quality prediction behavior is executed, the platform will record the following contents:

[0147] Input data state: current feature vector including image channel, gas channel, temperature and humidity channel, and microbial channel;

[0148] Model output: trend score value, predicted shelf period and risk level generated by the quality trend prediction model described in the present application;

[0149] Actual quality feedback: real quality state label from terminal sampling, consumer feedback or artificial sensory evaluation.

[0150] The matching behavior can be realized by simple label comparison, score difference analysis or error distribution evaluation, etc. For example, using the score difference method, the maximum allowed error (such as five percent score unit) between the predicted value and the measured value can be set to judge whether the current model is stable in prediction.

[0151] Prediction accuracy judgment criteria and retraining trigger mechanism: If the prediction deviation amplitude is greater than the prediction accuracy stability threshold for two consecutive periods, the platform automatically marks the current model for retraining.

[0152] Definition of prediction deviation: The prediction deviation is the numerical difference between the model output score and the true score. For example, if the model predicts that the shelf life is thirty hours, but the actual product becomes moldy after twenty hours, the deviation is ten hours.

[0153] Stability threshold explanation: The stability threshold can be set according to the historical model performance, such as taking the standard deviation of the past one hundred batches of prediction errors multiplied by a proportionality coefficient, or setting a fixed threshold (such as the error should not exceed ten percent).

[0154] Continuous deviation determination: The determination of two consecutive periods refers to the situation that the error threshold is exceeded in the same batch of products in two consecutive prediction behaviors, which is considered as the decline of the short-term prediction stability of the model, triggering the self-learning mechanism.

[0155] Selection logic of retraining samples and model update: The platform automatically marks the current model for retraining and selects the latest batch of stable samples as the training base.

[0156] Sample selection method: From the latest samples that do not appear deviation, risk level is "stable state" or "fluctuation warning" in the current period, the batches with complete data and clear evaluation are selected as the training base.

[0157] The training base can include: input feature vector (including image, gas, environment and microbial characteristics); quality trend score sequence; actual shelf life record or manual label.

[0158] Model update process: The platform sends the above training samples into the quality trend prediction model in the invention and performs incremental training or complete retraining. The training method can include:

[0159] Parameter iteration based on sliding window method; fine-tuning the model backend using transfer learning method; replacing the feature embedding layer structure to adapt to new packaging or processing conditions. The platform can set the training trigger frequency, such as automatically detecting the deviation state once a day, and once the cumulative number of marked batches reaches the threshold (such as three batches), the training behavior is forced to trigger.

[0160] In a specific example, batch number Z-123 is scored as seventy-five before transportation, with a risk level of "initial risk" and a predicted shelf life of forty-eight hours. However, through on-site sampling, mold occurs at thirty hours, with a prediction error of eighteen hours, exceeding the platform's maximum error threshold of ten hours.

[0161] The deviation is recorded as an unstable event. Then, another batch (No. Z-124) is predicted again in the next cycle, and the predicted shelf life is fifty hours, but the actual measured shelf life is only twenty-eight hours, with a deviation of twenty-two hours. Since the deviation exceeds the threshold for two consecutive times, it is automatically determined that the current model prediction ability has decreased. The platform immediately performs the following operations: marks the current model as "to be optimized"; automatically selects the data of stable state batches Z-127 and Z-130 in the past day as a new training set; triggers model retraining, and records the update log in the chain graph for user auditing.

[0162] The above formulas are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of collected data to obtain a formula of the latest real situation. The preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0163] The present application builds an intelligent quality monitoring method with multi-modal perception, trend modeling, intelligent judgment and dynamic control as the core around the need for mildew prevention and quality preservation of fresh corn from postharvest to circulation. First, multi-dimensional feature input is constructed through image acquisition, gas detection, humidity and temperature perception and rapid microbial identification to form a full data set that can be used for intelligent modeling. The system uses multi-level time series modeling to construct a quality degradation trend curve, predicts the shelf life and outputs a risk rating score, and simultaneously adjusts parameters such as image color recognition sensitivity and odor recognition threshold to improve the recognition accuracy of early "hidden mildew".

[0164] Based on the operation of the quality trend prediction model, the present application introduces a data cross-analysis mechanism and an abnormal sample identification logic, uses vector similarity and time series difference algorithm to identify potential boundary samples, and triggers secondary quality inspection and data refresh training. At the same time, the system constructs a five-level risk level decision system according to the comparison of the prediction result and the actual feedback, automatically matches the corresponding control strategy for different levels, including packaging replacement, circulation suspension, cold chain reinforcement and other operations, and realizes the transformation from "passive monitoring" to "active intervention".

[0165] To realize intelligent closed-loop management throughout the whole process, the present application builds a visual chain traceability structure on the platform side, records the processing path and quality trend of each batch of corn. When the continuous prediction deviation exceeds the stable threshold, the platform automatically identifies the model degradation state, triggers model retraining and selects the latest stable sample to update the learning basis, realizes the self-evolution of the quality trend prediction model and the sustainable improvement of the long-term accuracy of the system.

[0166] It should be understood that the size of the sequence number of each process described above in various embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0167] Those skilled in the art can clearly understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are executed in hardware or software depends on specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0168] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described devices and units can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0169] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for preventing mold and preserving fresh corn circulation based on intelligent monitoring, characterized in that: The following steps are involved: By configuring an image acquisition module, a gas detection module, a humidity and temperature sensor module, and a rapid microbial detection device, multi-source quality data of fresh corn can be collected simultaneously at different distribution links, including color, skin texture, gas volatile concentration, temperature and humidity changes, and pathogen metabolism indicators. Using a quality trend prediction algorithm based on time series modeling, the above multi-dimensional data is trained sequentially to construct a quality degradation trend curve, automatically determine the degree of quality deterioration and future shelf life, and assign a risk level score; When high humidity fluctuations or shocks occur in the storage and transportation environment, the color recognition sensitivity, gas warning threshold, and microbial detection frequency are adjusted in real time to achieve adaptive reconstruction of the parameter recognition model to improve the recognition accuracy of the early stages of mold; A multi-dimensional cross-analysis is performed on the multimodal data of each batch of fresh corn. When the risk level predicted by the identification model exceeds the standard warning value, a circulation suspension signal is triggered, and potentially latent moldy batches are prioritized to prevent products that appear normal but are internally spoiled from entering the sales process. Upload the collected data, prediction results and processing records at each stage to the centralized management platform through the data interface, and automatically archive them to form a quality traceability chain; The gas detection module is built based on the principle of an electronic nose and uses multi-channel metal oxide gas sensors to identify volatile organic compounds in the corn storage environment, including aldehydes, ketones, and low-molecular-weight fatty acid components in mold metabolites. Odor pattern recognition is achieved by constructing characteristic response vectors based on different odor intensity combinations. A sensitivity coefficient for spoilage odor shift is also set. When this coefficient value exceeds the gas perception sensitivity response warning threshold, the prediction model's mold progression weight is adjusted, assigning a high-risk tendency to recent data segments to achieve early identification of mold paths. The spoilage odor shift sensitivity coefficient is calculated using any one of three discrimination algorithms: Mahalanobis distance, Manhattan distance, and Shannon entropy. The rapid microbial detection device uses a non-destructive detection mechanism based on impedance or bioluminescence to capture metabolic activity indicators of living colonies on the surface of corn kernels or inside packaging. Before the test data enters the prediction model, it first performs high-frequency sample fluctuation difference fitting. When the difference value exceeds the microbial metabolic trend deviation threshold, it enters the critical sample backtracking stage for historical curve comparison. This comparison uses one of two strategies: moving average fitting and perturbation trend regression to determine whether there is a significant deviation event. When the data in the deviation range is within three consecutive periods, the prediction model will raise the alert level of the data batch from low risk to medium-high risk. The quality trend prediction algorithm adopts a multi-level time series modeling strategy, combining short-term trend windows with periodic trend windows for dual evolution analysis. A causal weight optimization mechanism is introduced at the core of the model. This mechanism adaptively and dynamically adjusts the weight ratio of the data involved in the modeling based on the strength of the causal relationship between different data channels. Data cross-analysis adopts a mechanism based on vector similarity matching and time series difference fusion. Feature vectors are constructed for image, gas, and temperature and humidity data channels respectively, and an omnidirectional data cross-section map is formed for each batch of fresh corn within a time period. If the similarity index between the maps is detected to decrease and the predicted risk score is close to the quality abnormality trigger threshold, the batch will be automatically marked as a key monitoring object. Such batches will trigger a secondary quality inspection mechanism and undergo data refresh training before entering the warehousing or transportation process to enhance the model's ability to accurately identify boundary samples.

2. The method for preventing mold and preserving fresh corn based on intelligent monitoring according to claim 1, characterized in that: The image acquisition module includes a high-resolution image sensor and a near-infrared reflection unit. The reflection image of a preset band is used to extract the subtle color difference of the corn kernel skin and the change signal of the local wet spot area. When there is no obvious abnormality in normal visible light detection, the potential signs of mold are determined through spectral enhancement feature recognition. The image data undergoes color space conversion and region segmentation operations in the preprocessing stage, converting the image from the red, green and blue color space into the brightness, saturation and hue space, and combining the edge gradient change with the color difference distribution mean to perform dynamic difference calculation, thereby improving the probability of identifying early mold micro-clusters.

3. The method for preventing mold and preserving fresh corn based on intelligent monitoring according to claim 2, characterized in that: When the shelf life is predicted, the risk level is divided into five levels, namely stable state, fluctuation warning, initial risk, moderate risk and high-risk spoilage state. Each level corresponds to different warning response and disposal strategies.

4. The method for preventing mold and preserving fresh corn based on intelligent monitoring according to claim 3, characterized in that: The adaptive reconstruction process of the parameter recognition model refers to automatically lowering the response threshold of subtle grayscale spot variations in the color recognition model after detecting a high-humidity environment trigger, and at the same time improving the discrimination factor of odor intensity changes; when the environmental fluctuations do not return to normal within three consecutive cycles, the current model prediction boundary is adjusted based on the actual judgment results of the previous cycle.

5. The method for preventing mold and preserving fresh corn based on intelligent monitoring according to claim 4 is characterized in that: The centralized management platform uses a visual chain structure to graphically display the post-harvest processing, packaging form, transportation route and quality change trends of different batches of fresh corn. Each time a prediction cycle is processed, the current input data status and output prediction results are recorded and matched with the actual quality feedback. If the prediction deviation is greater than the prediction accuracy stability threshold for two consecutive cycles, the platform automatically marks the current model as requiring retraining and selects the latest batch of stable samples as the training basis.

Images