A method and system for monitoring the quality of agricultural products in the logistics of fresh agricultural products

Abstract

The application relates to the transportation monitoring technical field and discloses a fresh agricultural product quality monitoring method and system in an in-transit logistics, which comprises the following steps: extracting vehicle vibration features in transportation vehicle driving data, analyzing the interaction information amount between the vehicle vibration features and the biomechanical damage state of fresh agricultural products by using a mutual information feature selection algorithm, and calculating a mechanical damage index caused to the fresh agricultural products in the transportation vehicle driving process based on the interaction information amount; calculating an environmental stress index based on the carriage environment features; inputting the mechanical damage index and the environmental stress index into a pre-trained agricultural product damage prediction model to output the cumulative damage index of the fresh agricultural products in a future period; predicting the remaining preservation time length of the fresh agricultural products, and formulating a fresh agricultural product preservation regulation strategy according to the remaining preservation time length. The application fully considers the coupling deterioration effect of the mechanical damage and the environmental stress on the agricultural product quality, and improves the accuracy and timeliness of the damage evaluation.

Description

Technical Field

[0001] This invention relates to the field of transportation monitoring technology, and more specifically, to a method and system for monitoring the quality of fresh agricultural products during transit logistics. Background Technology

[0002] In-transit logistics of fresh agricultural products refers to the entire process of transporting perishable agricultural products such as fruits, vegetables, and cut flowers from their harvesting sites to their destinations via road, rail, or combined transport. During transportation, the mechanical damage caused by vehicle bumps, combined with environmental stresses such as temperature fluctuations and abnormal humidity within the vehicle, accelerates product quality deterioration and causes significant economic losses. Therefore, quality monitoring during in-transit logistics, and timely understanding of the damage status and deterioration trends of agricultural products, is a crucial link in ensuring the commercial value of fresh agricultural products.

[0003] Existing methods for monitoring the quality of agricultural products still have the following shortcomings: On the one hand, most methods only focus on single environmental records of temperature and humidity in the vehicle compartment, or only monitor whether the vibration and impact amplitude exceeds the standard, failing to integrate mechanical damage with environmental stress for analysis, and making it difficult to reflect the coupled deterioration effect of the two on the quality of agricultural products; on the other hand, the monitoring results are mostly passive records of the current state, lacking the ability to predict future damage trends. Even if some methods attempt to introduce predictive models, they often rely solely on statistical inference from historical data in a single dimension, making it difficult to meet the accuracy and timeliness of the prediction results in actual logistics scenarios.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] In response to the problems in related technologies, this invention proposes a method and system for monitoring the quality of fresh agricultural products during transit logistics, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows: According to one aspect of the present invention, a method for monitoring the quality of fresh agricultural products during transit logistics is provided, the method comprising: S1. Collect multi-source datasets of fresh agricultural product transportation during the current time period. The multi-source datasets include vehicle driving data and environmental data inside the transport compartment. S2. Extract vehicle vibration characteristics from the driving data of transport vehicles, use mutual information feature selection algorithm to analyze the amount of interactive information between vehicle vibration characteristics and the biomechanical damage state of fresh agricultural products, and calculate the mechanical damage index of fresh agricultural products caused by transport vehicles during driving based on the amount of interactive information. S3. Extract the environmental features of the transport compartment from the environmental data inside the transport compartment, and calculate the environmental stress index based on the environmental features of the transport compartment; S4. Input the mechanical damage index and the environmental stress index into the pre-trained agricultural product damage prediction model, and output the cumulative damage index of fresh agricultural products in the future period; wherein, the agricultural product damage prediction model is trained using a multi-task learning strategy. S5. Based on the cumulative damage index of fresh agricultural products in the future period, predict the remaining shelf life of fresh agricultural products, and formulate fresh agricultural product preservation control strategies according to the remaining shelf life.

[0007] Preferably, S2 includes: S21. Perform time-frequency domain analysis on the driving data of transport vehicles, extract vehicle vibration features from the driving data of transport vehicles based on the time-frequency domain analysis results, and integrate them into a vehicle vibration feature set. S22. Using the mutual information feature selection algorithm, each feature in the vehicle vibration feature set is associated and matched with the biomechanical damage state of fresh agricultural products in the target database. Based on the association and matching results, a damage feature set is selected from the vehicle vibration feature set. The damage feature set includes a subset of instantaneous impact features and a subset of cumulative fatigue features; S23. Calculate the instantaneous impact damage index and the cumulative fatigue damage index based on the instantaneous impact feature subset and the cumulative fatigue feature subset, respectively. Combine the coupling deterioration effect between instantaneous impact damage and cumulative fatigue damage to calculate the mechanical damage index caused to fresh agricultural products by transport vehicles during operation.

[0008] Preferably, S22 includes: S221. Obtain the critical impact damage threshold and critical fatigue damage threshold of fresh agricultural products from the target database; S222. Compare each feature in the vehicle vibration feature set with the critical impact damage threshold and the critical fatigue damage threshold respectively. Based on the comparison results, divide the vehicle vibration feature set into an instantaneous impact feature set and a cumulative fatigue feature set. S223. Calculate the redundancy and local maximum interaction between each impact feature and other impact features in the instantaneous impact feature set, obtain the initial score of each impact feature according to the interaction expansion principle, and sort each impact feature according to the initial score. S224. Identify pairs of impact features that have positive interaction in the sorting order to form impact feature interaction groups. Continue to traverse the remaining impact features and add features that have positive synergy with the impact feature interaction groups to the group one by one until all impact features in the instantaneous impact feature set have been traversed to form a set of impact feature interaction groups. S225. Calculate the interaction weights between paired impact features within each impact feature interaction group, and correct the initial score of each impact feature. Based on the corrected comprehensive score, select the best subset of instantaneous impact features from the instantaneous impact feature set. S226. Select a subset of cumulative fatigue features from the cumulative fatigue feature set in accordance with the steps S223-S225.

[0009] Preferably, S225 includes: S2251. Initialize the selected feature set, and add the impact feature with the highest initial score as the first selected feature to the selected feature set; S2252. Among the remaining unselected impact features, perform the following steps sequentially for each candidate impact feature: Based on the amount of interaction information between pairs of impact features within the interaction group of the candidate impact feature, and combined with the initial scores of each impact feature, calculate the pairwise interaction weight between the candidate impact feature and the features in the same group in the selected feature set. Calculate the amount of interaction information between the current candidate impact feature interaction group and the currently selected feature set to obtain the interaction group weight; The initial score, pairwise interaction weights, and interaction group weights of the candidate impact feature are multiplied together to obtain the current comprehensive score of the candidate impact feature. S2253. Select the candidate impact feature with the highest current comprehensive score and add it to the selected feature set; S2254. Repeat S2252 to S2253 until the number of selected feature sets reaches the preset threshold or the comprehensive score of no candidate impact features is higher than the set lower limit. Output the final selected feature set as the instantaneous impact feature subset.

[0010] Preferably, S23 includes: S231. Input the subset of instantaneous impact features into the preset damage assessment model, and output the instantaneous impact damage index through the damage assessment model; S232. Input the cumulative fatigue feature subset into the preset damage assessment model, and output the cumulative fatigue damage index through the damage assessment model; S233. Based on the coupled deterioration effect of instantaneous impact damage and cumulative fatigue damage on fresh agricultural products, determine the dynamic fusion weight of instantaneous impact damage index and cumulative fatigue damage index. S234. Based on dynamic fusion weights, the instantaneous impact damage index and the cumulative fatigue damage index are weighted and fused to obtain the mechanical damage index that characterizes the entire transportation process of fresh agricultural products.

[0011] Preferably, S233 includes: S2331. Obtain the instantaneous impact damage index sequence and cumulative fatigue damage index sequence for the target time step; S2332. Calculate the impact-fatigue transfer entropy from the instantaneous impact damage index sequence to the cumulative fatigue damage index sequence. S2333. Calculate the fatigue-impact transfer entropy from the cumulative fatigue damage index sequence to the instantaneous impact damage index sequence; S2334. Combining the impact-fatigue transfer entropy and the fatigue-impact transfer entropy, calculate the coupling strength coefficient that characterizes the causal driving proportion of instantaneous impact damage to cumulative fatigue damage. S2335. The coupling strength coefficient is used as the dynamic fusion weight of the instantaneous impact damage index, and the difference between the target value and the coupling strength coefficient is used as the dynamic fusion weight of the cumulative fatigue damage index.

[0012] Preferably, S4 includes: S41. Obtain a multi-source historical dataset and divide it into a training set and a validation set. S42. Using the historical mechanical damage index and historical environmental stress index in the training set as model inputs and the historical cumulative damage measured values ​​as model output targets, the agricultural product damage prediction model is trained by combining a multi-task learning strategy to obtain the trained agricultural product damage prediction model. S44. The trained agricultural product damage prediction model is evaluated using a validation set. The evaluation metrics include the root mean square error and the coefficient of determination between the cumulative damage prediction value and the measured value. S45. After completing the model evaluation, the mechanical damage index and environmental stress index of the current period are used as accelerated deterioration factors and input into the trained agricultural product damage prediction model. The agricultural product damage prediction model outputs the cumulative damage index of fresh agricultural products in future periods.

[0013] Preferably, S41 includes: S411. Obtain a multi-source dataset of fresh agricultural product transportation within a historical time period, wherein the multi-source dataset includes historical vehicle driving data and historical environmental data inside the transport compartment. S412. Following the steps S2-S3, calculate the historical mechanical damage index and historical environmental stress index based on historical transport vehicle driving data and historical environmental data. S413. Using the historical cumulative damage measured values ​​of fresh agricultural products on the corresponding road section as labels, the historical mechanical damage index and historical environmental stress index are paired with the labels to construct a training sample set, and the training sample set is divided into a training set and a validation set.

[0014] Preferably, S42 includes: S421. Input the historical mechanical damage index and historical environmental stress index from the training set into the shared feature extraction network of the agricultural product damage prediction model to extract damage feature representations; the shared feature extraction network is a combination of fully connected layers or convolutional layers.

[0015] S422. Input the damage feature representation into the main task prediction head and the auxiliary task prediction head respectively; the main task prediction head outputs the cumulative damage prediction value, and the auxiliary task prediction head outputs the damage dominance type classification result, wherein the damage dominance type includes mechanical dominance type and environmental dominance type. S423. The mean square error between the predicted value of the cumulative damage index and the measured value of the historical cumulative damage is used as the main task loss, and the cross-entropy loss between the damage dominant type classification result and the true dominant type label is used as the auxiliary task loss. The main task loss and the auxiliary task loss are weighted and summed to obtain the joint loss function. S424. Minimize the joint loss function using the backpropagation algorithm and gradient descent optimizer, and iteratively update the parameters of the shared feature extraction network, the main task prediction head, and the auxiliary task prediction head until the preset convergence condition is met, thus obtaining the trained agricultural product damage prediction model.

[0016] According to another aspect of the present invention, a system for monitoring the quality of fresh agricultural products during transit is also provided, the system comprising: The multi-source data acquisition module is used to collect multi-source datasets of fresh agricultural products transportation during the current time period. The multi-source datasets include transportation vehicle driving data and environmental data inside the transportation compartment. The mechanical damage quantification module is used to extract vehicle vibration characteristics from the driving data of transport vehicles, use the mutual information feature selection algorithm to analyze the amount of interactive information between vehicle vibration characteristics and the biomechanical damage state of fresh agricultural products, and calculate the mechanical damage index caused to fresh agricultural products during the driving process of transport vehicles based on the amount of interactive information. The environmental stress quantification module is used to extract the environmental characteristics of the transport compartment from the environmental data inside the transport compartment, and calculate the environmental stress index based on the environmental characteristics of the compartment. The cumulative damage prediction module is used to input the mechanical damage index and the environmental stress index into a pre-trained agricultural product damage prediction model and output the cumulative damage index of fresh agricultural products in the future period; wherein, the agricultural product damage prediction model is trained using a multi-task learning strategy. The freshness control decision module is used to predict the remaining shelf life of fresh agricultural products based on the cumulative damage index of fresh agricultural products in future time periods, and to formulate freshness control strategies for fresh agricultural products based on the remaining shelf life.

[0017] The beneficial effects of this invention are as follows: 1. This invention collects multi-source vibration and environmental data during real-time transportation, quantifying the mechanical damage stress and environmental stress on fresh agricultural products from two dimensions: driving vibration and the vehicle environment. These two factors are then input into a pre-trained agricultural product damage prediction model as accelerating degradation factors, enabling proactive prediction of cumulative damage trends. Compared to single-dimensional monitoring, this invention fully considers the coupled degradation effect of mechanical damage and environmental stress on agricultural product quality, significantly improving the accuracy and timeliness of damage assessment. Furthermore, a mutual information feature selection algorithm is used to screen vibration features highly correlated with biomechanical damage states, effectively filtering out irrelevant noise and making the quantification of the mechanical damage index more accurate.

[0018] 2. This invention employs a mutual information feature selection algorithm to precisely correlate vehicle vibration characteristics with the biomechanical damage state of agricultural products, effectively eliminating redundant and irrelevant features and selecting a subset of differentiated features highly sensitive to impact damage and cumulative fatigue damage. Building upon this, the invention fully explores the positive synergistic and complementary effects between features, maximizing damage characterization capabilities with a limited number of sensitive features. Furthermore, it quantifies the causal coupling strength between impact and fatigue through transfer entropy, achieving adaptive allocation of dynamic fusion weights. This accurately captures the coupled deterioration effect of these two factors on agricultural product quality, significantly improving the accuracy and reliability of mechanical damage index quantification and providing a high-precision input foundation for subsequent damage prediction and preservation control.

[0019] 3. This invention employs a multi-task learning strategy to train an agricultural product damage prediction model. In the process of simultaneously predicting cumulative damage and identifying the dominant damage type, the shared feature extraction network effectively captures mechanical-environment coupling information. The joint optimization of the main and auxiliary tasks significantly improves the model's generalization ability and prediction accuracy. Based on this agricultural product damage prediction model, the future cumulative damage index can be output by inputting the current accelerating deterioration factor in real time, providing a reliable quantitative basis for subsequent shelf life prediction and differentiated control decisions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a method for monitoring the quality of fresh agricultural products during transit logistics according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a fresh agricultural product quality monitoring system in transit logistics according to an embodiment of the present invention.

[0022] In the picture: 1. Multi-source data acquisition module; 2. Mechanical damage quantification module; 3. Environmental stress quantification module; 4. Cumulative damage prediction module; 5. Preservation control decision module. Detailed Implementation

[0023] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0024] According to embodiments of the present invention, a method and system for monitoring the quality of fresh agricultural products during transit logistics are provided.

[0025] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a method for monitoring the quality of fresh agricultural products during transit logistics is provided, the method comprising: S1. Collect multi-source datasets of fresh agricultural product transportation during the current time period. The multi-source datasets include vehicle driving data and environmental data inside the transport compartments, specifically including: Accelerometers deployed on the chassis or cargo compartment of transport vehicles collect triaxial vibration acceleration signals in real time, which are used as driving data for the transport vehicles.

[0026] Temperature, relative humidity, and ethylene concentration data are collected in real time inside the transport vehicle using temperature, humidity, and ethylene concentration sensors installed inside the vehicle. These data serve as the environmental data for the transport vehicle's interior. The temperature sensor is a PT100 platinum resistance temperature sensor with a measurement range of -20℃ to 60℃ and an accuracy of ±0.3℃. The humidity sensor is a capacitive humidity sensor with a measurement range of 0-100%RH and an accuracy of ±3%RH. The ethylene concentration sensor is an electrochemical ethylene sensor with a measurement range of 0-100ppm and a resolution of 0.1pp.

[0027] The vehicle driving data and environmental data inside the transport compartment at the same sampling time are timestamped and integrated into a multi-source dataset. The data subset corresponding to the current time period is extracted as the multi-source dataset for the transportation of fresh agricultural products in the current time period.

[0028] S2. Extract vehicle vibration characteristics from the driving data of transport vehicles, use the mutual information feature selection algorithm to analyze the amount of interactive information between vehicle vibration characteristics and the biomechanical damage state of fresh agricultural products, and calculate the mechanical damage index caused to fresh agricultural products during the driving process of transport vehicles based on the amount of interactive information.

[0029] Wherein, S2 includes: S21. Perform time-frequency domain analysis on the transport vehicle driving data, extract vehicle vibration features from the transport vehicle driving data based on the time-frequency domain analysis results, and integrate them into a vehicle vibration feature set, specifically including: The acquired triaxial vibration acceleration signals were preprocessed, including detrending and bandpass filtering. Detrending was performed using least-squares fitting to remove DC components and slow drift from the signal; bandpass filtering used a Butterworth filter with a passband frequency of 0.5-500Hz to remove low-frequency vehicle body sway noise and high-frequency electromagnetic interference.

[0030] The preprocessed vibration signal of each axis is divided into frames with 2048 sampling points and a frame shift of 1024 sampling points. The Hanning window is used for frame segmentation to obtain multiple frames of vibration signal segments.

[0031] Temporal features are extracted from each frame of vibration signal segment, including peak acceleration, root mean square value, waveform factor, peak factor, and kurtosis, forming a temporal feature vector, where: Peak acceleration: the maximum absolute value of the vibration signal, used to capture instantaneous impact intensity; Kurtosis: The fourth-order normalized moment of the vibration signal distribution, which is sensitive to impact vibration; Peak factor: The ratio of peak acceleration to root mean square value, used to distinguish between impact signals and stationary signals.

[0032] Wavelet packet transform is performed on the same frame of vibration signal segments, using the db8 wavelet basis function to decompose the signal into several frequency bands. The wavelet packet energy proportion and wavelet packet energy entropy of each frequency band are calculated to form the time-frequency domain fusion feature vector of the signal in this frame. The frequency band division of the wavelet packet transform matches the frequency bands sensitive to agricultural product damage, and can simultaneously preserve the time-domain transient information and frequency-domain energy distribution information of the vibration signal.

[0033] The time-domain feature vector of each frame is concatenated with the time-frequency domain fused feature vector to form the complete vibration feature vector of that frame. The complete vibration feature vectors of all frames are arranged in chronological order to form the vehicle vibration feature set.

[0034] S22. Using the mutual information feature selection algorithm, each feature in the vehicle vibration feature set is associated and matched with the biomechanical damage state of fresh agricultural products in the target database. Based on the association and matching results, a damage feature set is selected from the vehicle vibration feature set.

[0035] Wherein, S22 includes: S221. Obtain the critical impact damage threshold and critical fatigue damage threshold of fresh agricultural products from the target database; S222. Compare each feature in the vehicle vibration feature set with the critical impact damage threshold and the critical fatigue damage threshold respectively. Based on the comparison results, divide the vehicle vibration feature set into an instantaneous impact feature set and a cumulative fatigue feature set.

[0036] It should be noted that each feature vector in the vehicle vibration feature set is identified and classified according to its source type. For features extracted from time-domain analysis, including peak acceleration, peak factor, and kurtosis, they are compared with the critical impact damage threshold: if the peak acceleration exceeds the critical impact damage threshold, or if the peak factor and kurtosis both exceed their corresponding statistical benchmark values, then the feature vector is classified into the instantaneous impact feature set.

[0037] For features extracted from time-frequency domain fusion analysis, including wavelet packet energy proportion and wavelet packet energy entropy, they are compared with the critical fatigue damage threshold: if the wavelet packet energy proportion of a specific frequency band exceeds the critical fatigue damage threshold, or the wavelet packet energy entropy is lower than the preset lower limit of entropy value, then the feature vector is assigned to the cumulative fatigue feature set.

[0038] The comparison process employs a frame-by-frame approach. For each frame of vibration signal feature vectors, the two types of comparisons described above are performed, and a category label is assigned to the feature vector of that frame based on the comparison results. After all frames have been processed, the vehicle vibration feature set is divided into an instantaneous impact feature set and a cumulative fatigue feature set according to the category labels. If a feature vector in a frame simultaneously satisfies both criteria, it is determined based on the degree to which its feature values ​​exceed a threshold, and it is assigned to the feature set with the higher exceedance ratio.

[0039] S223. Calculate the redundancy and local maximum interaction between each impact feature and other impact features in the instantaneous impact feature set, obtain the initial score of each impact feature according to the interaction expansion principle, and sort each impact feature according to the initial score.

[0040] It should be noted that redundancy measures the degree of overlap in the information carried by two features; the higher the overlap, the stronger the redundancy. Specifically, it is measured using the absolute value of the Pearson correlation coefficient, a method used in existing technologies. Interactivity measures whether the ability to identify the target damage state when two features are combined is greater than the simple sum of their individual capabilities. Specifically, it is measured using the amount of interactive information, including: Traversing each pair of impact features and Calculate their relationship with the target variable. Y The amount of interactive information between (i.e., the binary state of whether agricultural products have exceeded the damage threshold) .

[0041] Simultaneously calculate their individual values ​​and... Y mutual information and .

[0042] Local interactivity = .

[0043] if This indicates the existence of a positive interaction (synergistic effect) for each impact characteristic. The maximum value of its interaction with other features is taken as its local maximum interaction. The higher this value, the easier it is for the feature to form a strong synergistic combination with other features. Conversely, it indicates a negative interaction, meaning that the effect of combining two features is not as good as the sum of their individual effects.

[0044] The interaction expansion principle is a comprehensive scoring strategy, and its formula is: Initial score = Relevance × (1 - Redundancy) × (1 + Local maximum interactivity); where relevance is the impact feature. With target variable Y Mutual information between (1-Redundancy) represents the penalty term. The higher the redundancy, the smaller the value of this term, which encourages the selection of independent features and non-overlapping information; (1+Local Maximum Interactivity) represents the reward term.

[0045] S224. Identify pairs of impact features that have positive interaction in the sorting order to form impact feature interaction groups. Continue to traverse the remaining impact features and add features that have positive synergy with the impact feature interaction groups to the group one by one until all impact features in the instantaneous impact feature set have been traversed to form a set of impact feature interaction groups. S225. Calculate the interaction weights between paired impact features within each impact feature interaction group, and correct the initial score of each impact feature. Based on the corrected comprehensive score, select the best subset of instantaneous impact features from the instantaneous impact feature set.

[0046] Wherein, S225 includes: S2251. Initialize the selected feature set, and add the impact feature with the highest initial score as the first selected feature to the selected feature set; S2252. Among the remaining unselected impact features, perform the following steps sequentially for each candidate impact feature: Based on the amount of interaction information between pairs of impact features within the interaction group of the candidate impact feature, and combined with the initial scores of each impact feature, the pairwise interaction weights between the candidate impact feature and features in the same group of the selected feature set are calculated, specifically including: ①Required prerequisite data: The amount of interactive information of each pair of impact features stored in step S223 ; The initial scores of each impact characteristic calculated in step S223 ; The set of impact feature interaction groups constructed in step S224; ② Implementation steps: Determine the current candidate impact characteristics The interaction group I belong to ; Iterate through the currently selected feature set and filter out those that also belong to the interaction group. The features constitute the set of selected features in the same group. ; like If the value is empty (i.e., there are no members of the interaction group in the selected feature set), then the pairwise interaction weights are returned directly. Skip the subsequent steps; if Not empty, for each selected feature Calculate the pairwise interaction weights : Retrieve from the stored interaction information table Retrieve the initial score from the S223 results. and ; Calculate the strength of cooperation ; like ,but: ; like If the features do not exhibit positive collaboration, then the pair of features has no positive collaboration. .

[0047] Multiply the one-to-one weights of all selected features in the same group to obtain the joint pairwise interaction weights:

[0048] The product form ensures that the collaborative contribution of each selected feature in the same group is taken into account, and if any pair of weights is zero, the joint weight is also zero.

[0049] Calculate the amount of interaction information between the current candidate impact feature's impact feature interaction group and the currently selected feature set to obtain the interaction group weight, specifically including: ① The required prerequisite data includes the set of impact feature interaction groups constructed in step S224, as well as the impact features and target variables. Y Mutual information value; ② Implementation steps: Determine the current candidate impact characteristics The interaction group it belongs to (i.e., the impact feature interaction group, hereinafter referred to as the interaction group). .

[0050] Calculate the currently selected feature set With target variable Y Joint information This value measures the overall predictive power of the selected feature set for the damage state.

[0051] Computational Interaction Group With the selected feature set The union of the target variable and the target variable Y Joint information This value is used to measure the interaction group The overall predictive power after adding all information to the selected feature set.

[0052] Computational Interaction Group Complementary information increments to the selected feature set: ; The complementary information increment represents the interaction group. How much new, uncovered damage-related information can be provided to the selected feature set?

[0053] Computational Interaction Group Total amount of information of itself That is, the mutual information of all features within the interaction group and Y, and based on the interaction group The weight of the interaction group is calculated by combining its own total information volume with the incremental complementary information. ; The initial score, pairwise interaction weights, and interaction group weights of the candidate impact feature are multiplied together to obtain the current comprehensive score of the candidate impact feature, which specifically includes: ①Required prerequisite data: Initial scoring of candidate impact features in step S223 Pairwise interaction weights and interaction group weights .

[0054] ② Implementation steps: Multiply the three factors directly:

[0055] The weight values ​​range from 0 to 1; the closer the value is to 1, the stronger the interaction group. The more new information provided, the stronger the complementarity; the closer the value is to 0, the more the information has been fully covered by the selected feature set.

[0056] S2253. Select the candidate impact feature with the highest current comprehensive score and add it to the selected feature set; S2254. Repeat S2252 to S2253 until the number of selected feature sets reaches the preset threshold or the comprehensive score of no candidate impact features is higher than the set lower limit. Output the final selected feature set as the instantaneous impact feature subset.

[0057] S226. Select a subset of cumulative fatigue features from the cumulative fatigue feature set in accordance with the steps S223-S225.

[0058] The following describes step S22 in further detail with reference to specific embodiments.

[0059] Suppose that the instantaneous impact feature set extracted from a certain transportation process contains four features: feature A (peak acceleration), feature B (peak factor), feature C (kurtosis), and feature D (waveform factor).

[0060] Phase 1, Initial Scoring and Ranking: The mutual information between each feature and the target variable Y (a binary state indicating whether agricultural products exceed the critical shock damage threshold) is calculated as a correlation score. The calculated correlation scores are: feature A has the highest correlation, followed by feature C, then feature B, and feature D has the lowest.

[0061] Calculate the absolute values ​​of the Pearson correlation coefficients between each feature and other features, and take the mean as the redundancy. The calculations show that feature A is highly correlated with feature B, resulting in the highest redundancy; feature D has low correlations with all other features, resulting in the lowest redundancy.

[0062] For each pair of features, the local interactivity is calculated. When features A and C are combined, the joint mutual information is significantly greater than the sum of their individual mutual information, resulting in a large positive local interactivity value. This indicates a strong positive synergy between the two features; the peak acceleration characterizes the impact intensity, and the kurtosis characterizes the transient characteristics of the impact. Their combination provides a more comprehensive portrayal of the impact damage pattern. However, when features A and B are combined, due to their high correlation and significant information overlap, the local interactivity is negative.

[0063] The maximum local interaction of each feature with other features is taken as the local maximum interaction of that feature. The local maximum interaction of features A and C both originate from each other, resulting in the highest value among all feature pairs; the local maximum interaction of features B and D is lower. Following the interaction expansion principle formula: Initial Score = Relevance × (1 - Redundancy) × (1 + Local Maximum Interaction), the initial scores for each feature are calculated. Although feature C's relevance is slightly lower than feature A's, its redundancy is lower, and it shares the same local maximum interaction with feature A, resulting in a slightly higher overall score and ranking first. Feature A ranks second, and features D and B rank third and fourth respectively. The final ranking is: C > A > D > B.

[0064] The second stage involves constructing a set of feature interaction groups. Interaction groups are constructed according to the sorting order. Feature C, ranked first, is taken as the core of the first interaction group. In the remaining features {A, D, B}, features with positive local interactions with C are searched sequentially. It is found that feature A has a positive local interaction with C, so A is paired with C to form the initial interaction group {C, A}. The remaining features {D, B} are then traversed, checking if they have a positive local interaction with any feature in the current interaction group. Feature D has a positive local interaction with both feature A and feature C, satisfying the inclusion condition, so D is added to the interaction group, updating it to {C, A, D}. Feature B has a negative local interaction with feature A, not satisfying the positive interaction condition, and is not included in the group.

[0065] At this point, the first interaction group is fixed as: interaction group ① = {C, A, D}, the remaining ungrouped feature is {B}, which constitutes interaction group ②, and the final set of interaction groups is {interaction group ①, interaction group ②}.

[0066] The third stage involves iterative selection of feature subsets: Select the feature C with the highest initial score as the first selected feature and add it to the selected feature set. At this point, the selected feature set = {C}, and execute the first round of iteration: The remaining candidate features are {A, D, B}. The current comprehensive score is calculated for each feature. Taking candidate feature A as an example, A and the selected feature C belong to the same interaction group ①. The pairwise interaction weights between A and C are calculated. I(A, C; Y) is retrieved from the stored interaction information table, and the initial score S is retrieved from the S223 results. A and S C Calculate the synergy strength Δ A,C Calculations show that Δ A,C >0 indicates positive collaboration. Since the local interaction between A and C is the highest among all feature pairs, and their initial score product is the largest within interaction group ①, the pairwise interaction weight is relatively large. Because only C belongs to interaction group ① in the selected feature set, the joint pairwise interaction weight is W. A,C .

[0067] Simultaneously, the interaction group weights between interaction group ① and the currently selected feature set {C} are calculated. In interaction group ①, A and D are still unselected, resulting in a large complementary information increment ΔI, and thus the interaction group weights are... If the value is higher, the three factors are multiplied together to obtain the current comprehensive score of A, which is a higher score.

[0068] The same calculation is performed on candidate feature D; its local interaction with C is lower than that between A and C, S D ×S CThe weight of pairwise interactions is relatively low; the weight of the interaction group is the same as that of A. The current comprehensive score of D is lower than that of A. For candidate feature B, there are no selected features in its interaction group ②, and the weight of pairwise interactions is 1. Interaction group ② contains only one feature, B, and the incremental complementary information with the selected feature set {C} is limited, so the weight of the interaction group is low. The current comprehensive score of B is lower than that of A but higher than that of D.

[0069] This round's ranking: Score A Score B Score D Select A and add it to the selected feature set, updating it to {C,A}.

[0070] Repeat the second iteration: The selected feature set is {C, A}, and the remaining candidate features are {D, B}. For candidate feature D, the pairwise interaction weights between D and C, and between D and A, need to be calculated separately. The single pairwise weight W between D and C is... D,C In the middle, S D ×S C The normalized value is relatively small; the single pair weights W of D and A are relatively small. D,A In this case, due to the existence of positive collaboration, the value is acceptable. The joint pairwise interaction weight is the product of the two weights compared to W in the previous round. A,C The weights have decreased somewhat. Regarding the interaction group weights, since two features in interaction group ① have already been selected, the remaining complementary information is reduced. Significant decline. D's current overall score has dropped dramatically.

[0071] For candidate feature B, the pairwise interaction weight remains 1, and the interaction group weight remains relatively stable. The current comprehensive score of B remains stable; the ranking in this round is based on Score. B Score D Select B and add it to the selected feature set, updating it to {C,A,B}.

[0072] Termination judgment and output: If the preset feature quantity threshold is 3, then the threshold has been reached and the iteration stops; if the preset score lower limit is higher than the current comprehensive score of the remaining candidate feature D, the iteration also stops; the final selected feature set {C,A,B} is output as the instantaneous impact feature subset.

[0073] Following the same approach as steps S223 to S225, the interactive expansion principle and iterative selection method are applied to the cumulative fatigue feature set, and a subset of cumulative fatigue features is selected through iterative selection. This subset, together with the subset of instantaneous impact features, constitutes a complete damage feature set.

[0074] The damage feature set includes a subset of instantaneous impact features and a subset of cumulative fatigue features; S23. Calculate the instantaneous impact damage index and the cumulative fatigue damage index based on the instantaneous impact feature subset and the cumulative fatigue feature subset, respectively. Combine the coupling deterioration effect between instantaneous impact damage and cumulative fatigue damage to calculate the mechanical damage index caused to fresh agricultural products by transport vehicles during operation.

[0075] Wherein, S23 includes: S231. Input the subset of instantaneous impact features into the preset damage assessment model, and output the instantaneous impact damage index through the damage assessment model; S232. Input the cumulative fatigue feature subset into the preset damage assessment model, and output the cumulative fatigue damage index through the damage assessment model; S233. Based on the coupled deterioration effect of instantaneous impact damage and cumulative fatigue damage on fresh agricultural products, determine the dynamic fusion weight of instantaneous impact damage index and cumulative fatigue damage index.

[0076] It should be noted that the damage assessment model contains two independent and parallel computational branches: an impact damage calculation branch and a fatigue damage calculation branch. The two branches share the model's input interface, but their internal logic is completely decoupled. When the input is a subset of instantaneous impact features, the impact damage calculation branch is triggered; when the input is a subset of cumulative fatigue features, the fatigue damage calculation branch is triggered, specifically including: ① Construction of the impact damage calculation branch: The core of the impact damage calculation branch is an impact response spectrum analysis module, which integrates impact response spectrum curves of fresh agricultural products pre-calibrated through standard impact drop tests. These curves describe the probability of damage to agricultural products under different impact acceleration amplitudes. The curve calibration process is as follows: a standard sample of the same variety as the target agricultural product is selected and placed on an impact test bench. Single impacts are applied to the sample at different acceleration amplitudes, progressively increasing from low to high. After each impact, the sample is checked for damage such as bruising or cracking. The proportion of damage occurring at each acceleration amplitude is recorded. The impact response spectrum curve is obtained through data fitting and written into this branch.

[0077] ② Construction of the fatigue damage calculation branch: The core of the fatigue damage calculation branch is the modified Miner linear cumulative damage model, which integrates the SN curve of agricultural products pre-calibrated through fixed-frequency sweep fatigue tests. This curve describes the number of cycles required for fatigue failure of agricultural products under different vibration stress levels. The curve calibration process is as follows: standard samples of the same variety as the target agricultural product are selected, placed on a vibration test bench, and fixed-frequency fatigue tests are conducted at different vibration stress levels. The number of cycles required for fatigue damage to occur at each stress level is recorded. The SN curve is obtained through data fitting and written into this branch.

[0078] ③ Model Validation: Simulated transportation tests are conducted using independent test samples that have not been calibrated. Vibration data during the test is input into the model, which outputs the instantaneous impact damage index and the cumulative fatigue damage index, respectively. The output results are compared with the actual damage level of the samples after the test. If the prediction error is within the preset range, the model passes validation; if the deviation is large, the parameters of the response spectrum curve or SN curve are fine-tuned and the model is revalidated.

[0079] Wherein, S233 includes: S2331. Obtain the instantaneous impact damage index sequence and cumulative fatigue damage index sequence for the target time step; S2332. Calculate the impact-fatigue transfer entropy from the instantaneous impact damage index sequence to the cumulative fatigue damage index sequence. S2333. Calculate the fatigue-impact transfer entropy from the cumulative fatigue damage index sequence to the instantaneous impact damage index sequence; S2334. Combining the impact-fatigue transfer entropy and the fatigue-impact transfer entropy, calculate the coupling strength coefficient that characterizes the causal driving proportion of instantaneous impact damage to cumulative fatigue damage. S2335. The coupling strength coefficient is used as the dynamic fusion weight of the instantaneous impact damage index, and the difference between the target value and the coupling strength coefficient is used as the dynamic fusion weight of the cumulative fatigue damage index.

[0080] S234. Based on dynamic fusion weights, the instantaneous impact damage index and the cumulative fatigue damage index are weighted and fused to obtain the mechanical damage index that characterizes the entire transportation process of fresh agricultural products.

[0081] It should be noted that the coupling strength coefficient is a scalar value between 0 and 1, representing the proportion of the causal driving force of instantaneous impact damage on cumulative fatigue damage in the past N time steps. The coupling strength coefficient is used as the dynamic fusion weight of the instantaneous impact damage index at the current moment, and the difference between 1 and the coupling strength coefficient is used as the dynamic fusion weight of the cumulative fatigue damage index at the current moment.

[0082] The dynamic fusion weights are determined based on the coupled deterioration effect of instantaneous impact damage and cumulative fatigue damage on fresh agricultural products. Specifically, the instantaneous impact damage index sequence and the cumulative fatigue damage index sequence over the past N time steps are obtained; the transfer entropy from impact damage to fatigue damage and from fatigue damage to impact damage are calculated. The transfer entropy in these two directions quantifies the causal driving force of impact on fatigue and the inverse amplification degree of fatigue on impact sensitivity, respectively; the impact-fatigue transfer entropy is divided by the sum of the transfer entropies in the two directions to obtain the coupling strength coefficient, which represents the proportion of causal driving force of impact damage on fatigue damage; the coupling strength coefficient is used as the dynamic fusion weight of the instantaneous impact damage index, and the difference between 1 and this coefficient is used as the dynamic fusion weight of the cumulative fatigue damage index. The calculation of transfer entropy preferentially adopts the discrete transfer entropy estimation method, which estimates the probability distribution by the occurrence frequency of matching configurations in the statistical time series and then applies it to the definition of transfer entropy for solution. The discrete transfer entropy estimation method is an existing technology and will not be elaborated on here.

[0083] S3. Extract the environmental features of the transport vehicle from the environmental data inside the vehicle, and calculate the environmental stress index based on the environmental features, specifically including: Real-time time-series data on temperature, humidity, and ethylene concentration are extracted from the environmental data inside the transport vehicle. This environmental data is collected and transmitted by sensors deployed within the vehicle at a preset frequency.

[0084] Based on the preset suitable temperature range for this fresh agricultural product Calculate the temperature stress index Specifically, this involves integrating the temperature deviation beyond a suitable range over time within a set time window. For example, when the temperature... Higher than or below At that time, the amount of stress can be expressed as This index is used to reflect the degree of cumulative heat or cold damage caused by temperature deviation.

[0085] Based on the preset suitable humidity range for this fresh agricultural product Calculate the humidity stress index Specifically, this includes: within a preset time window, counting the cumulative time when the humidity value exceeds the suitable range, and calculating the proportion of this time to the total window duration, i.e. the proportion of humidity stress time. This index is used to reflect the degree of risk of condensation and disease growth due to high humidity or wilting due to low humidity.

[0086] Based on real-time collected ethylene concentration data Calculate the ethylene stress index Specifically, this involves calculating the integral of ethylene concentration over time within a preset time window, i.e., the ethylene exposure dose. This index is used to reflect the degree of stress caused by ethylene accumulation, which accelerates maturation and aging.

[0087] By employing a weighted summation method, the three stress indices mentioned above are fused to obtain the environmental stress index. : ; Among them, weight The environmental stress index is pre-set based on the sensitivity of this fresh agricultural product variety to various environmental factors. A scalar value is used to reflect in real time the overall stress of the current environment inside the transport compartment on the quality of agricultural products.

[0088] S4. Input the mechanical damage index and the environmental stress index into the pre-trained agricultural product damage prediction model, and output the cumulative damage index of fresh agricultural products in the future period; wherein, the agricultural product damage prediction model is trained using a multi-task learning strategy.

[0089] Wherein, S4 includes: S41. Obtain multi-source historical datasets and divide them into training and validation sets.

[0090] Wherein, S41 includes: S411. Obtain a multi-source dataset of fresh agricultural product transportation within a historical time period, wherein the multi-source dataset includes historical vehicle driving data and historical environmental data inside the transport compartment. S412. Following the steps S2-S3, calculate the historical mechanical damage index and historical environmental stress index based on historical transport vehicle driving data and historical environmental data. S413. Using the historical cumulative damage measured values ​​of fresh agricultural products on the corresponding road section as labels, the historical mechanical damage index and historical environmental stress index are paired with the labels to construct a training sample set, and the training sample set is divided into a training set and a validation set.

[0091] S42. Using the historical mechanical damage index and historical environmental stress index in the training set as model inputs and the historical cumulative damage measured values ​​as model output targets, the agricultural product damage prediction model is trained by combining a multi-task learning strategy to obtain the trained agricultural product damage prediction model.

[0092] Wherein, S42 includes: S421. Input the historical mechanical damage index and historical environmental stress index from the training set into the shared feature extraction network of the agricultural product damage prediction model to extract damage feature representations; the shared feature extraction network is a combination of fully connected layers or convolutional layers.

[0093] Specifically, the shared feature extraction network consists of a first fully connected layer, a batch normalization layer, an activation function layer, and a second fully connected layer connected sequentially. The first fully connected layer maps the input two-dimensional feature vector (mechanical damage index, environmental stress index) to a 64-dimensional feature space. The batch normalization layer normalizes the mapping result, accelerating training convergence and preventing gradient vanishing. The activation function layer uses the ReLU function to introduce non-linear expressive power into the network. The second fully connected layer maps the 64-dimensional features to a 32-dimensional damage feature representation vector. This vector serves as the shared input for the subsequent two prediction heads, containing the coupled feature information between mechanical damage and environmental stress.

[0094] S422. Input the damage feature representation into the main task prediction head and the auxiliary task prediction head respectively; the main task prediction head outputs the cumulative damage prediction value, and the auxiliary task prediction head outputs the damage dominance type classification result, wherein the damage dominance type includes mechanical dominance type and environmental dominance type.

[0095] Specifically, the main task prediction head consists of a fully connected layer and a linear output layer connected in sequence. The fully connected layer maps the 32-dimensional damage feature representation to a 16-dimensional feature space using the ReLU activation function; the linear output layer maps the 16-dimensional features into a single scalar value, which serves as the predicted cumulative damage index value, representing the expected cumulative damage level under the current mechanical damage and environmental stress conditions.

[0096] The auxiliary task prediction head consists of a fully connected layer and a softmax classification layer connected sequentially. The fully connected layer maps the 32-dimensional damage feature representation to an 8-dimensional feature space, using the ReLU activation function. The softmax classification layer maps the 8-dimensional features to a two-dimensional probability distribution, outputting the damage dominance type classification result, and selecting the category corresponding to the highest probability. The damage dominance type includes mechanically dominant and environmentally dominant. Mechanically dominant indicates that the current accumulated damage is mainly caused by vibration and impact factors, while environmentally dominant indicates that the current accumulated damage is mainly caused by environmental factors such as temperature, humidity, or ethylene.

[0097] S423. The mean squared error between the predicted cumulative damage index and the measured historical cumulative damage is used as the main task loss, and the cross-entropy loss between the damage dominant type classification result and the true dominant type label is used as the auxiliary task loss. The joint loss function is obtained by weighted summing of the main task loss and the auxiliary task loss. The true dominant type label is determined based on the contribution ratio of the mechanical damage index and the environmental stress index to the cumulative measured damage value in historical road sections. Specifically, this involves normalizing the historical mechanical damage index and the historical environmental stress index to the 0-1 interval, and calculating the normalized mechanical damage index using partial least squares regression. k mThe contribution coefficients to the measured cumulative damage values ​​and the contribution coefficients to the environmental stress index. k e ;like k m ×Normalized Mechanical Damage Index> k e If the normalized environmental stress index is ×, it is marked as mechanically dominant; otherwise, it is marked as environmentally dominant.

[0098] S424. Minimize the joint loss function using the backpropagation algorithm and gradient descent optimizer, and iteratively update the parameters of the shared feature extraction network, the main task prediction head, and the auxiliary task prediction head until the preset convergence condition is met, thus obtaining the trained agricultural product damage prediction model.

[0099] The gradient descent optimizer used is the Adam optimizer, with an initial learning rate of 0.001 and a batch size of 32 samples. In each training round, training set samples are input into the model in batches according to the batch size. After calculating the joint loss function, the gradient of the loss function with respect to each parameter in the shared feature extraction network, the main task prediction head, and the auxiliary task prediction head is calculated using the backpropagation algorithm. The Adam optimizer adaptively adjusts the parameter values ​​based on the first and second moments of the gradient. After one round of training, the joint loss function value on the validation set is calculated. The preset convergence condition is: the decrease in the joint loss function value on the validation set is less than 0.001 for five consecutive training rounds, or the training rounds reach the preset maximum number of iterations (500 rounds). After meeting the convergence condition, the model parameters are saved, resulting in the trained agricultural product damage prediction model.

[0100] It should be further explained that, through the design of the shared feature extraction network, the main task (cumulative damage prediction) and the auxiliary task (damage-dominant type classification) mutually promote each other during training. The classification loss of the auxiliary task forces the shared network to retain the differential information between mechanical damage and environmental stress when learning feature representations. This information, in turn, enhances the main task's ability to predict the cumulative damage index, solving the problem that traditional single-task models struggle to capture cross-factor coupling effects.

[0101] S44. The trained agricultural product damage prediction model is evaluated using a validation set. The evaluation metrics include the root mean square error and the coefficient of determination between the cumulative damage prediction value and the measured value. S45. After completing the model evaluation, the mechanical damage index and environmental stress index of the current period are used as accelerating deterioration factors and input into the trained agricultural product damage prediction model. The agricultural product damage prediction model outputs the cumulative damage index of fresh agricultural products in future periods, specifically including: The mechanical damage index calculated for the current time period is obtained from step S2, and the environmental stress index calculated for the current time period is obtained from step S3. The two together constitute a two-dimensional input vector (mechanical damage index, environmental stress index), which serves as the accelerated deterioration factor for the current time period.

[0102] The two-dimensional input vector is fed into the shared feature extraction network of the trained agricultural product damage prediction model. The shared feature extraction network consists of a first fully connected layer, a batch normalization layer, an activation function layer, and a second fully connected layer connected sequentially. The input vector is mapped to a 64-dimensional feature space by the first fully connected layer, normalized by the batch normalization layer, subjected to a nonlinear transformation by the ReLU activation function layer, and mapped to a 32-dimensional damage feature representation vector by the second fully connected layer.

[0103] The damage feature representation vector is fed into the main task prediction head, which consists of a fully connected layer and a linear output layer connected sequentially. After the damage feature representation vector is mapped to a 16-dimensional feature space by the fully connected layer, the linear output layer outputs a scalar value, which is the cumulative damage index of fresh agricultural products in the future time period predicted under the current time condition.

[0104] The auxiliary task prediction head can be selectively used during the inference phase. If it is necessary to simultaneously obtain the dominant damage type, the damage feature representation vector is fed into the auxiliary task prediction head in parallel, and the classification result of the dominant damage type (mechanical or environmental) for the current time period is output after passing through a fully connected layer and a Softmax classification layer, providing additional decision-making reference for the formulation of subsequent preservation control strategies.

[0105] By repeating the above reasoning process along the transportation timeline for each period, the cumulative damage index sequence for each future period can be output, thus enabling the prediction of damage trends throughout the transportation process.

[0106] S5. Based on the cumulative damage index of fresh agricultural products in the future, predict the remaining shelf life of fresh agricultural products, and formulate fresh agricultural product preservation control strategies based on the remaining shelf life, specifically including: Obtain the cumulative damage index for future periods output by the model; this index is a scalar value between 0 and 1. Match this index with a pre-defined damage-quality degradation curve for the agricultural product variety, which describes the mapping relationship between the cumulative damage index and the percentage of shelf life already consumed. Based on the current cumulative damage index, find the corresponding percentage of shelf life already consumed on the curve and calculate the remaining shelf life.

[0107] Differentiated strategies are formulated based on the threshold range of the remaining shelf life: a remaining shelf life greater than 12 hours (each threshold needs to be determined according to the specific agricultural product type in actual implementation) is the safe range, and the current transportation parameters remain unchanged; 4 to 12 hours is the warning range, triggering dynamic preservation control, lowering the compartment temperature by 2 to 3°C, and taking targeted measures according to the damage type—for mechanically-driven damage, priority is given to slowing down or adjusting the route to reduce vibration, and for environmentally-driven damage, priority is given to adjusting temperature and humidity control parameters to suppress environmental degradation, while a warning message is sent to the logistics dispatch center to prompt priority delivery; less than 4 hours is the emergency range, triggering emergency control, lowering the compartment temperature to the lowest safe temperature that the variety can tolerate and starting the maximum power refrigeration, while searching for the nearest cold storage or transit station to generate emergency transfer suggestions, marking this batch as the highest priority and notifying the recipient to prepare in advance.

[0108] According to another embodiment of the invention, such as Figure 2 As shown, a fresh agricultural product quality monitoring system for in-transit logistics is also provided. This system includes: Multi-source data acquisition module 1 is used to collect multi-source datasets of fresh agricultural products transportation during the current time period. The multi-source datasets include transportation vehicle driving data and environmental data inside the transportation compartment. Mechanical damage quantification module 2 is used to extract vehicle vibration characteristics from the driving data of transport vehicles, analyze the interaction information between vehicle vibration characteristics and the biomechanical damage state of fresh agricultural products using the mutual information feature selection algorithm, and calculate the mechanical damage index caused to fresh agricultural products during the driving process of transport vehicles based on the interaction information. The environmental stress quantification module 3 is used to extract the environmental features of the transport compartment from the environmental data inside the transport compartment, and calculate the environmental stress index based on the environmental features of the transport compartment. The cumulative damage prediction module 4 is used to input the mechanical damage index and the environmental stress index into the pre-trained agricultural product damage prediction model and output the cumulative damage index of fresh agricultural products in the future period; wherein, the agricultural product damage prediction model is trained using a multi-task learning strategy. The preservation control decision module 5 is used to predict the remaining shelf life of fresh agricultural products based on the cumulative damage index of fresh agricultural products in the future period, and to formulate preservation control strategies for fresh agricultural products based on the remaining shelf life.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the quality of fresh agricultural products during transit logistics, characterized in that, The method includes: S1. Collect multi-source datasets of fresh agricultural product transportation during the current time period. The multi-source datasets include vehicle driving data and environmental data inside the transport compartment. S2. Extract vehicle vibration characteristics from the driving data of transport vehicles, use mutual information feature selection algorithm to analyze the amount of interactive information between vehicle vibration characteristics and the biomechanical damage state of fresh agricultural products, and calculate the mechanical damage index of fresh agricultural products caused by transport vehicles during driving based on the amount of interactive information. S3. Extract the environmental features of the transport compartment from the environmental data inside the transport compartment, and calculate the environmental stress index based on the environmental features of the transport compartment; S4. Input the mechanical damage index and the environmental stress index into the pre-trained agricultural product damage prediction model, and output the cumulative damage index of fresh agricultural products in the future period; wherein, the agricultural product damage prediction model is trained using a multi-task learning strategy. S5. Based on the cumulative damage index of fresh agricultural products in the future period, predict the remaining shelf life of fresh agricultural products, and formulate fresh agricultural product preservation control strategies according to the remaining shelf life.

2. The method for monitoring the quality of fresh agricultural products during transit logistics according to claim 1, characterized in that, S2 includes: S21. Perform time-frequency domain analysis on the driving data of transport vehicles, extract vehicle vibration features from the driving data of transport vehicles based on the time-frequency domain analysis results, and integrate them into a vehicle vibration feature set. S22. Using the mutual information feature selection algorithm, each feature in the vehicle vibration feature set is associated and matched with the biomechanical damage state of fresh agricultural products in the target database. Based on the association and matching results, a damage feature set is selected from the vehicle vibration feature set. The damage feature set includes a subset of instantaneous impact features and a subset of cumulative fatigue features; S23. Calculate the instantaneous impact damage index and the cumulative fatigue damage index based on the instantaneous impact feature subset and the cumulative fatigue feature subset, respectively. Combine the coupling deterioration effect between instantaneous impact damage and cumulative fatigue damage to calculate the mechanical damage index caused to fresh agricultural products by transport vehicles during operation.

3. The method for monitoring the quality of fresh agricultural products during transit logistics according to claim 2, characterized in that, S22 includes: S221. Obtain the critical impact damage threshold and critical fatigue damage threshold of fresh agricultural products from the target database; S222. Compare each feature in the vehicle vibration feature set with the critical impact damage threshold and the critical fatigue damage threshold respectively. Based on the comparison results, divide the vehicle vibration feature set into an instantaneous impact feature set and a cumulative fatigue feature set. S223. Calculate the redundancy and local maximum interaction between each impact feature and other impact features in the instantaneous impact feature set, obtain the initial score of each impact feature according to the interaction expansion principle, and sort each impact feature according to the initial score. S224. Identify pairs of impact features that have positive interaction in the sorting order to form impact feature interaction groups. Continue to traverse the remaining impact features and add features that have positive synergy with the impact feature interaction groups to the group one by one until all impact features in the instantaneous impact feature set have been traversed to form a set of impact feature interaction groups. S225. Calculate the interaction weights between paired impact features within each impact feature interaction group, and correct the initial score of each impact feature. Based on the corrected comprehensive score, select the best subset of instantaneous impact features from the instantaneous impact feature set. S226. Select a subset of cumulative fatigue features from the cumulative fatigue feature set in accordance with the steps S223-S225.

4. The method for monitoring the quality of fresh agricultural products during transit logistics according to claim 3, characterized in that, S225 includes: S2251. Initialize the selected feature set, and add the impact feature with the highest initial score as the first selected feature to the selected feature set; S2252. Among the remaining unselected impact features, perform the following steps sequentially for each candidate impact feature: Based on the amount of interaction information between pairs of impact features within the interaction group of the candidate impact feature, and combined with the initial scores of each impact feature, calculate the pairwise interaction weight between the candidate impact feature and the features in the same group in the selected feature set. Calculate the amount of interaction information between the current candidate impact feature interaction group and the currently selected feature set to obtain the interaction group weight; The initial score, pairwise interaction weights, and interaction group weights of the candidate impact feature are multiplied together to obtain the current comprehensive score of the candidate impact feature. S2253. Select the candidate impact feature with the highest current comprehensive score and add it to the selected feature set; S2254. Repeat S2252 to S2253 until the number of selected feature sets reaches the preset threshold or the comprehensive score of no candidate impact features is higher than the set lower limit. Output the final selected feature set as the instantaneous impact feature subset.

5. A method for monitoring the quality of fresh agricultural products during transit logistics according to claim 4, characterized in that, S23 includes: S231. Input the subset of instantaneous impact features into the preset damage assessment model, and output the instantaneous impact damage index through the damage assessment model; S232. Input the cumulative fatigue feature subset into the preset damage assessment model, and output the cumulative fatigue damage index through the damage assessment model; S233. Based on the coupled deterioration effect of instantaneous impact damage and cumulative fatigue damage on fresh agricultural products, determine the dynamic fusion weight of instantaneous impact damage index and cumulative fatigue damage index. S234. Based on dynamic fusion weights, the instantaneous impact damage index and the cumulative fatigue damage index are weighted and fused to obtain the mechanical damage index that characterizes the entire transportation process of fresh agricultural products.

6. A method for monitoring the quality of fresh agricultural products during transit logistics according to claim 5, characterized in that, S233 includes: S2331. Obtain the instantaneous impact damage index sequence and cumulative fatigue damage index sequence for the target time step; S2332. Calculate the impact-fatigue transfer entropy from the instantaneous impact damage index sequence to the cumulative fatigue damage index sequence. S2333. Calculate the fatigue-impact transfer entropy from the cumulative fatigue damage index sequence to the instantaneous impact damage index sequence; S2334. Combining the impact-fatigue transfer entropy and the fatigue-impact transfer entropy, calculate the coupling strength coefficient that characterizes the causal driving proportion of instantaneous impact damage to cumulative fatigue damage. S2335. The coupling strength coefficient is used as the dynamic fusion weight of the instantaneous impact damage index, and the difference between the target value and the coupling strength coefficient is used as the dynamic fusion weight of the cumulative fatigue damage index.

7. A method for monitoring the quality of fresh agricultural products during transit logistics according to claim 1, characterized in that, S4 includes: S41. Obtain a multi-source historical dataset and divide it into a training set and a validation set. S42. Using the historical mechanical damage index and historical environmental stress index in the training set as model inputs and the historical cumulative damage measured values ​​as model output targets, the agricultural product damage prediction model is trained by combining a multi-task learning strategy to obtain the trained agricultural product damage prediction model. S44. The trained agricultural product damage prediction model is evaluated using a validation set. The evaluation metrics include the root mean square error and the coefficient of determination between the cumulative damage prediction value and the measured value. S45. After completing the model evaluation, the mechanical damage index and environmental stress index of the current period are used as accelerated deterioration factors and input into the trained agricultural product damage prediction model. The agricultural product damage prediction model outputs the cumulative damage index of fresh agricultural products in future periods.

8. A method for monitoring the quality of fresh agricultural products during transit logistics according to claim 7, characterized in that, S41 includes: S411. Obtain a multi-source dataset of fresh agricultural product transportation within a historical time period, wherein the multi-source dataset includes historical vehicle driving data and historical environmental data inside the transport compartment. S412. Following the steps S2-S3, calculate the historical mechanical damage index and historical environmental stress index based on historical transport vehicle driving data and historical environmental data. S413. Using the historical cumulative damage measured values ​​of fresh agricultural products on the corresponding road section as labels, the historical mechanical damage index and historical environmental stress index are paired with the labels to construct a training sample set, and the training sample set is divided into a training set and a validation set.

9. A method for monitoring the quality of fresh agricultural products during transit logistics according to claim 8, characterized in that, S42 includes: S421. Input the historical mechanical damage index and historical environmental stress index from the training set into the shared feature extraction network of the agricultural product damage prediction model to extract damage feature representations; the shared feature extraction network is a combination of fully connected layers or convolutional layers. S422. Input the damage feature representation into the main task prediction head and the auxiliary task prediction head respectively; the main task prediction head outputs the cumulative damage prediction value, and the auxiliary task prediction head outputs the damage dominance type classification result, wherein the damage dominance type includes mechanical dominance type and environmental dominance type. S423. The mean square error between the predicted value of the cumulative damage index and the measured value of the historical cumulative damage is used as the main task loss, and the cross-entropy loss between the damage dominant type classification result and the true dominant type label is used as the auxiliary task loss. The main task loss and the auxiliary task loss are weighted and summed to obtain the joint loss function. S424. Minimize the joint loss function using the backpropagation algorithm and gradient descent optimizer, and iteratively update the parameters of the shared feature extraction network, the main task prediction head, and the auxiliary task prediction head until the preset convergence condition is met, thus obtaining the trained agricultural product damage prediction model.

10. A system for monitoring the quality of fresh agricultural products during transit logistics, used to implement the method for monitoring the quality of fresh agricultural products during transit logistics as described in any one of claims 1-9, characterized in that, The system includes: The multi-source data acquisition module is used to collect multi-source datasets of fresh agricultural products transportation during the current time period. The multi-source datasets include transportation vehicle driving data and environmental data inside the transportation compartment. The mechanical damage quantification module is used to extract vehicle vibration characteristics from the driving data of transport vehicles, use the mutual information feature selection algorithm to analyze the amount of interactive information between vehicle vibration characteristics and the biomechanical damage state of fresh agricultural products, and calculate the mechanical damage index caused to fresh agricultural products during the driving process of transport vehicles based on the amount of interactive information. The environmental stress quantification module is used to extract the environmental characteristics of the transport compartment from the environmental data inside the transport compartment, and calculate the environmental stress index based on the environmental characteristics of the compartment. The cumulative damage prediction module is used to input the mechanical damage index and the environmental stress index into a pre-trained agricultural product damage prediction model and output the cumulative damage index of fresh agricultural products in the future period; wherein, the agricultural product damage prediction model is trained using a multi-task learning strategy. The freshness control decision module is used to predict the remaining shelf life of fresh agricultural products based on the cumulative damage index of fresh agricultural products in future time periods, and to formulate freshness control strategies for fresh agricultural products based on the remaining shelf life.

Images