An organic vegetable cold chain loss early warning system based on internet of things

Abstract

The application discloses a kind of organic vegetables cold chain loss early warning system of fusion internet of things, and system operation process specifically includes the following steps: collecting cold chain transport environment data;Environment state feature set is generated based on cold chain transport environment data;Based on environment state feature set and preset loss model, calculate current loss risk level;If current loss risk level is high risk level, generate multi-level early warning signal and send to user terminal.The application has the following advantages and effects: through dynamic weight distribution, principal component feature fusion and hybrid model collaborative prediction, the coupling effect of multiple environmental parameters is accurately quantified, effectively solve the problem of insufficient evaluation accuracy of multiple parameter dynamic correlation loss caused by static threshold and fixed weight in traditional method.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of cold chain logistics, in particular to an organic vegetable cold chain loss early warning system fusing Internet of Things. BACKGROUND

[0002] In the process of organic vegetable cold chain transportation, abnormal fluctuations in environmental parameters will directly lead to deterioration of vegetable quality and increase in loss rate. In the prior art, the traditional monitoring system usually adopts a static threshold early warning mechanism of a single environmental parameter, for example, only triggers an alarm when the temperature exceeds a fixed range. However, the loss of organic vegetables is the result of the nonlinear coupling of multiple environmental factors, for example, the increase in ethylene concentration in a high humidity environment may accelerate ripening and rotting, and low temperature and insufficient light may cause metabolic imbalance. This multi-factor synergistic effect cannot be accurately characterized by a single parameter threshold or a static model with fixed weights, resulting in the following core problems of existing systems:

[0003] The loss risk assessment precision under the dynamic multi-parameter coupling is insufficient. The existing method does not consider the real-time deviation degree of different environmental parameters and their causal contribution difference to the loss, and there is a high correlation between parameters, such as the synergistic change of carbon dioxide and ethylene concentration, which will interfere with model judgment. At the same time, the traditional static weight distribution is difficult to adapt to the dynamic scene of environmental mutation in the transportation process, resulting in early warning lag or false alarm. This defect is particularly prominent in long-distance, multi-node organic vegetable transportation, which cannot provide reliable decision basis for timely regulation, causing economic loss and resource waste. SUMMARY

[0004] The purpose of the present application is to provide an organic vegetable cold chain loss early warning system fusing Internet of Things to solve the problems raised in the background.

[0005] The above technical purpose of the present application is realized by the following technical scheme:

[0006] An organic vegetable cold chain loss early warning system fusing Internet of Things, the system running process specifically includes the following steps:

[0007] S100, collecting cold chain transportation environment data; wherein the cold chain transportation environment data includes real-time temperature data, real-time humidity data, real-time light intensity data and real-time gas concentration data of the transportation container where the organic vegetables are located, and the real-time gas concentration data includes carbon dioxide concentration data and ethylene concentration data;

[0008] S200, generating an environment state feature set based on the cold chain transportation environment data; wherein the environment state feature set includes a plurality of environment state features and dynamic weights corresponding to the environment state features, and the environment state features are associated with a preset loss threshold range;

[0009] S300, calculate a current loss risk level based on the set of environmental state features and a preset loss model, wherein the preset loss model is generated by training historical transportation data, and the current loss risk level includes a low risk level, a medium risk level, and a high risk level;

[0010] S400, if the current loss risk level is a high risk level, generate a multi-level early warning signal and send it to a user terminal, wherein the multi-level early warning signal includes a first-level early warning signal, a second-level early warning signal, and a third-level early warning signal, the first-level early warning signal is used to prompt an environmental parameter anomaly, the second-level early warning signal is used to recommend an adjustment strategy, and the third-level early warning signal is used to trigger an emergency intervention mechanism.

[0011] By adopting the above technical solution, by constructing a complete cold chain transportation environment data acquisition and multi-level early warning mechanism, dynamic monitoring and intelligent grading response of loss risk in organic vegetable transportation process are realized; by integrating real-time acquisition capability of temperature, humidity, light intensity and gas concentration multi-dimensional environmental parameters, the limitation of traditional monitoring technology only focusing on single parameter is broken through, and key environmental factors affecting vegetable quality are covered; the dynamic weight distribution mechanism automatically adjusts the contribution weight of each feature according to the degree of parameter deviation from the safety threshold, ensuring rapid positioning in the event of environmental anomaly; the three-level early warning system realizes seamless connection from preliminary alarm to emergency intervention through progressive signal triggering mechanism, avoiding the lag problem of traditional single-level early warning due to rigid response strategy; the hierarchical design of multi-level early warning signal combined with real-time interaction of user terminal significantly improves the disposal efficiency and accuracy of transportation environmental abnormal events; therefore, the system can significantly solve the evaluation distortion, response lag and false alarm problems caused by static threshold and fixed weight in the background technology, and provides a high-precision, adaptive and low-latency intelligent loss prevention and control scheme for organic vegetable cold chain transportation.

[0012] Further setting is that S200 specifically includes the following sub-steps:

[0013] S210, according to a preset weight distribution rule, an initial weight is assigned to each environmental state feature; including:

[0014] The historical environmental feature data set includes a plurality of historical environmental state features and corresponding actual loss rates of organic vegetables;

[0015] For each environmental state feature, the causal contribution degree thereof to the actual loss rate of organic vegetables is calculated;

[0016] Based on the causal contribution degree and the information entropy of the environmental state feature, an initial weight corresponding to each environmental state feature is assigned;

[0017] S220, dynamically adjusting the initial weight based on a deviation degree of the real-time environment parameter value and a boundary value of the preset loss threshold range, to generate a dynamic weight; wherein the deviation degree is obtained by calculating the real-time environment parameter value and the boundary value of the preset loss threshold range, and if the deviation degree exceeds a preset deviation threshold, an exponential correction is made to the dynamic weight corresponding to the environment state feature;

[0018] S230, combining each environment state feature and its corresponding dynamic weight to generate an environment state feature set; wherein each environment state feature in the environment state feature set is mapped to a specific interval of the preset loss threshold range and labeled with a distance value of the real-time environment parameter value to the nearest boundary;

[0019] S240, performing redundant feature elimination and feature fusion processing on the environment state feature set; wherein the redundant feature elimination is achieved by a variance filtering method to eliminate features with a variance below a preset variance threshold, and the feature fusion processing combines multiple environment state features with high correlation into a comprehensive feature by a principal component analysis method and recalculates the dynamic weight of the comprehensive feature.

[0020] By adopting the above technical solutions, a data basis is provided for subsequent risk prediction; the initial weight allocation method based on causal contribution degree eliminates the subjective bias of artificial experience weighting and ensures the objectivity of weight allocation by mining the correlation between each environment parameter and loss rate from historical data; the exponential dynamic weight correction algorithm strengthens the sensitivity of the system to critical state parameters, enabling the model to provide early warning when the parameters approach the safety boundary and avoiding the response delay caused by traditional linear correction; the redundant feature elimination mechanism filters out effective features by variance filtering, reduces the interference of data noise on model training, and improves the model convergence speed and generalization ability; the feature fusion processing integrates high-correlation parameters using principal component analysis, solves the multicollinearity problem of multidimensional data, retains the core information of the original data, and optimizes the information density and computational efficiency of the feature set; the boundary distance value labeled in the environment state feature set provides intuitive quantitative basis for subsequent risk level determination, enhancing the explainability of the system output; the dual labeling mechanism of dynamic weight and boundary distance constructs a multidimensional risk assessment framework, supporting comprehensive judgment of the current transportation state from two dimensions of parameter deviation degree and risk contribution degree; the standardized feature preprocessing process ensures the compatibility of different dimension parameters, avoiding the negative impact of numerical scale difference on model learning.

[0021] Further, in the S240, the feature fusion processing combines multiple environment state features with high correlation into a comprehensive feature by a principal component analysis method, and recalculates the dynamic weight of the comprehensive feature, including:

[0022] calculate a Pearson correlation coefficient matrix among all environment state features in the environment state feature set, and screen out environment state feature groups with absolute values of correlation coefficients exceeding a preset threshold, environment state features in each environment state feature group having high linear correlation;

[0023] perform standardization processing on the environment state features in each environment state feature group, to eliminate dimensional influence;

[0024] perform principal component analysis, and extract a first principal component as a comprehensive environment state feature;

[0025] sum dynamic weight values of the environment state features in each environment state feature group to obtain an original weight sum, the dynamic weight value of the comprehensive environment state feature being a product of a variance contribution rate and the original weight sum;

[0026] add the generated comprehensive environment state feature, the dynamic weight value thereof, and a distance value from a nearest boundary of a preset loss threshold range to the environment state feature set, and remove the original environment state features that are merged from the environment state feature set.

[0027] By adopting the technical solution, high-correlation environment features are deeply integrated through principal component analysis, and information density and calculation efficiency of the feature set are optimized; the Pearson correlation coefficient matrix screening mechanism can identify feature groups with strong linear correlation, avoid the influence of invalid feature combinations on dimension reduction effect, and improve the pertinence of feature fusion; the standardization processing eliminates data distribution differences of different dimensional parameters, ensures mathematical rationality of the principal component extraction process, and prevents principal component deviation caused by dimensional differences; the first principal component, as a comprehensive feature, retains variance information of original data to the greatest extent, reduces information loss, realizes feature dimension reduction, and solves the dimension problem caused by high-dimensional data; the dynamic weight inheritance rule calculates a new weight through a product of a variance contribution rate and an original weight sum, maintains the continuity of feature importance evaluation, and ensures that risk representation capability of the fused feature is not weakened due to data reconstruction; migration labeling of the boundary distance value guarantees the continuity of risk distance calculation after feature fusion, avoids a risk assessment fault caused by feature reconstruction, and maintains logical consistency of risk level determination.

[0028] Further, the S300 specifically includes the following sub-steps:

[0029] S310, input the environment state feature set into the preset loss model; wherein, the preset loss model includes a regression prediction module constructed based on a deep neural network and a classification decision module constructed based on a random forest;

[0030] S320, weight fusion is performed on each original environment state feature in the set of environment state features and the comprehensive environment state feature generated by principal component analysis based on the dynamic weight, to generate a weighted feature vector; wherein the elements of the weighted feature vector include the product of the real-time observation value of the original environment state feature and the comprehensive environment state feature and the corresponding dynamic weight;

[0031] S330, input the weighted feature vector into the regression prediction module, and calculate the real-time loss rate prediction value by forward propagation;

[0032] S340, input the real-time loss rate prediction value into the classification decision module, and determine the current loss risk level according to the preset risk division rule;

[0033] S350, based on the distance value between the real-time observation value of each original environment state feature and the comprehensive environment state feature in the set of environment state features and the nearest boundary of the preset loss threshold range, the confidence of the current loss risk level is obtained.

[0034] By adopting the above technical scheme, the regression module of the deep neural network captures the nonlinear relationship between the environment parameters and the loss rate, solves the problem of insufficient modeling ability of traditional linear models for complex interaction effects, and realizes high-precision fitting of multi-factor coupling. The random forest classification module improves the stability and anti-interference ability of risk level determination through multi-decision tree integration, reduces the sensitivity of a single model to noise data, and enhances the robustness of the classification result; the weighted feature vector generation mechanism strengthens the decision-making influence of high-weight features, ensures that key abnormal parameters play a dominant role in the prediction process, and improves the identification accuracy of the model for core risk sources; the confidence calculation module combines feature boundary distance and dynamic weight to quantify the credibility of the risk assessment result, provides a reference basis for artificial review, and balances the synergistic effect of automatic decision and manual intervention; the real-time loss rate prediction value and the risk level classification result output by the model form a complement, and a multi-dimensional risk situation awareness system is constructed; distributed training meets the real-time monitoring needs of large-scale cold chain transportation networks; the cooperative mechanism of forward propagation and decision tree voting realizes the balance between prediction efficiency and accuracy, and ensures the stable operation of the system in a high-concurrency scenario.

[0035] Further setting is, in the S340, the preset risk division rule is:

[0036] If the real-time loss rate prediction value ≤T low , the current loss risk level is low risk level;

[0037] If T low < real-time loss rate prediction value ≤T high , the current loss risk level is medium risk level;

[0038] If the real-time loss rate prediction value > T high , the current loss risk level is a high risk level;

[0039] wherein, T low and T high are threshold values set by the percentiles of historical transportation data, corresponding to the boundary values of the low risk level and the high risk level respectively.

[0040] By adopting the above technical solution, the dynamic adaptation of the risk level is realized through the data-driven threshold division method, and the generalization ability of the system to different transportation scenarios is enhanced; the threshold setting strategy based on the percentiles of historical data ensures that the risk level boundary is highly consistent with the loss distribution characteristics of the real business scenario, avoiding the evaluation deviation caused by subjective threshold setting; the quantile threshold calibration mechanism allows the system to automatically update the judgment standard as the historical data accumulates, realizing the dynamic optimization of the threshold parameters and adapting to the data distribution changes brought by the transportation environment or equipment upgrade.

[0041] Further setting is that in the S320, the weighted fusion process is as follows:

[0042]

[0043] wherein, F weighted is a weighted feature vector; X k,real represents the real-time observation value of the kth original environment state feature; W dynamic,k is the dynamic weight of the kth original environment state feature; Z z is the zth standardized comprehensive environment state feature. represents the dynamic weight of the zth comprehensive environment state feature; m is the total number of original environment state features; p is the total number of comprehensive environment state features.

[0044] By adopting the above technical solution, the product calculation rule of the dynamic weight and the real-time observation value realizes the dual expression of the parameter value influence and the risk contribution degree, ensuring that the model can simultaneously perceive the absolute numerical change of the parameter and its relative importance in the current environment; the weighted accumulation operation of the original environment state feature and the comprehensive environment state feature realizes the unified representation of heterogeneous data, ensuring that the feature vector after fusion completely carries the multi-dimensional environment state information, avoiding the performance degradation of the model caused by information loss; the standardized comprehensive feature eliminates the negative impact of dimension difference on model training, improves the convergence efficiency of the gradient descent algorithm, and speeds up the model training process.

[0045] Further setting is that in the S350, the calculation process of the confidence degree is as follows:

[0046]

[0047] wherein, Distance k is the minimum distance between the real-time observation value of the kth original environmental state feature and the boundary of the preset loss threshold range; is the minimum distance between the standardized value of the zth comprehensive environmental state feature and the boundary of the preset loss threshold range; Confidence represents the confidence.

[0048] By adopting the above technical solutions, the weighted distance summation mechanism comprehensively reflects the overall situation of each feature deviating from the safety boundary, avoiding the misjudgment risk caused by single parameter anomaly, and improving the comprehensiveness and objectivity of risk assessment; The design of dynamic weight participating in calculation gives greater influence to high importance features, so that the confidence index and the actual risk severity maintain strong correlation, ensuring that the confidence significantly decreases when high-weight parameter anomaly occurs; The multi-dimensional fusion characteristics of the confidence index enhance the system's representation ability for complex risk scenarios, for example, it can still accurately reflect the credibility of the overall risk when multiple parameters deviate slightly.

[0049] Further, the S400 specifically includes the following sub-steps:

[0050] S410, determining whether the current loss risk level is a high risk level; if it is determined to be a high risk level, triggering a multi-level early warning signal generation process, otherwise returning to S300 for continuous monitoring;

[0051] S420, generating a first-level early warning signal, including:

[0052] Based on the deviation degree of the real-time observation value of each original environmental state feature and comprehensive environmental state feature in the environmental state feature set from the preset loss threshold range, screening out the original environmental state feature or comprehensive environmental state feature whose deviation degree exceeds the first deviation threshold but does not reach the second deviation threshold , generating a primary abnormal parameter list;

[0053] The screening result is packaged as a first-level early warning signal and sent to the user terminal through the wireless communication module;

[0054] S430, generating a second-level early warning signal, including:

[0055] If the deviation degree of at least one original environmental state feature and / or comprehensive environmental state feature in the primary abnormal parameter list exceeds the second deviation threshold or the confidence of the real-time loss rate prediction value is lower than the first preset confidence threshold Confidence1, a second-level early warning is triggered;

[0056] Matching the adjustment strategy corresponding to the out-of-limit feature from the preset adjustment strategy library, and prioritizing according to its dynamic weight to generate a recommended strategy set;

[0057] The recommended strategy is encapsulated as a secondary early warning signal, which is sent to the user terminal with high priority;

[0058] S440, generating a tertiary early warning signal, comprising:

[0059] If the deviation degree of a certain original environment state feature and / or comprehensive environment state feature in the primary abnormal parameter list continues to exceed the third deviation threshold If the preset time length or the confidence of the real-time loss rate prediction value is lower than the second preset confidence threshold Confidence2, a tertiary early warning is triggered;

[0060] A tertiary early warning signal is generated and a forced control instruction is sent through a redundant communication link.

[0061] The above technical solution realizes gradient identification of risk severity through a deviation threshold layering mechanism, ensures that the early warning level strictly matches the actual situation, avoids the problems of insufficient response or excessive control, realizes closed-loop management from abnormal discovery to forced intervention through a progressive early warning process, controls the disposal cost through a progressive upgrading mechanism, and avoids triggering excessive response for low-risk events.

[0062] Further, in the S400 substep, the following is provided:

[0063] The first deviation threshold The second deviation threshold And the third deviation threshold Satisfy the following hierarchical relationship:

[0064]

[0065] The first preset confidence threshold Confidence1 and the second preset confidence threshold Confidence2 satisfy the following hierarchical relationship:

[0066] The second preset confidence threshold Confidence2 is less than the first preset confidence threshold Confidence1.

[0067] In summary, the present application has the following beneficial effects:

[0068] Through dynamic weight distribution, principal component feature fusion, and collaborative prediction of a hybrid model, the coupling effect of multiple environment parameters is accurately quantified, effectively solving the problem of insufficient evaluation accuracy of dynamic correlation loss of multiple parameters caused by static threshold and fixed weight in traditional methods. BRIEF DESCRIPTION OF DRAWINGS

[0069] Fig. 1 The main flowchart of the embodiment;

[0070] Fig. 2A schematic diagram of a main flow for an embodiment;

[0071] Fig. 3 A schematic diagram of a flow for a preset risk classification rule in S340 in an embodiment;

[0072] Fig. 4 A schematic diagram of a flow for S400 in an embodiment. DETAILED DESCRIPTION

[0073] The application will be further described in detail below with reference to the accompanying drawings.

[0074] As shown in the accompanying Figs. 1 to 4 ;

[0075] The embodiment discloses an organic vegetable cold chain loss early warning system fusing Internet of Things, and a system running process specifically comprises the following steps:

[0076] S100, collecting cold chain transportation environment data; wherein the cold chain transportation environment data comprises real-time temperature data, real-time humidity data, real-time light intensity data and real-time gas concentration data of a transportation container where the organic vegetables are located, and the real-time gas concentration data comprises carbon dioxide concentration data and ethylene concentration data;

[0077] The internal environment parameters of the transportation container are acquired in real time through an Internet of Things sensor array; wherein the Internet of Things sensor array comprises a temperature sensor, a humidity sensor, a light sensor and a gas concentration sensor, and the gas concentration sensor is configured to synchronously detect carbon dioxide concentration and ethylene concentration;

[0078] The internal environment parameters are uploaded to a cloud server through a wireless communication module; wherein the wireless communication module supports at least one of 5G network, LoRa protocol and NB-IoT protocol communication modes;

[0079] The internal environment parameters are subjected to denoising processing and normalization processing in the cloud server to obtain standardized cold chain transportation environment data; the denoising processing and normalization processing belong to a conventional data processing mode, which is a prior art field and will not be described here.

[0080] Embodiment 1

[0081] The transportation container is a refrigerated container with a multi-layer temperature insulation structure, and an Internet of Things sensor array is installed inside; the Internet of Things sensor array comprises:

[0082] A PT100 platinum resistance temperature sensor is disposed at the top, middle and bottom of the refrigerated container, with an interval of 2 meters between each layer, and a total of 6 measuring points;

[0083] A capacitive humidity sensor of model SHT35 is installed at the ventilation port of the vegetable stacking area;

[0084] BH1750 digital light intensity sensor, deployed inside the transparent observation window of the refrigerated container, to avoid external light interference;

[0085] Multi-channel gas detection module model SGP41, synchronously detects carbon dioxide concentration and ethylene concentration, installed at the outlet of the refrigerated container internal circulation air duct.

[0086] The main link uses 5G network, and the secondary link uses NB-IoT. When the 5G signal strength is lower than -90dBm, it automatically switches to NB-IoT. The data upload frequency is every 5 minutes, and each transmission contains the original readings of all sensors and time stamps.

[0087] In the cloud server, the data preprocessing process includes: using wavelet threshold denoising for temperature data, using moving average filtering for humidity data, and using median filtering for light and gas data.

[0088] Then, using Min-Max normalization, each parameter is mapped to the interval [0, 1].

[0089] S200, generating an environment state feature set based on cold chain transportation environment data; wherein the environment state feature set includes multiple environment state features and dynamic weights corresponding to the environment state features, and the environment state features are associated with a preset loss threshold range;

[0090] S300, calculating the current loss risk level based on the environment state feature set and the preset loss model; wherein the preset loss model is generated by training historical transportation data, and the current loss risk level includes low risk level, medium risk level and high risk level;

[0091] S400, if the current loss risk level is high risk level, generating a multi-level warning signal and sending it to the user terminal; wherein the multi-level warning signal includes a first-level warning signal, a second-level warning signal and a third-level warning signal, the first-level warning signal is used to prompt the environment parameter anomaly, the second-level warning signal is used to recommend adjustment strategy, and the third-level warning signal is used to trigger the emergency intervention mechanism.

[0092] Specifically, S200 specifically includes the following sub-steps:

[0093] S210, according to the preset weight distribution rule, assigning an initial weight to each environment state feature; including:

[0094] Extracting a historical environment feature data set from a historical transportation database, the historical environment feature data set including multiple historical environment state features and their corresponding actual loss rates of organic vegetables;

[0095] For each environment state feature, calculate its causal contribution to the actual loss rate of organic vegetables;

[0096] The calculation process of the causal contribution degree is as follows:

[0097]

[0098] wherein C k is the causal contribution degree, representing the causal influence intensity of the kth environmental state feature on the organic vegetable loss rate; X k,i represents the observation value of the kth environmental state feature in the ith transportation; which represents the deviation of the observation value of the kth environmental state feature from its historical mean value represents the loss rate in the ith transportation predicted by the gradient boosting tree model; which represents the deviation of the actual loss rate Y i from its historical mean value represents the partial derivative of the predicted loss rate with respect to the feature X k,i calculated by the automatic differentiation technique; n represents the total number of historical transportation data samples.

[0099] Based on the causal contribution degree and the information entropy of the environmental state feature, an initial weight corresponding to each environmental state feature is assigned;

[0100] The assignment process of the initial weight is as follows:

[0101]

[0102] wherein, which represents the information entropy of the kth environmental state feature, P b is the probability of the feature value falling into the bth bin, and the bin number B is adaptively adjusted according to the feature distribution; W total represents the total weight value; C k is the causal contribution degree; C j represents the cyclic variable of the causal contribution degree; H(X j ) is the information entropy of the jth environmental state feature; W total is the weight sum, and the fixed value is 1; W init,k is the initial weight value assigned to the kth environmental state feature; γ is the entropy value adjustment coefficient.

[0103] Example 2

[0104] Data of 200 transportation tasks in the past year, including environmental parameters and corresponding vegetable loss rates, were extracted from the historical database.

[0105] In one transportation, the average temperature was 4.2℃, corresponding to a historical mean value of 4.5℃; ​​value of 0.15; AX temp value of -0.3°C; AY value of 8.5%-7.2%=1.3%; after calculation, the causal contribution degree C temp value of -0.0346.

[0106] In the information entropy calculation, the temperature data is divided into 5 bins, and the probability of each bin is calculated to calculate H(X temp value of 1.23 bits; the total weight W total is 1, and γ=0.5, then the initial weight W init,temp value is about 0.12.

[0107] S220, based on the deviation of the real-time environmental parameter value and the boundary value of the preset loss threshold range, dynamically adjusting the initial weight to generate a dynamic weight; wherein the deviation is obtained by calculating the real-time environmental parameter value and the boundary value of the preset loss threshold range, if the deviation exceeds the preset deviation threshold, the dynamic weight of the corresponding environmental state feature is exponentially corrected;

[0108] It should be noted that the real-time environmental parameter value refers to the real-time temperature data, real-time humidity data, real-time light intensity data and real-time gas concentration data, specifically the real-time observation value of each environmental state feature;

[0109] The deviation calculation formula is:

[0110]

[0111] Wherein, d k represents the deviation of the kth environmental state feature; U k and L k respectively represent the upper limit and lower limit of the preset loss threshold range of the kth environmental state feature; X k,real represents the real-time observation value of the kth environmental state feature;

[0112] If the deviation d k is greater than the preset deviation threshold d threshold , the initial weight W init,k is exponentially corrected:

[0113]

[0114] Wherein, α is the exponential correction coefficient; W dynamic,k is the dynamic weight value allocated to the kth environmental state feature;

[0115] If the deviation d k is less than or equal to the preset deviation threshold d threshold , the initial weight remains unchanged.

[0116] Embodiment 3

[0117] In a real-time monitoring, the carbon dioxide concentration is 1800ppm, and the preset threshold range is [400, 1500]ppm. The value of the deviation degree calculated after is 300ppm;

[0118] The preset deviation threshold d threshold is 200ppm; and the dynamic weight adjustment is The value of about 0.41, where a takes 0.01.

[0119] S230, combining each environment state feature and its corresponding dynamic weight to generate an environment state feature set; wherein each environment state feature in the environment state feature set is mapped to a specific interval of the preset loss threshold range, and the distance value of the real-time environment parameter value and the nearest boundary is labeled;

[0120] The real-time observation value X k,real of each environment state feature is mapped to the preset loss threshold range [L k , U k ], and the distance value Distance k of the nearest boundary is labeled:

[0121] Distance k = min(|X k,real -L k |,|X k,real -U k |)

[0122] Wherein, Distance k is the minimum distance between the real-time observation value of the kth environment state feature and the boundary of the preset loss threshold range;

[0123] The environment state feature set is generated as follows:

[0124] F dynamic = {(X k,real ,W dynamic,k ,Distance k )|k=1,2,…,m}

[0125] Wherein, F dynamic is the environment state feature set; m is the total number of environment state features.

[0126] S240, the redundant feature elimination and feature fusion processing are performed on the environment state feature set; wherein the redundant feature elimination is realized by variance filtering method, and the features with variance lower than a preset variance threshold are eliminated; the feature fusion processing is realized by principal component analysis method, and multiple environment state features with high correlation are combined into a comprehensive feature, and the dynamic weight of the comprehensive feature is recalculated.

[0127] Specifically, in S240, the feature fusion processing is realized by principal component analysis method, and multiple environment state features with high correlation are combined into a comprehensive feature, and the dynamic weight of the comprehensive feature is recalculated, including:

[0128] A Pearson correlation coefficient matrix between all environment state features in the environment state feature set is calculated, and environment state feature groups with absolute value of correlation coefficient exceeding a preset threshold are screened out, and the environment state features in each environment state feature group have high linear correlation;

[0129] The environment state features in each environment state feature group are normalized to eliminate the dimension influence, and Z-score normalization is specifically adopted;

[0130] Principal component analysis is performed, and a first principal component is extracted as a comprehensive environment state feature;

[0131] The expression is:

[0132]

[0133] Wherein, Z z is the zth normalized comprehensive environment state feature, which is generated by principal component analysis; G z is the zth environment state feature group screened out by Pearson correlation coefficient; k∈G z represents each environment state feature in G z ; β z,k represents the load coefficient of the first principal component in the principal component analysis of the zth comprehensive environment state feature; X k represents the kth environment state feature in G z ;

[0134] The dynamic weight values of the environment state features in each environment state feature group are summed to obtain an original weight sum, and the dynamic weight value of the comprehensive environment state feature is the product of the variance contribution rate and the original weight sum;

[0135]

[0136] Wherein, is the original weight sum in the zth environment state feature group; V z represents the variance contribution rate of the zth comprehensive environment state feature; a dynamic weight representing the zth integrated environmental state feature;

[0137] The generated integrated environmental state feature, its dynamic weight value, and the distance value from the nearest boundary of the preset loss threshold range are added to the environmental state feature set, and the merged original environmental state feature is removed from the environmental state feature set.

[0138] Embodiment 4

[0139] The correlation coefficient r between temperature and CO2 concentration is 0.87, which exceeds the preset threshold value 0.8; after standardization, the first principal component Z1 is extracted, and the variance contribution rate V1 is 78%; the original weight sum is 0.53, and the value of the integrated feature weight is about 0.41 after calculation.

[0140] Specifically, S300 specifically includes the following sub-steps:

[0141] S310, inputting the environmental state feature set into a preset loss model; wherein the preset loss model includes a regression prediction module constructed based on a deep neural network and a classification decision module constructed based on a random forest; the regression prediction module is used to output a real-time loss rate prediction value of the organic vegetables, and the classification decision module is used to divide the current loss risk level according to the real-time loss rate prediction value;

[0142] S320, based on the dynamic weight, weighting and fusing each original environmental state feature in the environmental state feature set and the integrated environmental state feature generated by the principal component analysis method to generate a weighted feature vector; wherein the elements of the weighted feature vector include the product of the real-time observation value of the original environmental state feature and the integrated environmental state feature and its corresponding dynamic weight; the real-time observation value of the integrated environmental state feature is represented by the principal component score;

[0143] S330, inputting the weighted feature vector into the regression prediction module to obtain the real-time loss rate prediction value by forward propagation; wherein the network structure of the regression prediction module includes an input layer, three fully connected hidden layers and an output layer, the activation function adopts LeakyReLU, and the output layer adopts a linear activation function;

[0144] S340, inputting the real-time loss rate prediction value into the classification decision module to determine the current loss risk level according to the preset risk division rule;

[0145] S350, based on the real-time observation value of each original environmental state feature and the integrated environmental state feature in the environmental state feature set and the distance value from the nearest boundary of the preset loss threshold range, obtaining the confidence of the current loss risk level.

[0146] Specifically, in S340, the preset risk division rule is:

[0147] if the real-time loss rate prediction value ≤ T low , the current loss risk level is a low risk level;

[0148] if T low < real-time loss rate prediction value ≤ T high , the current loss risk level is a medium risk level;

[0149] if the real-time loss rate prediction value > T high , the current loss risk level is a high risk level;

[0150] wherein, T low and T high are threshold values set by the percentiles of historical transportation data, corresponding to the boundary values of the low risk level and the high risk level respectively.

[0151] Specifically, in S320, the weighted fusion process is as follows:

[0152]

[0153] wherein, F weighted is a weighted feature vector; X k,real represents the real-time observation value of the kth original environmental state feature; W dynamic,k is the dynamic weight of the kth original environmental state feature; Z z is the zth standardized comprehensive environmental state feature; represents the dynamic weight of the zth comprehensive environmental state feature; m is the total number of original environmental state features; p is the total number of comprehensive environmental state features.

[0154] Specifically, in S350, the calculation process of the confidence is as follows:

[0155]

[0156] wherein, Distance k is the minimum distance between the real-time observation value of the kth original environmental state feature and the boundary of the preset loss threshold range; is the minimum distance between the standardized value of the zth comprehensive environmental state feature and the boundary of the preset loss threshold range; Confidence represents the confidence.

[0157] Specifically, S400 specifically includes the following sub-steps:

[0158] S410, determining whether the current loss risk level is a high risk level; if it is determined to be a high risk level, triggering a multi-level early warning signal generation process, otherwise returning to S300 for continuous monitoring;

[0159] S420, generating a primary early warning signal, comprising:

[0160] Based on the deviation degree of each original environmental state feature and the real-time observation value of the comprehensive environmental state feature from the preset loss threshold range, the original environmental state feature or the comprehensive environmental state feature whose deviation degree exceeds the first deviation threshold But not reaching the second deviation threshold , a primary abnormal parameter list is generated;

[0161] The deviation degree calculation needs to cover all environmental state features, including comprehensive environmental state features, and the preset loss threshold range of which is defined by the following method:

[0162]

[0163] Among them, U z And L z The upper and lower limits of the preset loss threshold range of the zth comprehensive environmental state feature; d z Indicates the deviation degree of the zth comprehensive environmental state feature;

[0164] The screening results are packaged as a primary early warning signal and sent to the user terminal through the wireless communication module;

[0165] S430, generating a secondary early warning signal, comprising:

[0166] If the deviation degree of at least one original environmental state feature and / or comprehensive environmental state feature in the primary abnormal parameter list exceeds the second deviation threshold Or the confidence of the real-time loss rate prediction value is lower than the first preset confidence threshold Confidence1, a secondary early warning is triggered;

[0167] Match the adjustment strategy corresponding to the exceeding feature from the preset adjustment strategy library, and prioritize according to its dynamic weight to generate a recommended strategy set;

[0168] The dynamic weight needs to reflect the exponential correction result in S220 simultaneously: if the deviation degree of a certain environmental state feature triggers weight correction in S220, its dynamic weight directly affects the strategy sorting of the secondary early warning;

[0169] The recommended strategy is packaged as a secondary early warning signal and sent to the user terminal with high priority;

[0170] S440, generating a tertiary early warning signal, comprising:

[0171] If the deviation degree of a certain original environmental state feature and / or comprehensive environmental state feature in the primary abnormal parameter list continuously exceeds the third deviation threshold If the confidence level of the predicted real-time loss rate or the preset duration is lower than the second preset confidence threshold Confidence2, a level 3 warning will be triggered.

[0172] A three-level early warning signal is generated and a mandatory control command is sent through a redundant communication link; the redundant communication link is marked as a redundant communication link to ensure real-time signal transmission.

[0173] Specifically, in the S400 sub-step:

[0174] First deviation threshold Second deviation threshold and the third deviation threshold The following hierarchical relationship must be satisfied:

[0175]

[0176] The first preset confidence threshold Confidence1 and the second preset confidence threshold Confidence2 satisfy the following hierarchical relationship:

[0177] The second preset confidence threshold Confidence2 < the first preset confidence threshold Confidence1.

[0178] Example 5

[0179] When CO2 concentration deviates The value exceeds 300ppm A value of 200ppm generates a list of abnormal parameters;

[0180] If the CO2 deviation further exceeds The value is 250ppm, and the policy is matched with "Activate standby ventilation system" from the policy library. The dynamic weight of 0.41 determines that it is the highest priority policy.

[0181] When the CO2 concentration continues to exceed the preset time for 10 minutes, a command to force the cooling equipment to start is sent through the 5G+NB-IoT dual link.

[0182] Example 6

[0183] In 30 actual transportation tasks, this method, compared with traditional threshold alarms:

[0184] The high-risk underreporting rate decreased from 18% to 3.2%;

[0185] The average early warning response time has been reduced from 5.2 minutes to 1.8 minutes;

[0186] Vegetable spoilage rate decreased by 23.7%.

[0187] The embodiments are only used to explain the present application, and are not used to limit the present application, and any modification without creative contribution made by the person skilled in the art according to the embodiments after reading the specification is protected by the patent law as long as it is within the scope of the claims of the present application.

Claims

1. A fusion Internet of Things (IoT) organic vegetable cold chain loss early warning system, characterized in that, The system operation process specifically comprises the following steps: S100, collecting cold chain transportation environment data; wherein the cold chain transportation environment data comprises real-time temperature data, real-time humidity data, real-time light intensity data and real-time gas concentration data of the transportation container in which the organic vegetables are located, and the real-time gas concentration data comprises carbon dioxide concentration data and ethylene concentration data; S200, generating an environment state feature set based on the cold chain transportation environment data; wherein the environment state feature set comprises a plurality of environment state features and dynamic weights corresponding to the environment state features, the environment state features are associated with a preset loss threshold range; comprising: S210, assigning an initial weight to each environment state feature according to a preset weight distribution rule; comprising: extracting a historical environment feature data set from a historical transportation database, the historical environment feature data set comprising a plurality of historical environment state features and their corresponding actual loss rates of organic vegetables; For each environment state feature, calculate its causal contribution to the actual loss rate of organic vegetables; Based on the causal contribution and the information entropy of the environment state feature, assign an initial weight to each environment state feature; S220, dynamically adjusting the initial weight based on the deviation of the real-time environment parameter value from the boundary value of the preset loss threshold range, and generating a dynamic weight; wherein the deviation is obtained by calculating the real-time environment parameter value and the boundary value of the preset loss threshold range, if the deviation exceeds a preset deviation threshold, the dynamic weight of the corresponding environment state feature is exponentially corrected; comprising: The deviation calculation formula is: wherein, represents a deviation degree of the environmental state feature; and represent an upper limit and a lower limit of a preset loss threshold range of the environmental state feature, respectively; represents a real-time observation value of the original environmental state feature; deviation degree greater than a preset deviation threshold then the initial weight is exponentially corrected: wherein, is an exponential correction factor; is a dynamic weight value assigned to the original environmental state feature; deviation degree less than or equal to a preset deviation threshold then the initial weight is kept unchanged; S230, combining each environment state feature and its corresponding dynamic weight to generate an environment state feature set; wherein each environment state feature in the environment state feature set is mapped to a specific interval of the preset loss threshold range, and the distance value between the real-time environment parameter value and the nearest boundary is labeled; S240, performing redundant feature elimination and feature fusion processing on the environment state feature set; wherein the redundant feature elimination is realized by variance filtering method, and the features with variance below a preset variance threshold are removed, and the feature fusion processing combines multiple environment state features with high correlation into a comprehensive feature by principal component analysis method, and recalculates the dynamic weight of the comprehensive feature; S300, calculating the current loss risk level based on the environment state feature set and a preset loss model; wherein the preset loss model is generated by training historical transportation data, and the current loss risk level comprises a low risk level, a medium risk level and a high risk level; S400, if the current loss risk level is a high risk level, generating a multi-level warning signal and sending it to a user terminal; wherein the multi-level warning signal comprises a first-level warning signal, a second-level warning signal and a third-level warning signal, the first-level warning signal is used to prompt the environment parameter abnormality, the second-level warning signal is used to recommend adjustment strategy, and the third-level warning signal is used to trigger an emergency intervention mechanism.

2. The organic vegetable cold chain loss early warning system integrated with Internet of Things according to claim 1, characterized in that: In the S240, the feature fusion processing combines multiple environmental state features with high correlation into a comprehensive feature through principal component analysis, and recalculates the dynamic weight of the comprehensive feature, including: calculating the Pearson correlation coefficient matrix between all environmental state features in the environmental state feature set, and screening out environmental state feature groups with absolute correlation coefficients exceeding a preset threshold, the environmental state features in each environmental state feature group having high linear correlation; standardizing the environmental state features in each environmental state feature group to eliminate the influence of dimension; performing principal component analysis to extract the first principal component as a comprehensive environmental state feature; summing the dynamic weight values of the environmental state features in each environmental state feature group to obtain an original weight sum, and the dynamic weight value of the comprehensive environmental state feature being the product of its variance contribution rate and the original weight sum; adding the generated comprehensive environmental state feature, its dynamic weight value, and the distance value from the nearest boundary of the preset loss threshold range to the environmental state feature set, and removing the original environmental state features that are combined from the environmental state feature set.

3. The organic vegetable cold chain loss early warning system integrated with Internet of Things according to claim 2, characterized in that: The S300 specifically includes the following sub-steps: S310, input the environmental state feature set into the preset loss model; wherein the preset loss model includes a regression prediction module constructed based on a deep neural network and a classification decision module constructed based on a random forest; S320, based on the dynamic weight, weighting fusion is performed on each original environmental state feature in the environmental state feature set and the comprehensive environmental state feature generated through principal component analysis to generate a weighted feature vector; wherein the elements of the weighted feature vector include the product of the real-time observation value of the original environmental state feature and the comprehensive environmental state feature and their corresponding dynamic weight; S330, input the weighted feature vector into the regression prediction module to calculate the real-time loss rate prediction value through forward propagation; S340, input the real-time loss rate prediction value into the classification decision module to determine the current loss risk level according to a preset risk division rule; S350, based on the real-time observation value of each original environmental state feature and the comprehensive environmental state feature in the environmental state feature set and the distance value from the nearest boundary of the preset loss threshold range, the confidence of the current loss risk level is obtained.

4. The organic vegetable cold chain loss early warning system integrated with Internet of Things according to claim 3, characterized in that: In the S340, the preset risk division rule is: If the real-time loss rate prediction value ≤ the current loss risk level is a low risk level; If < Real-time loss rate prediction value ≤ the current loss risk level is a medium risk level; If the real-time loss rate prediction value is > If so, the current loss risk level is high risk level; wherein, and are threshold values set by percentiles of historical transportation data, corresponding to the boundary values of the low risk level and the high risk level, respectively.

5. The organic vegetable cold chain loss early warning system integrated with Internet of Things according to claim 3, characterized in that: In the S320, the weighting fusion process is as follows: wherein, is a weighted feature vector; represents a real-time observation value of the th original environment state feature; is a dynamic weight of the th original environment state feature; is a normalized comprehensive environment state feature of the th original environment state feature; represents a dynamic weight of the th comprehensive environment state feature; is a total number of original environment state features; is a total number of comprehensive environment state features.

6. The organic vegetable cold chain loss early warning system integrated with Internet of Things according to claim 3, characterized in that: In the S350, the confidence calculation process is as follows: wherein, is the minimum distance of the real-time observation value of the first original environmental state feature to the boundary of the preset loss threshold range; is the minimum distance of the real-time observation value of the first original environmental state feature to the boundary of the preset loss threshold range; is the minimum distance of the standardized value of the first comprehensive environmental state feature to the boundary of the preset loss threshold range; is the minimum distance of the standardized value of the first comprehensive environmental state feature to the boundary of the preset loss threshold range; represents the confidence degree.

7. The organic vegetable cold chain loss early warning system integrated with Internet of Things according to claim 1, characterized in that: The S400 specifically includes the following sub-steps: S410, determine whether the current loss risk level is a high risk level; if it is determined to be a high risk level, trigger a multi-level early warning signal generation process, otherwise return to S300 for continuous monitoring; S420, generate a first-level early warning signal, including: filter out the original environment state features or the comprehensive environment state features whose deviation degree exceeds a first deviation threshold based on deviation degrees of real-time observation values of each original environment state feature and the comprehensive environment state feature in the environment state feature set from the preset loss threshold range but do not reach a second deviation threshold to generate a primary abnormality parameter list; packaging the screening result as a first-level early warning signal and sending it to the user terminal through the wireless communication module; S430, generate a second-level early warning signal, including: if at least one of the deviation degrees of the original environmental state features and / or the comprehensive environmental state features in the primary abnormal parameter list exceeds a second deviation threshold value or the confidence of the real-time loss rate prediction value is lower than a first preset signal threshold value, a secondary early warning is triggered. matching the adjustment strategy corresponding to the out-of-limit feature from the preset adjustment strategy library, and prioritizing according to its dynamic weight to generate a recommended strategy set; The recommended strategy is encapsulated as a second-level early warning signal and sent to the user terminal with high priority; S440, generating a third-level early warning signal, comprising: If the deviation degree of a certain original environment state feature and / or comprehensive environment state feature in the primary abnormal parameter list continues to exceed the third deviation threshold for a preset length of time or the confidence of the real-time loss rate prediction value is lower than a second preset signal threshold , a tertiary early warning is triggered. Generating a third-level early warning signal and sending a forced regulation instruction through a redundant communication link.

8. The organic vegetable cold chain loss early warning system integrated with Internet of Things according to claim 7, characterized in that: In the S400 substep: the first deviation threshold the second deviation threshold and the third deviation threshold satisfy the following hierarchical relationship: first deviation threshold < < ; the first pre-set signal threshold and the second pre-set signal threshold satisfy the following hierarchical relationship: second preset signal threshold first preset signal threshold .

Images