A fruit sorting method and system based on group interaction effect modeling

Abstract

The application provides a fruit sorting method and system based on group interaction effect modeling, and relates to the technical field of fruit intelligent sorting. The method comprises the following steps: acquiring multi-modal detection data of a target object, state data of a candidate container, and physical transmission time delay data; predicting the reference prediction state of the target object at a plurality of prediction nodes based on the multi-modal detection data; for any candidate container, calculating the base group interaction effect based on the inventory state data of the existing objects in the container, and combining the nonlinear saturation constraint to correct the reference prediction state of the target object to obtain the target effective prediction state; incorporating the target effective prediction state into the candidate container and updating the global prediction state; evaluating the candidate sorting action based on the updated global prediction state, determining the target sorting action, and performing sorting control. The technical problem of increasing fruit damage due to maturity prediction deviation in the prior art is solved.

Description

Technical Field

[0001] This application relates to the field of intelligent fruit sorting technology, and in particular to a fruit sorting method and system based on group interaction effect modeling. Background Technology

[0002] In intelligent sorting scenarios for easily ripening fruits like durian, the accuracy of sorting decisions depends on the precision of predicting the future quality state of the fruit after it enters the warehouse. Existing sorting methods typically use the current quality state of the fruit as the sole basis for decision-making, or treat all fruits in the sorting warehouse as a homogeneous group for a general quality evolution prediction. In scenarios involving biological interactions such as ethylene ripening among fruits in the sorting warehouse, existing methods fail to distinguish the roles of the fruit to be sorted and the existing fruits in the warehouse in the interaction effects when predicting the future state of the fruit after it enters the warehouse. This leads to the fruit being repeatedly affected by its own released interaction effects during the prediction process, forming a self-catalyzing logical loop. This causes the predicted future quality state to deviate from the actual physical laws, resulting in distorted evaluation of sorting actions, biased selection of target sorting warehouses, and affecting the overall decision-making reliability of the sorting system. Therefore, a method is urgently needed to solve the technical problem of increased fruit loss due to the bias in maturity prediction in existing technologies. Summary of the Invention

[0003] This application provides a fruit sorting method and system based on group interaction effect modeling, which solves the technical problem of increased fruit loss due to maturity prediction deviation in the prior art.

[0004] To achieve the above objectives, this application adopts the following technical solution:

[0005] Firstly, a fruit sorting method based on group interaction effect modeling is provided. The method includes: acquiring multimodal detection data of the target item, state data of candidate containers, and physical transmission delay data; predicting the baseline prediction state of the target item at multiple prediction nodes based on the multimodal detection data; calculating the basis group interaction effect for any candidate container based on the existing state data of the items in the candidate container; correcting the baseline prediction state of the target item based on the basis group interaction effect combined with nonlinear saturation constraints to obtain the effective prediction state of the target item; incorporating the effective prediction state of the target item into the candidate container to update the global prediction state; evaluating candidate sorting actions based on the updated global prediction state, determining the target sorting action, and executing sorting control.

[0006] In conjunction with the first aspect mentioned above, in one possible implementation, based on multimodal detection data, the baseline predicted state of the target item at multiple prediction nodes is predicted, including: extracting and fusing features from the multimodal detection data to obtain the current state features of the target item; acquiring historical state data corresponding to the target item and inputting the current state features and historical state data into a time-series prediction model; predicting the state of the target item at multiple preset prediction nodes based on the time-series prediction model to obtain a baseline predicted state sequence of the target item; and when historical state data of the target item is lacking, determining the prior state change trajectory based on the basic attributes of the target item and generating a baseline predicted state sequence based on the prior state change trajectory.

[0007] In conjunction with the first aspect above, in one possible implementation, for any candidate container, the basic group interaction effect is calculated based on the stock status data of existing items within the candidate container, including: obtaining the predicted status data and interaction parameters corresponding to each existing item within the candidate container; determining the individual interaction effect generated by each existing item at the corresponding prediction node based on the predicted status data and corresponding interaction parameters of each existing item; and for each prediction node, aggregating the individual interaction effects of each existing item within the candidate container to obtain the basic group interaction effect of the candidate container at each prediction node.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, for each prediction node, the individual interaction effects of each existing item in the candidate container are aggregated, including: for any prediction node, obtaining the number of existing items in the candidate container; based on the number of existing items, normalizing and aggregating the individual interaction effects of each existing item at the prediction node; and determining the normalized aggregation result as the base group interaction effect corresponding to the prediction node.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the base group interaction effect... Satisfy the following formula:

[0010] .

[0011] Where N is the total number of items in the candidate container, and r is the index of the existing item. This is the basic release coefficient for a single product. As a maturity regulator, This represents the basic predicted state of the r-th existing item. For predicting node indexes.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the baseline predicted state of the target item is corrected based on the base group interaction effect combined with nonlinear saturation constraints to obtain the effective predicted state of the target item. This includes: obtaining the interaction sensitivity parameters corresponding to the target item; for each prediction node, determining the state correction amount generated by existing items in the candidate container on the target item based on the base group interaction effect and the interaction sensitivity parameters; constraining the state correction amount through a nonlinear saturation function to obtain the target state increment that meets the preset state change range; and correcting the baseline predicted state of the target item at the corresponding prediction node based on the target state increment to obtain the effective predicted state of the target item.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, the target effectively predicts the state. Satisfy the following formula:

[0014] .

[0015] in, The baseline predicted state for the target item. The interaction-sensitive parameters for the target item. It is the hyperbolic tangent function. As a base group interaction effect, For predicting node indexes.

[0016] In conjunction with the first aspect mentioned above, in one possible implementation, the effective predicted state of the target is incorporated into the candidate container, and the global predicted state is updated. This includes: incorporating the effective predicted state of the target as a new item state into the set of existing item states in the candidate container; redetermining the group interaction effect of each item in the candidate container at each prediction node based on the updated set of states; updating the predicted state of each item in the candidate container based on the redetermined group interaction effect; and generating the global predicted state corresponding to the candidate container based on the updated predicted state of each item.

[0017] In conjunction with the first aspect mentioned above, in one possible implementation, the process of evaluating candidate sorting actions based on the updated global predicted state, determining the target sorting action, and executing sorting control includes: obtaining the target predicted state corresponding to each candidate container; comparing the updated global predicted state with the target predicted state of the corresponding candidate container for each candidate sorting action to obtain the predicted state deviation; determining the predicted state concentration and container load state of the corresponding candidate container based on the global predicted state; standardizing the predicted state deviation, predicted state concentration, and container load state, and determining the comprehensive evaluation result of each candidate sorting action based on the standardization result; determining the target sorting action based on the comprehensive evaluation result of each candidate sorting action; and determining the triggering time of the sorting execution mechanism based on physical transmission delay data and the target sorting action, and controlling the sorting execution mechanism to perform sorting.

[0018] Secondly, a fruit sorting system based on group interaction effect modeling is provided. The system includes: a data acquisition module for acquiring multimodal detection data of the target item, state data of candidate containers, and physical transmission delay data; a state prediction module for predicting the baseline prediction state of the target item at multiple prediction nodes based on the multimodal detection data; a base group effect calculation module for calculating the base group interaction effect for any candidate container based on the existing state data of the items in the candidate container; a target effective state prediction module for correcting the baseline prediction state of the target item based on the base group interaction effect combined with nonlinear saturation constraints to obtain the target effective prediction state; a global state prediction module for incorporating the target effective prediction state into the candidate container and updating the global prediction state; and a sorting control module for evaluating candidate sorting actions based on the updated global prediction state, determining the target sorting action, and executing sorting control.

[0019] This application strictly limits the calculation scope of the base group interaction effect to the existing state data of items within the candidate container. This ensures that goods awaiting sorting do not participate in the generation of interaction effects during the first stage of the warehouse entry simulation, acting only as recipients of these effects. This cuts off the path for goods awaiting sorting to accelerate their own development, ensuring that the predicted future quality state conforms to real physical laws and improving the interpretability and reliability of sorting decisions. By limiting the state correction amount through nonlinear saturation constraints, the state increment caused by interaction effects is constrained within a preset range of state changes, avoiding the problem of predicted state overflow due to the accumulation of interaction effects and ensuring the numerical stability of the simulation results in long-term, multi-prediction-node scenarios. After the goods awaiting sorting are corrected and incorporated into the candidate container, the warehouse-wide group interaction effect is recalculated based on the updated state set. This incorporates the interaction effects between existing goods and the interaction effects of newly added goods on existing goods into the global predicted state update process, fully covering the real physical scenario of interaction effects between items after goods enter the warehouse, thus improving the accuracy of the global predicted state.

[0020] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0021] Figure 1 A system architecture diagram of a fruit sorting system based on group interaction effect modeling is provided for embodiments of this application;

[0022] Figure 2 A flowchart illustrating a fruit sorting method based on group interaction effect modeling provided in this application embodiment. Figure 1 ;

[0023] Figure 3 A flowchart illustrating a fruit sorting method based on group interaction effect modeling provided in this application embodiment. Figure 2 ;

[0024] Figure 4A flowchart illustrating a fruit sorting method based on group interaction effect modeling provided in this application embodiment. Figure 3 ;

[0025] Figure 5 A flowchart illustrating a fruit sorting method based on group interaction effect modeling provided in this application embodiment. Figure 4 ;

[0026] Figure 6 A flowchart illustrating a fruit sorting method based on group interaction effect modeling provided in this application embodiment. Figure 5 . Detailed Implementation

[0027] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0028] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0029] Example 1:

[0030] This embodiment provides a fruit sorting system based on group interaction effect modeling. The system includes a data acquisition module 101, a state prediction module 102, a base group effect calculation module 103, a target effective state prediction module 104, a global state prediction module 105, and a sorting control module 106. The data acquisition module 101 acquires multimodal detection data of the target item, state data of candidate containers, and physical transmission delay data. The state prediction module 102 predicts the baseline prediction state of the target item at multiple prediction nodes based on the multimodal detection data. The base group effect calculation module 103 calculates the base group interaction effect for any candidate container based on the existing state data of items within the candidate container. The target effective state prediction module 104 corrects the baseline prediction state of the target item based on the base group interaction effect combined with nonlinear saturation constraints to obtain the target effective prediction state. The global state prediction module 105 incorporates the target effective prediction state into the candidate container and updates the global prediction state. The sorting control module 106 is used to evaluate candidate sorting actions based on the updated global predicted state, determine the target sorting action, and execute sorting control.

[0031] Example 2:

[0032] To address the technical problem of increased fruit loss due to misjudgment of maturity in existing technologies, this application provides a fruit sorting method based on group interaction effect modeling. Figure 2 A flowchart illustrating the fruit sorting method based on group interaction effect modeling provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the method includes:

[0033] S201. Obtain multimodal detection data of the target item, status data of the candidate container, and physical transmission delay data.

[0034] Here, "target item" refers to a single item currently entering the sorting decision-making process, awaiting determination of its warehouse destination. "Multimodal detection data" refers to a collection of detection data covering different physical or chemical dimensions, obtained through simultaneous collection of data from various types of detection equipment on the target item. This data may include at least two of the following: visible light image data, spectral detection data, weight data, three-dimensional morphological contour data, and environmental temperature and humidity data of the detection area. "Candidate containers" refers to multiple optional containers available for the target item to be warehoused. The status data of the candidate containers includes the quantity of items already inside, the identification of each existing item, historical detection data for each existing item, and the rated capacity of the candidate container. "Physical transmission delay data" refers to the time required for the target item to be transmitted from the sorting decision-making location to the sorting execution mechanism corresponding to each candidate container.

[0035] In one possible implementation, when a target item enters the inspection area, the sorting system simultaneously collects multimodal inspection data of the target item using equipment such as industrial cameras, spectrometers, weighing sensors, 3D contour scanners, and temperature and humidity sensors deployed in the inspection area. Simultaneously, it collects real-time status data of candidate containers by reading container status records from the warehouse management system or by using sensors deployed at each candidate container. Physical transmission delay data is obtained through measured calibration of the conveyor line speed or by calculating the ratio of conveyor line length to speed. The timing of acquiring these three types of data can be flexibly configured according to the actual sorting cycle. For example, multimodal inspection data and candidate container status data can be acquired simultaneously when the target item enters the inspection area, and physical transmission delay data can be acquired after the target item completes inspection; alternatively, candidate container status data and physical transmission delay data can be acquired first, followed by multimodal inspection data of the target item.

[0036] It should be noted that this step acquires all the necessary data at the beginning of the sorting decision-making process, so that subsequent steps do not need to access external data sources again during the simulation and evaluation process, ensuring the consistency of data sources and the continuity of the simulation process, and avoiding the drift of the simulation benchmark due to inconsistent data acquisition timing.

[0037] As an example, when a durian enters the sorting line detection area, the sorting system simultaneously collects the durian's RGB image, close-up image of the stem area, surface texture data, net weight data, three-dimensional morphological contour data, spectral detection data, and real-time temperature and humidity data of the detection area. At the same time, it reads the current quantity of goods in each sorting bin, the set of goods identification tags, the rated capacity, and the historical detection data of goods in the bin, and marks the physical transmission delay from the conveyor line decision point to each sorting execution mechanism, forming a structured data set for subsequent steps.

[0038] Based on the above steps, the sorting system completes the unified acquisition of multimodal detection data of target items, candidate container status data and physical transmission delay data at the beginning of the decision-making process, so that subsequent steps are based on the same data benchmark and avoid interference to the inference results caused by inconsistent data sources.

[0039] S202. Based on multimodal detection data, predict the baseline prediction state of the target item at multiple prediction nodes.

[0040] In this context, prediction nodes refer to multiple pre-defined time points on a future timeline after the target item has entered the warehouse. The number and interval of prediction nodes can be configured according to business needs, such as tiered time points like 12 hours, 24 hours, 36 hours, 48 ​​hours, and 72 hours. The baseline prediction state refers to the predicted state value of the target item at each prediction node, without considering the interaction between the target item and existing items in the candidate container.

[0041] In one possible implementation, the sorting system first extracts and fuses features from the multimodal detection data to obtain the current state features of the target item. These current state features are then input into a temporal prediction model, which outputs a baseline predicted state sequence for the target item across multiple prediction nodes. The temporal prediction model can be implemented using recurrent neural networks, long short-term memory networks, gated recurrent units, or attention-based temporal models. For target items lacking historical state data, the sorting system can determine the prior state change trajectory based on the target item's fundamental attributes and generate a baseline predicted state sequence based on this trajectory.

[0042] It should be noted that the baseline predicted state output in this step only reflects the state change trend of the target item under its own inherent evolutionary laws, and does not include any interaction effects from existing items in the candidate container. This data boundary provides a clean input baseline for the independent calculation of the base group interaction effect and the step-by-step correction of the effective predicted state of the target item in subsequent steps, ensuring a clear separation between the target item's own evolutionary laws and external interaction effects during the deduction process.

[0043] Based on the above steps, the sorting system transforms multimodal detection data into a baseline predicted state sequence with a time dimension, expanding the future state of the target item from a static description of a single current moment to a dynamic trajectory of multiple predicted nodes, providing time-aligned baseline data for subsequent node calculation and correction of interactive effects.

[0044] S203. For any candidate container, calculate the basic group interaction effect based on the existing stock status data of the items in the candidate container.

[0045] Among them, the existing state data refers to the state data of the items already existing in the candidate container at the current moment, excluding any state data of the target item itself. The basic group interaction effect is a quantitative representation of the interaction influence of the existing item group in the candidate container on the target item.

[0046] In one possible implementation, the sorting system first acquires the predicted state data and interaction parameters corresponding to each existing item in the candidate container. Then, based on the predicted state data and corresponding interaction parameters of each existing item, it determines the individual interaction effect of each existing item at the corresponding prediction node. Finally, for each prediction node, the individual interaction effects of each existing item in the candidate container are aggregated to obtain the base group interaction effect of the candidate container at each prediction node. The interaction parameters include the basic release coefficient of a single item and a maturity adjustment factor. The basic release coefficient of a single item characterizes the baseline ability of an existing item to release interaction media per unit state, while the maturity adjustment factor characterizes the nonlinear adjustment effect of the state value of an existing item on the amount of interaction media released.

[0047] It should be noted that in calculating the basic group interaction effect, the scope of the calculation is strictly limited to the existing state data of items within the candidate container; the data of the target item itself is not included in any calculation in this step. This limitation cuts off the path for the target item to interact with itself at the data source level, so that the basic group interaction effect, in a physical sense, only represents the group interaction environment that already existed within the candidate container before the target item entered the warehouse.

[0048] Based on the above steps, the sorting system uses the existing inventory status data of the items in the candidate container as the sole basis for calculation, and generates a basic group interaction effect that does not include the contribution of the target item itself. This provides a clearly defined interaction effect input for the subsequent correction of the effective prediction status of the target, avoiding the paradox of the target item's self-catalysis.

[0049] S204. Based on the base group interaction effect combined with nonlinear saturation constraints, the baseline prediction state of the target item is corrected to obtain the effective prediction state of the target.

[0050] Nonlinear saturation constraint refers to limiting the state correction amount through a nonlinear function with saturation characteristics, causing the state correction amount to tend to saturate and grow nonlinearly when the input is large. The effective predicted state of the target item refers to the predicted state value at each prediction node after the target item has received corrections from the basis group interaction effects of existing items in the candidate container.

[0051] In one possible implementation, the sorting system first acquires the interaction-sensitive parameters corresponding to the target item. Then, for each prediction node, based on the basis group interaction effect and the interaction-sensitive parameters, it determines the state correction amount generated by existing items in the candidate container on the target item. The state correction amount is then constrained by a nonlinear saturation function to obtain the target state increment that satisfies a preset state change range. Finally, based on the target state increment, the baseline predicted state of the target item at the corresponding prediction node is corrected to obtain the effective predicted state of the target item. The interaction-sensitive parameters characterize the state response amplitude of the target item to the unit basis group interaction effect, and the nonlinear saturation function is implemented using functions with upper and lower bound convergence properties, such as the hyperbolic tangent function.

[0052] It should be noted that in this step, when correcting the baseline predicted state of the target item, only the base group interaction effect output in S203 is used as the source of the interaction influence. The target item itself does not participate in the generation of the interaction effect; the target item only acts as the receiver of the interaction effect during the correction phase. The effective predicted state of the target obtained after correction will then participate in the recalculation of the whole warehouse interaction effect as the agent of the interaction effect in the subsequent S205. This role switching order is consistent with the timing of goods entering the warehouse in the real physical scenario.

[0053] Based on the above steps, the sorting system limits the state correction amount caused by the interaction effect to a preset state change range through nonlinear saturation constraints, avoiding the problem of predicted state overflow caused by the accumulation of interaction effects. At the same time, it ensures that the target item only acts as the recipient of the interaction effect during the correction stage, thus guaranteeing rationality.

[0054] S205. Merge the effective prediction state of the target into the candidate container and update the global prediction state.

[0055] The global prediction state refers to the set of predicted states of all items in the candidate container at each prediction node after the candidate container accepts the target item.

[0056] In one possible implementation, the sorting system first incorporates the target effective prediction state as the state of the new item into the state set of existing items in the candidate container. Then, based on the updated state set, it redetermines the group interaction effect of each item in the candidate container at each prediction node. Based on the redetermined group interaction effect, it updates the prediction state of each item in the candidate container. Finally, based on the updated prediction state of each item, it generates the global prediction state corresponding to the candidate container.

[0057] In another possible implementation, the sorting system adopts an iterative update method. In the (m+1)th iteration, the group interaction effect is recalculated based on the state set updated in the mth iteration and the state of each item is corrected. When the maximum value of the change in the state of each item in two adjacent iterations is less than a preset threshold, convergence is determined, the iteration is terminated, and the current state set is output as the global predicted state. If the number of iterations reaches the preset upper limit and convergence is still not achieved, the weighted average of the results of the last two iterations is output.

[0058] It should be noted that after the target's effective prediction state is incorporated into the candidate container, this step recalculates the overall warehouse group interaction effect, including the target item. At this point, the target item has completed its role switch, changing from the recipient of the interaction effect to the exertor of the interaction effect. The interaction effects between existing goods and the interaction effects of new goods on existing goods are all included in the update process.

[0059] Based on the above steps, after the target item is corrected and incorporated into the candidate container, the sorting system recalculates the overall warehouse group interaction effect based on the updated state set. This ensures that the interaction between existing goods and the interaction between new goods and existing goods are both included in the global prediction state update process, improving the completeness and accuracy of the global prediction state.

[0060] S206. Based on the updated global predicted state, evaluate the candidate sorting actions, determine the target sorting action, and execute sorting control.

[0061] Among them, candidate sorting action refers to the optional action of sorting the target item into different candidate containers, and target sorting action refers to the optimal sorting action selected from the candidate sorting actions.

[0062] In one possible implementation, the sorting system first obtains the target predicted state corresponding to each candidate container. Then, for each candidate sorting action, it compares the updated global predicted state with the target predicted state of the corresponding candidate container to obtain the predicted state deviation. Based on the global predicted state, it determines the predicted state concentration and container load state of the corresponding candidate container. Then, it standardizes the predicted state deviation, predicted state concentration, and container load state, and determines the comprehensive evaluation result of each candidate sorting action based on the standardization result. Then, based on the comprehensive evaluation result of each candidate sorting action, it determines the target sorting action. Finally, based on the physical transmission delay data and the target sorting action, it determines the trigger time of the sorting execution mechanism and controls the sorting execution mechanism to perform sorting.

[0063] It should be noted that in this step, when determining the trigger time of the sorting execution mechanism, the physical transmission delay data is used to reverse compensate for the action time of the sorting execution mechanism, so that the sorting execution mechanism performs the sorting action when the target item actually arrives at the sorting port, avoiding sorting misalignment caused by transmission delay.

[0064] Based on the above steps, the sorting system comprehensively evaluates candidate sorting actions by standardizing the predicted state deviation, predicted state concentration, and container load status. This allows evaluation indicators of different dimensions to participate in decision-making on the same numerical scale, avoiding the problem of evaluation weight failure caused by differences in indicator magnitude. At the same time, physical transmission delay compensation ensures the timing accuracy of sorting actions.

[0065] Example 3:

[0066] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S202 can be specifically implemented through the following S301 to S304, which are explained in detail below:

[0067] S301. Perform feature extraction and fusion on the multimodal detection data to obtain the current state features of the target item.

[0068] Among them, the current state feature refers to the vectorized representation of the target item's current state obtained after feature extraction and fusion of the multimodal detection data of the target item.

[0069] In one possible implementation, the sorting system calls corresponding feature extraction networks to extract independent features from multimodal detection data of different dimensions. Visual image data is processed using convolutional neural networks or visual transformers to extract visual features; spectral detection data is processed using one-dimensional convolutional networks or fully connected networks to extract spectral features; weight data and three-dimensional morphology data are processed using multilayer perceptrons to extract joint features; and environmental temperature and humidity data are directly used as numerical features in subsequent fusion. The sorting system concatenates the feature vectors extracted from each modality and inputs them into a multimodal fusion network for fusion encoding to obtain the current state features of the target item. The multimodal fusion network can use an attention mechanism to weightedly fuse the features of each modality, or it can use a gating mechanism to adaptively adjust the contribution of each modality feature.

[0070] It should be noted that the specific network structure for feature extraction and fusion can be flexibly selected according to the type of target item and the modality of the detection data; this embodiment does not impose any limitations on this. When detection data for a certain modality is missing, the sorting system can fuse features based on the features of the remaining valid modalities, or it can first estimate the features of the missing modality through a cross-modal reconstruction network before fusing them, in order to ensure the continuity and integrity of the current state features.

[0071] As an example, the sorting system calls the corresponding feature extraction network to complete independent feature extraction for multimodal detection data of different dimensions. Visible light RGB images (resolution 1920×1080 pixels) are processed using a ResNet-50 convolutional neural network to extract visual features. This network is pre-trained on the ImageNet dataset and then fine-tuned on a durian image dataset, outputting a 2048-dimensional visual feature vector. Near-infrared spectral data (wavelength range 900-1700nm, sampling interval 2nm, total 401 dimensions) are processed using a one-dimensional convolutional neural network to extract spectral features. The network structure is as follows: First convolutional layer (kernel size 7, number of channels 64), batch normalization, ReLU, max pooling (window 2); Second convolutional layer (kernel size 5, number of channels 128), batch normalization, ReLU, global average pooling, fully connected layer (output 256 dimensions), resulting in a 256-dimensional spectral feature vector. Weight data and 3D morphological contour data (sampled to 1024 points from the farthest point) are used to extract joint features through a PointNet++ network, outputting a 512-dimensional morphological feature vector. Environmental temperature and humidity data (2D) are encoded into a 32-dimensional environmental feature vector through a fully connected layer. The sorting system concatenates the above four types of feature vectors along the channel dimension to form a 2848-dimensional original fused feature. The original fused feature is then input into a multimodal fusion network, which consists of two fully connected layers: the first layer has a 2848-dimensional input and a 1024-dimensional output, using ReLU activation and Dropout (dropout rate 0.3); the second layer has a 1024-dimensional input and a 256-dimensional output, using ReLU activation, ultimately yielding a 256-dimensional current state feature vector.

[0072] Based on the above steps, the sorting system transforms the multimodal detection data from raw heterogeneous data into a unified current state feature representation, enabling the information between different modalities to be aligned and complementary at the feature level. This provides a complete and dimensionally consistent input for the subsequent time series prediction model, avoiding the problem of inconsistent prediction model input caused by differences in the data formats of each modality.

[0073] S302. Obtain the historical state data corresponding to the target item, and input the current state characteristics and historical state data into the time series prediction model.

[0074] Historical status data refers to the set of status data recorded at previous detection nodes before the target item enters the current sorting process. Historical status data can include information such as the target item's status detection value at previous time points, the rate of status change, and the corresponding detection timestamp.

[0075] In one possible implementation, the sorting system retrieves the historical state data corresponding to the target item from a historical inspection database using the item's unique identifier. The unique identifier can be obtained through methods such as QR code scanning, RFID, or machine vision recognition. The sorting system aligns the retrieved historical state data with the current state features along the time dimension, forming a state data sequence containing both the current and historical moments. This state data sequence is then input into a time-series prediction model. The time-series prediction model can be implemented using recurrent neural networks, long short-term memory networks, gated recurrent units, or attention-based time-series models.

[0076] It should be noted that there are two scenarios regarding the acquisition of historical state data. When the target item has a record in the historical detection database, the sorting system inputs both the historical state data and the current state characteristics into the time series prediction model. When the target item does not have a record in the historical detection database, the sorting system can either input only the current state characteristics into the time series prediction model, or generate a baseline predicted state sequence based on the prior state change trajectory.

[0077] As an example, the time series prediction model employs a two-layer Long Short-Term Memory (LSTM) network. The input layer has a dimension of T×M, where T=10 (time step, i.e., inputting data from the 10 most recent time points) and M=258 (including a 256-dimensional current state feature vector, a 1-dimensional temperature value, and a 1-dimensional humidity value). The first LSTM layer has 128 hidden units and returns a sequence set to True; the second LSTM layer has 64 hidden units and returns a sequence set to False. A Dropout layer with a dropout rate of 0.2 is placed between the two LSTM layers. The fully connected output layer has a 64-dimensional input and a K-dimensional output (K is the preset number of prediction nodes; in this example, K=5). The model is trained before deployment using the Adam optimizer with an initial learning rate of 0.001, a batch size of 8, and a mean squared error loss function.

[0078] Based on the above steps, the sorting system uses the historical state data and current state characteristics of the target item as input to the time series prediction model, enabling the model to make predictions by simultaneously utilizing the long-term evolution pattern of the target item and the current state information, thereby improving the temporal consistency and prediction accuracy of the baseline predicted state sequence.

[0079] S303. Based on the time-series prediction model, predict the state of the target item at multiple preset prediction nodes to obtain the baseline prediction state sequence of the target item.

[0080] In one possible implementation, the sorting system uses the time information corresponding to preset prediction nodes as additional input to the time-series prediction model. The time-series prediction model outputs the baseline prediction status of the target item at each preset prediction node. The preset prediction nodes can be set to equally spaced time points, such as one prediction node every 12 hours, or non-equally spaced time points, such as one prediction node every 6 hours in the first 24 hours after warehousing, and one prediction node every 12 hours thereafter. The number and interval of preset prediction nodes can be flexibly configured according to factors such as the timeliness requirements of candidate containers and the rate of change of the target item's status.

[0081] It should be noted that the baseline predicted state output by the time series prediction model only reflects the state change trend of the target item under its own inherent evolution law, and does not include any interaction effects from existing items in the candidate container.

[0082] Based on the above steps, the sorting system expands the future state of the target item from a static description of the current moment to a dynamic trajectory of multiple prediction nodes through a time-series prediction model. This enables the calculation and correction of subsequent interaction effects at each node to be carried out on a unified time reference, avoiding the problem of time misalignment of the inference results caused by inconsistent prediction nodes.

[0083] S304. When historical state data of the target item is lacking, determine the prior state change trajectory based on the basic attributes of the target item, and generate a baseline predicted state sequence based on the prior state change trajectory.

[0084] Among them, basic attributes refer to the inherent attribute information of the target item that does not change with storage time, such as variety, origin, harvesting time, and initial specifications. The prior state change trajectory refers to a general trajectory that reflects the state change of this type of item over time, pre-established based on the statistical laws of similar items.

[0085] In one possible implementation, when the sorting system detects a lack of historical state data for a target item, it retrieves the average state change curve of similar items from a static prior knowledge base based on the target item's basic attributes. This retrieved average state change curve is then used as the target item's prior state change trajectory. The static prior knowledge base can be pre-built, and its construction can include statistical analysis of historical detection data for similar items and fitting and modeling the state change patterns of similar items. The sorting system uses the state values ​​corresponding to each preset prediction node of the prior state change trajectory as the baseline predicted state of the target item, generating a baseline predicted state sequence.

[0086] It should be noted that the method for determining the prior state change trajectory is not limited to retrieving the average state change curve of the group. In another possible implementation, the sorting system can calculate the prior state change trajectory based on the basic attributes of the target item using a parameterized state change model. The parameters of the parameterized state change model can be obtained by fitting historical detection data of similar items. In yet another possible implementation, the sorting system can retrieve the historical state change trajectories of several similar items most closely related to the target item from a static prior knowledge base based on the basic attributes of the target item, and use the weighted average of these historical state change trajectories as the prior state change trajectory.

[0087] Based on the above steps, when the target item lacks historical state data, the sorting system determines the prior state change trajectory through basic attributes and generates a baseline predicted state sequence, so that the target item in the cold start scenario can also obtain the baseline predicted state, avoiding the interruption of the prediction process due to the lack of historical data, and ensuring the continuity and availability of the sorting simulation process in the cold start scenario.

[0088] Example 4:

[0089] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S203 can be specifically implemented through the following S401 to S403, which are explained in detail below:

[0090] S401. Obtain the predicted state data and interaction parameters corresponding to each existing item in the candidate container.

[0091] Here, "existing items" refers to items that are currently in the candidate container, excluding target items that have not yet been added. "Interaction parameters" is a set of parameters used to quantify the contribution of each existing item to the overall interaction environment within the candidate container.

[0092] In one possible implementation, the sorting system extracts the identifiers of existing items within a candidate container from the container state data, and then retrieves the predicted state data of each existing item at each prediction node from the state prediction database based on these identifiers. Interaction parameters can be pre-stored in the item files of each existing item, and the sorting system reads them simultaneously when acquiring the predicted state data. These interaction parameters may include a basic release coefficient and a maturity adjustment factor. The basic release coefficient characterizes the baseline ability of an existing item to release interaction media per unit state, while the maturity adjustment factor characterizes the nonlinear adjustment effect of the existing item's state value on the amount of interaction media released. Taking durian sorting as an example, the basic release coefficient can be understood as the baseline rate of ethylene release from a single durian at a baseline maturity level, and the maturity adjustment factor can be understood as the degree to which the durian's own maturity amplifies or inhibits the ethylene release rate.

[0093] It should be noted that the specific source and form of the interaction parameters can be flexibly determined based on the type of item and the physicochemical properties of the interaction medium. For items such as durian, which produce ripening interactions by releasing ethylene, the interaction parameters can be obtained by regression mapping of the item's spectral and morphological characteristics, or they can be obtained by querying a preset parameter table based on the item's variety and origin. For other types of items, the interaction parameters can correspond to the release characteristic parameters of other types of interaction media; this embodiment does not limit this.

[0094] As an example, the sorting system extracts the identifiers of each existing item within a candidate container from the container status data for the currently being simulated candidate container. Then, based on these identifiers, it retrieves the predicted status data for each existing item at each prediction node from the status prediction database. Finally, based on the variety and origin information of each existing item, the sorting system queries the corresponding basic release coefficient for each item from a pre-built interaction parameter mapping table. and maturity regulators Values. The interaction parameter mapping table is pre-constructed as follows: For each variety and origin combination, select no fewer than 200 samples, place each sample in a constant temperature and humidity chamber (temperature 25±0.5°C, relative humidity 75±2%), and use an ethylene gas detector to measure the ethylene concentration in the chamber every 2 hours for 72 hours, recording the ethylene release rate of each sample. State value at the corresponding time The correspondence was established; for each sample, the ethylene release rate was regressed against the state value, and the regression model adopted... ,in, The error term is obtained by solving using the least squares method. and The estimated value; the above calibration process is performed on different varieties and origin combinations, and the calibration results are stored in the interactive parameter mapping table.

[0095] Based on the above steps, the sorting system has completed the acquisition of the predicted state data and interaction parameters of each existing item in the candidate container, so that the subsequent calculation of individual interaction effects is based on complete and consistent input data, avoiding the deviation in interaction effect calculation caused by incomplete data acquisition.

[0096] S402. Based on the predicted state data of each existing item and the corresponding interaction parameters, determine the individual interaction effect of each existing item at the corresponding prediction node.

[0097] Individual interaction effects refer to the contribution of a single existing item at a given prediction node to the overall interaction environment within the candidate container. The individual interaction effects of all existing items collectively form the basis of the overall interaction environment within the candidate container.

[0098] In one possible implementation, the sorting system, for each existing item in the candidate container, inputs the predicted state data of the existing item at a certain prediction node and the corresponding interaction parameters of the existing item into an individual interaction effect calculation function to obtain the individual interaction effect generated by the existing item at that prediction node. The specific form of the individual interaction effect calculation function can be determined according to the release law of the interaction medium. For ethylene ripening scenarios such as durian, the relationship between the individual interaction effect and the predicted state of the existing item is usually non-linear, which can be described by an exponential function, so that the individual interaction effect can reflect the non-linear amplification effect of the maturity of the existing item on the ethylene release rate.

[0099] Based on the above steps, the sorting system transforms the predicted state data and interaction parameters of each existing item into the individual interaction effects of each existing item at each prediction node. This enables the subsequent aggregation calculation of the group interaction effect to be carried out on the basis of a unified individual effect, avoiding deviations in the aggregation results caused by inconsistent calculation methods for individual effects.

[0100] S403. For each prediction node, aggregate the individual interaction effects of each existing item in the candidate container to obtain the basic group interaction effect of the candidate container at each prediction node.

[0101] Aggregation refers to combining the individual interaction effects of existing items within a candidate container at the same prediction node into a single value that represents the overall interaction level of the candidate container, according to preset rules. The base group interaction effect is a quantitative indicator obtained after aggregation, used to characterize the interaction influence of the group of existing items within the candidate container on the target item.

[0102] In one possible implementation, for any given prediction node, the sorting system obtains the quantity of existing items in the candidate container. Based on the quantity of existing items, it normalizes and aggregates the individual interaction effects of each existing item at that prediction node, and determines the normalized aggregation result as the base group interaction effect corresponding to that prediction node. Specifically, the normalization aggregation can be achieved by summing the individual interaction effects of each existing item and dividing by the quantity of existing items to obtain the mean of the individual interaction effects. Through mean normalization, the dimensions of the base group interaction effect are decoupled from the quantity of existing items in the candidate container, making the interaction effect levels comparable between candidate containers with different loading capacities.

[0103] Preferably, the basic group interaction effect Satisfy the following formula:

[0104] .

[0105] Where N is the total number of items in the candidate container, and r is the index of the existing item. This is the basic release coefficient for a single product. As a maturity regulator, This represents the basic predicted state of the r-th existing item. This is the index for the prediction node. The basic release coefficient of a single item represents the baseline ability of the r-th existing item to release interaction medium per unit state. The maturity adjustment factor represents the nonlinear adjustment effect of the state value of the r-th existing item on its release amount of interaction medium. The basic predicted state is the predicted state value of the r-th existing item at the k-th prediction node without considering the influence of group interaction.

[0106] It should be noted that the response characteristics of the exponential function form in the above formula to the maturity adjustment factor are as follows: when the maturity adjustment factor is positive, the higher the basic predicted state of existing items, the larger the value of the exponential term, and the stronger the individual interaction effect. This reflects the physical law that items with higher maturity release more interaction media. When the maturity adjustment factor is close to zero, the exponential term approaches 1, and the individual interaction effect is mainly determined by the basic release coefficient of the single item. This reflects the item types whose maturity has less impact on the release of interaction media. Mean normalization, by dividing by the total number of existing items N, eliminates the influence of the number of existing items on the dimension of the interaction effect. This ensures that the interaction effect of the base group reflects the average interaction level of existing items in the candidate container, rather than a simple sum of interaction effects, thus avoiding the problem of incomparable interaction effect values ​​due to differences in the loading capacity of candidate containers.

[0107] Based on the above steps, the sorting system merges the individual interaction effects of existing items in the candidate container into a basic group interaction effect through normalization aggregation, so that the basic group interaction effect can accurately represent the average group interaction environment that already exists in the candidate container before the target item enters the warehouse.

[0108] Example 5:

[0109] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S204 can be specifically implemented through the following S501 to S504, which are explained in detail below:

[0110] S501. Obtain the interaction-sensitive parameters corresponding to the target item.

[0111] The interaction sensitivity parameter refers to a parameter used to characterize the magnitude of the state response of a target item to the interaction effect of a unit base group. The physical meaning of the interaction sensitivity parameter can be understood as the upper limit of the state change that a target item can cause per unit interaction effect when influenced by the interaction of existing items within a candidate container. Taking durian sorting as an example, the interaction sensitivity parameter corresponds to a variety-specific ethylene sensitivity coefficient, representing the maximum ripening increase that this variety of durian can achieve under extreme ethylene concentrations.

[0112] In one possible implementation, the sorting system queries a preset parameter table for the interaction-sensitive parameters corresponding to the target item based on its basic attributes such as variety, origin, and harvest time. The preset parameter table can be pre-built based on historical detection data and interaction experiment data of similar items, and records the values ​​of interaction-sensitive parameters corresponding to different varieties and origins of items.

[0113] It should be noted that the methods for obtaining interaction-sensitive parameters are not limited to the two methods mentioned above. Interaction-sensitive parameters can also be dynamically adjusted by the sorting system based on the current state characteristics of the target item and the environmental parameters of the candidate container. For example, the interaction-sensitive parameter can be appropriately increased when the temperature inside the candidate container is high and appropriately decreased when the temperature is low, to reflect the influence of environmental factors on the interaction response amplitude of the target item. This embodiment does not limit the specific method for obtaining the interaction-sensitive parameter, as long as the parameter can characterize the response amplitude of the target item to the base group interaction effect.

[0114] As an example, the sorting system queries the corresponding interactive sensitive parameter values ​​from a pre-built interactive sensitive parameter mapping table based on the basic attributes of the target item, such as variety, origin, and harvest time. The interactive sensitive parameter mapping table is pre-built as follows: For each variety and origin combination, at least 100 samples are selected and placed in a constant temperature and humidity chamber (temperature 25±0.5°C, humidity 75±2%) with a saturated ethylene concentration (1000 ppm). The state value of the samples is measured every 6 hours for 72 hours. The maximum state change of each sample from its initial state to 72 hours is recorded, and the average of the maximum state changes of all samples under that variety and origin combination is calculated. This average is used as the calibration value of the interactive sensitive parameter for that combination. The calibration values ​​of each variety and origin combination are stored in the interactive sensitive parameter mapping table. Taking durian sorting as an example, the calibration value for Golden Pillow durian is 15-20 (dimensionless state value), and the calibration value for Musang King durian is 25-35.

[0115] S502. For each prediction node, based on the base group interaction effect and interaction sensitivity parameters, determine the state correction amount generated by the existing items in the candidate container on the target item.

[0116] Among them, the state correction amount refers to the state change amount corresponding to the interaction effect of the existing group of items in the candidate container on the target item. This state correction amount has not been subjected to nonlinear saturation constraints, and its value may exceed the reasonable range of change of the target item's state value.

[0117] In one possible implementation, the sorting system multiplies the base group interaction effect and the interaction sensitivity parameter for each prediction node. The product is then used as the state correction amount generated by the existing items in the candidate container for the target item at that prediction node. The state correction amount can be expressed as the product of the base group interaction effect and the interaction sensitivity parameter, where the base group interaction effect represents the average interaction level of the existing item group in the candidate container, and the interaction sensitivity parameter represents the response magnitude of the target item to a unit interaction effect. Multiplying the two yields the state change amount generated by the target item under the group interaction environment.

[0118] Based on the above steps, the sorting system transforms the baseline group interaction effect and interaction sensitivity parameters into state correction quantities corresponding to each prediction node. This transforms the group interaction impact from an abstract interaction effect index into a state change quantity with the same dimensions as the target item's state value, providing a directly usable input for subsequent nonlinear saturation constraints and baseline prediction state correction.

[0119] S503. Constrain the state correction amount through a nonlinear saturation function to obtain the target state increment that satisfies the preset state change range.

[0120] Here, the nonlinear saturation function refers to a nonlinear function with convergence characteristics at both upper and lower bounds. The preset state change range refers to the upper and lower limits of the state increment preset according to the reasonable range of values ​​of the target item's state value. The target state increment refers to the state correction amount that satisfies the preset state change range after being constrained by the nonlinear saturation function.

[0121] In one possible implementation, the sorting system uses the state correction value as input to a nonlinear saturation function. This function performs a nonlinear mapping on the state correction value, and the mapping result is used as the target state increment. The nonlinear saturation function can be implemented using a hyperbolic tangent function, whose output range is an open interval between -1 and +1, capable of mapping input values ​​of arbitrary size to this interval. The sorting system can directly use the state correction value as input to the hyperbolic tangent function, or it can first multiply the state correction value by a scaling factor before using it as input to adjust the sensitivity of the saturation function.

[0122] It should be noted that the mechanism of nonlinear saturation constraint is as follows: when the basis population interaction effect is large, the state correction is also large, but the output of the nonlinear saturation function approaches the upper limit of saturation and no longer increases linearly, thus limiting the target state increment to a preset range of state changes; when the basis population interaction effect is small, the output of the nonlinear saturation function is approximately equal to the input, and the target state increment and the state correction are approximately linearly related. This characteristic simulates the saturation phenomenon in biological enzyme kinetics, that is, after the concentration of the interaction medium reaches a certain level, the state response of the target item no longer increases linearly with the continued increase of the concentration of the interaction medium, but tends to a stable saturation value.

[0123] Based on the above steps, the sorting system constrains the state correction amount within a preset state change range through a nonlinear saturation function, so that the target state increment tends to saturate and grow nonlinearly when the interaction effect is strong, avoiding the problem of predicted state overflow caused by the accumulation of interaction effects, and ensuring the stability of the effective predicted state of the target in long-period multi-prediction node scenarios.

[0124] S504. Based on the incremental correction of the target state, the baseline prediction state of the target item at the corresponding prediction node is corrected to obtain the effective prediction state of the target.

[0125] In one possible implementation, the sorting system adds the target state increment to the baseline predicted state of the target item at each prediction node, and uses the addition result as the effective predicted state of the target item at that prediction node.

[0126] Preferably, the target effective prediction state Satisfy the following formula:

[0127] .

[0128] in, The baseline predicted state for the target item. The interaction-sensitive parameters for the target item. It is the hyperbolic tangent function. As a base group interaction effect, For predicting node indexes.

[0129] Based on the above steps, the sorting system superimposes the target state increment, after nonlinear saturation constraints, onto the baseline predicted state of the target item to obtain the effective predicted state of the target. The effective predicted state of the target retains the results of the inherent evolutionary laws of the target item itself, and also incorporates the unidirectional interactive influence of the existing item group in the candidate container on the target item. At the same time, the nonlinear saturation constraints ensure that the state value will not overflow the reasonable value range due to excessively strong interactive effects.

[0130] Example 6:

[0131] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 6 As shown, the above S206 can be implemented by the following S601 to S606, which are explained in detail below:

[0132] S601. Obtain the target prediction state corresponding to each candidate container.

[0133] In one possible implementation, the sorting system constructs a dynamically adapted target predicted state for each candidate container based on its order delivery time, historical shipping frequency, projected market sales volume, container rated capacity, logistics and transportation plan, and preset maturity matching standards. The target predicted state can be represented as a set of state distributions corresponding to each predicted node, such as the range of saleable item quantities, immature item quantities, and over-ripe item quantities corresponding to each predicted node. The target predicted state can be constructed based on historical shipping data statistics, business rule template configuration, or by solving an optimization model.

[0134] It should be noted that the target predicted state is acquired at the beginning of the sorting decision-making process, along with the state data of the candidate containers. The specific format of the target predicted state is consistent with that of the global predicted state, so that subsequent deviation comparisons can be performed on the same data structure.

[0135] Based on the above steps, the sorting system determines the target predicted state that matches the business needs for each candidate container, so that the evaluation of subsequent candidate sorting actions has a clear benchmark and avoids the problem of making decisions based solely on the absolute value of the global predicted state without business orientation.

[0136] S602. For each candidate sorting action, compare the updated global prediction state with the target prediction state of the corresponding candidate container to obtain the prediction state deviation.

[0137] The predicted state deviation refers to the quantified difference between the global predicted state and the target predicted state of a candidate container after executing a certain candidate sorting action. The larger the predicted state deviation, the greater the deviation between the future state of the candidate container after the candidate sorting action and the business expectation.

[0138] In one possible implementation, for each candidate sorting action, the sorting system compares the global predicted state corresponding to that action with the target predicted state of the corresponding candidate container at each prediction node, obtaining the state difference value for each prediction node. Then, the state difference values ​​of each prediction node are weighted and summed to obtain the predicted state deviation corresponding to the candidate sorting action. The weight of each prediction node can be configured according to the business timeliness priority; for example, the prediction node closer to the current time has a higher weight, reflecting the importance of recent state matching. The calculation method for the state difference value can include absolute difference, squared difference, or weighted absolute difference, etc.

[0139] It should be noted that both the global prediction state and the target prediction state contain state information corresponding to multiple prediction nodes. The calculation of the prediction state deviation needs to be performed on the same prediction node to avoid miscalculation of deviation due to inconsistency of prediction nodes.

[0140] Based on the above steps, the sorting system quantifies the difference between the global predicted state and the target predicted state of each candidate sorting action as the predicted state deviation, so that the matching accuracy of different candidate sorting actions has a unified quantitative comparison basis, providing the first evaluation index for subsequent comprehensive evaluation.

[0141] S603. Based on the global prediction status, determine the prediction status concentration and container load status of the corresponding candidate containers.

[0142] Among them, the predicted state concentration refers to the ratio of the peak number of items in the effective state interval of each predicted node to the rated capacity of the container after the candidate container accepts the target item. It is used to characterize the degree of risk of shipping congestion or over-ripeness of the candidate container in the future period.

[0143] In one possible implementation, for each candidate sorting action, the sorting system extracts the number of items in the effective state range of each predicted node from the global predicted state corresponding to that candidate sorting action. The maximum value of the corresponding number for each predicted node is taken as the peak quantity. The peak quantity is divided by the rated capacity of the candidate container to obtain the predicted state concentration. The predicted state concentration is a dimensionless ratio, typically ranging from 0 to 1. The sorting system adds 1 to the current number of items in the candidate container and divides it by the rated capacity to obtain the container load state. The container load state is also a dimensionless ratio, ranging from 0 to 1.

[0144] It should be noted that both the predicted state concentration and the container load state are determined based on the global predicted state, which has been updated in the aforementioned embodiments. It includes the full warehouse interaction effect after the target item is incorporated into the candidate container. Therefore, the predicted state concentration and the container load state can reflect the true future state of the target item after it is placed in the warehouse.

[0145] Based on the above steps, the sorting system extracts two evaluation indicators from the global predicted state: the concentration of predicted state and the container load state. This allows the evaluation of candidate sorting actions to consider not only state deviations but also future shipping congestion risks and current capacity occupancy, providing a second and third evaluation indicator for subsequent comprehensive evaluation.

[0146] S604. Standardize the predicted state deviation, predicted state concentration, and container load state, and determine the comprehensive evaluation result of each candidate sorting action based on the standardization results.

[0147] In one possible implementation, the sorting system performs standardization on the predicted state deviation, predicted state concentration, and container load status. Standardization can be achieved using Z-Score standardization to obtain a standardized score. The sorting system then multiplies the standardized predicted state deviation, predicted state concentration, and container load status by their respective evaluation weights and sums them to obtain a comprehensive evaluation result for each candidate sorting action. The evaluation weights can be dynamically adjusted according to the business scenario; for example, increasing the weight of predicted state deviation during peak shipping periods and increasing the weight of container load status when container resources are scarce.

[0148] Based on the above steps, the sorting system transforms the predicted state deviation, predicted state concentration, and container load state to the same numerical scale through standardization processing. This allows evaluation indicators of different dimensions to participate fairly in the decision-making process according to preset weights in the comprehensive evaluation, avoiding the problem of evaluation weight failure caused by differences in indicator magnitudes and improving the rationality of the comprehensive evaluation results.

[0149] S605. Based on the comprehensive evaluation results of each candidate sorting action, determine the target sorting action.

[0150] The target sorting action refers to the sorting action with the best overall evaluation result selected from all candidate sorting actions. The criteria for determining the best overall evaluation result can be configured according to business needs; for example, the best result can be the smallest or the largest overall evaluation result.

[0151] In one possible implementation, the sorting system ranks the comprehensive evaluation results of each candidate sorting action and selects the candidate sorting action with the best comprehensive evaluation result as the target sorting action. When the comprehensive evaluation result is dominated by deviation-type indicators, the smaller the comprehensive evaluation result, the better the candidate sorting action. The sorting system selects the candidate sorting action with the smallest comprehensive evaluation result as the target sorting action. When multiple candidate sorting actions have the same or similar comprehensive evaluation results, the sorting system can perform secondary screening by combining auxiliary factors such as the priority of candidate containers and batch information of target items.

[0152] It should be noted that the determination of the target sorting action is based solely on the comprehensive evaluation results obtained from S604. The comprehensive evaluation results have integrated three indicators: predicted state deviation, predicted state concentration, and container load state. Therefore, the target sorting action is the optimal choice after comprehensive consideration under multiple constraints, rather than a one-sided decision based solely on a single indicator.

[0153] Based on the above steps, the sorting system selects the target sorting action from the candidate sorting actions based on a comprehensive evaluation of multiple constraints. This enables the sorting decision to simultaneously consider the accuracy of state matching, the risk of future shipping congestion, and the current capacity occupancy, thus avoiding the one-sidedness of decision-making based on a single indicator.

[0154] S606. Based on physical transmission delay data and target sorting actions, determine the trigger time of the sorting execution mechanism and control the sorting execution mechanism to perform sorting.

[0155] Physical transmission delay data refers to the time required for the target item to be transferred from the sorting decision location to the sorting execution mechanism corresponding to each candidate container. The sorting execution mechanism refers to the mechanical actuation component used to move the target item from the conveyor line into the target container, such as a fork, push rod, cylinder, or swing arm. The trigger moment refers to the point in time when the sorting execution mechanism begins to perform the sorting action.

[0156] In one possible implementation, the sorting system, based on the candidate container corresponding to the target sorting action, queries the physical transmission delay data for that candidate container. It then adds the physical transmission delay to the current decision time and subtracts the mechanical response safety margin of the sorting actuator to obtain the trigger time of the sorting actuator. When the trigger time arrives, the sorting system sends a sorting control command to the sorting actuator, controlling it to perform the sorting action and move the target item into the target container. The mechanical response safety margin of the sorting actuator refers to the time required for the actuator to complete the sorting action from receiving the control command; this time can be pre-calibrated.

[0157] It should be noted that the physical transmission delay data is used for reverse compensation of the sorting execution mechanism's action timing. Without delay compensation, if the sorting execution mechanism executes the sorting action immediately at the decision moment, but the target item has not yet arrived at the sorting port, the sorting action will be missed or misaligned. By delaying the trigger moment by the physical transmission delay and increasing the mechanical response safety margin, the sorting execution mechanism can complete the sorting action at the moment the target item actually arrives at the sorting port, achieving precise alignment between the sorting action and the arrival time of the target item.

[0158] Based on the above steps, the sorting system uses physical transmission delay data to reverse-compensate the triggering time of the sorting actuator, so that the action time of the sorting actuator is precisely aligned with the actual arrival time of the target item at the sorting port. This avoids the problem of sorting actions being missed or misaligned due to transmission delay, and ensures the timing accuracy of sorting control.

[0159] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0160] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

Claims

1. A fruit sorting method based on group interaction effect modeling, characterized in that, The method includes: Acquire multimodal detection data of the target item, state data of the candidate container, and physical transmission latency data; Based on the multimodal detection data, the baseline prediction state of the target item at multiple prediction nodes is predicted; For any candidate container, the base group interaction effect is calculated based on the existing stock status data of the items in the candidate container; Based on the aforementioned base group interaction effect combined with nonlinear saturation constraints, the baseline prediction state of the target item is corrected to obtain the effective prediction state of the target. The effective prediction state of the target is incorporated into the candidate container, and the global prediction state is updated. Based on the updated global predicted state, candidate sorting actions are evaluated, target sorting actions are determined, and sorting control is executed.

2. The method according to claim 1, characterized in that, The step of predicting the baseline prediction state of the target item at multiple prediction nodes based on the multimodal detection data includes: Feature extraction and fusion are performed on the multimodal detection data to obtain the current state features of the target item; Obtain the historical state data corresponding to the target item, and input the current state features and the historical state data into the time series prediction model; Based on the time-series prediction model, the state of the target item at multiple preset prediction nodes is predicted to obtain the baseline prediction state sequence of the target item. When the target item lacks historical state data, a priori state change trajectory is determined based on the basic attributes of the target item, and the baseline predicted state sequence is generated based on the priori state change trajectory.

3. The method according to claim 1, characterized in that, The calculation of the base group interaction effect based on the existing inventory status data of items within any candidate container includes: Obtain the predicted state data and interaction parameters corresponding to each existing item in the candidate container; Based on the predicted state data of each existing item and the corresponding interaction parameters, determine the individual interaction effect of each existing item at the corresponding prediction node. For each prediction node, the individual interaction effects of existing items within the candidate container are aggregated to obtain the basic group interaction effect of the candidate container at each prediction node.

4. The method according to claim 3, characterized in that, The aggregation of individual interaction effects of existing items within the candidate container for each prediction node includes: For any prediction node, obtain the number of items already in the candidate container; Based on the quantity of the existing items, the individual interaction effects of each existing item at the prediction node are normalized and aggregated. The normalized aggregation result is determined as the basis population interaction effect corresponding to the predicted node.

5. The method according to claim 3, characterized in that, The basic group interaction effect Satisfy the following formula: ； Where N is the total number of items already in the candidate container, and r is the index of the existing items. This is the basic release coefficient for a single product. As a maturity regulator, This represents the basic predicted state of the r-th existing item. For predicting node indexes.

6. The method according to claim 1, characterized in that, The process of correcting the baseline prediction state of the target item based on the base group interaction effect combined with nonlinear saturation constraints to obtain the effective prediction state of the target includes: Obtain the interaction-sensitive parameters corresponding to the target item; For each prediction node, based on the base group interaction effect and the interaction sensitivity parameter, determine the state correction amount generated by the existing items in the candidate container on the target item; The state correction amount is constrained by a nonlinear saturation function to obtain the target state increment that satisfies the preset state change range; Based on the target state increment, the baseline prediction state of the target item at the corresponding prediction node is corrected to obtain the effective prediction state of the target.

7. The method according to claim 6, characterized in that, The target effective prediction state Satisfy the following formula: ； in, The baseline predicted state for the target item. The interaction-sensitive parameters for the target item. It is the hyperbolic tangent function. As a base group interaction effect, For predicting node indexes.

8. The method according to claim 1, characterized in that, The step of incorporating the effective prediction state of the target into the candidate container and updating the global prediction state includes: The effective predicted state of the target is added as the state of the new item and incorporated into the state set of existing items in the candidate container. Based on the updated state set, the group interaction effect of each item in the candidate container at each prediction node is re-determined; Based on the redefined group interaction effect, the predicted state of each item in the candidate container is updated. Based on the updated predicted states of each item, a global predicted state corresponding to the candidate container is generated.

9. The method according to claim 1, characterized in that, The process of evaluating candidate sorting actions based on the updated global predicted state, determining the target sorting action, and executing sorting control includes: Obtain the target prediction state corresponding to each candidate container; For each candidate sorting action, the updated global prediction state is compared with the target prediction state of the corresponding candidate container to obtain the prediction state deviation. Based on the global prediction state, the prediction state concentration and container load state of the corresponding candidate containers are determined. The predicted state deviation, the predicted state concentration, and the container load state are standardized, and the comprehensive evaluation result of each candidate sorting action is determined based on the standardization result. Based on the comprehensive evaluation results of each candidate sorting action, the target sorting action is determined; Based on the physical transmission delay data and the target sorting action, the trigger time of the sorting execution mechanism is determined, and the sorting execution mechanism is controlled to perform sorting.

10. A fruit sorting system based on group interaction effect modeling, characterized in that, The system includes: The data acquisition module is used to acquire multimodal detection data of the target item, status data of the candidate container, and physical transmission delay data; The state prediction module is used to predict the baseline prediction state of the target item at multiple prediction nodes based on the multimodal detection data. The basic group effect calculation module is used to calculate the basic group interaction effect for any candidate container based on the stock status data of existing items in the candidate container. The target effective state prediction module is used to correct the baseline prediction state of the target item based on the basis group interaction effect combined with nonlinear saturation constraints, so as to obtain the target effective prediction state. The global state prediction module is used to incorporate the effective predicted state of the target into the candidate container and update the global prediction state. The sorting control module is used to evaluate candidate sorting actions based on the updated global predicted state, determine the target sorting action, and execute sorting control.

Images