A real-time adjustment method and system for a fresh stewed bird's nest periodic distribution path

Abstract

The present application relates to the technical field of cold chain cycle distribution path optimization, and discloses a real-time adjustment method and system for fresh stewed bird's nest cycle distribution path, which comprises collecting order data, package temperature sequence and recovery temperature control box temperature sequence; inputting batch parameters and real-time temperature into Arrhenius recursion relationship to generate freshness index and preservation hard constraint of each package; calculating predicted arrival freshness along path time line and converting it into preservation potential; forming a preservation vulnerability thermal map through counterfactual reasoning, cutting a disturbance subgraph when a dynamic disturbance occurs, triggering temperature sampling or road condition confirmation for fragile nodes; inputting the subgraph into a graph attention strategy network with physical modulation to modulate attention weight caused by service order difference, adding a non-negative part of potential change to strategy loss, and outputting the adjusted distribution path. The present application realizes real-time preservation constraint adjustment of cycle distribution path.

Description

Technical Field

[0001] This invention relates to the field of cold chain cycle delivery route optimization technology, specifically to a method and system for real-time adjustment of the cycle delivery route of fresh stewed bird's nest. Background Technology

[0002] For short-shelf-life foods such as freshly stewed bird's nest, subscription-based delivery typically requires meeting customer requirements regarding delivery time windows, vehicle routes, temperature control box temperatures, and product freshness simultaneously. Adjustments to existing delivery routes are usually based on order location, estimated travel time, and traffic conditions. Delivery sequences are replanned when new orders are inserted, addresses change, or temporary traffic disruptions occur.

[0003] However, in the cold chain delivery scenario of freshly stewed bird's nest, product freshness is not only related to delivery time, but is also affected by product batch parameters, real-time temperature of the temperature-controlled box, temperature disturbances caused by opening and closing the door, the temperature-controlled environment in which the package is located, and the residual temperature state after receipt. If route adjustments are calculated only based on distance or time windows, it is difficult to reflect the constraints of parcel-by-parcel freshness decay on the delivery sequence, and it is also difficult to identify which order nodes are more likely to experience a decline in freshness upon delivery due to route disturbances.

[0004] Furthermore, when new orders are inserted, addresses change, or temperatures become abnormal during delivery, directly resolving the global path can easily introduce node changes unrelated to the current disturbance; conversely, adjusting only local locations may overlook the freshness vulnerability of neighboring orders. Existing methods also typically treat temperature sampling, driver traffic confirmation, and route rescheduling as independent processes, lacking a coordinated mechanism to determine whether to add sampling or confirmation based on changes in freshness uncertainty.

[0005] Therefore, this application provides a method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest to solve the above-mentioned technical problems. Summary of the Invention

[0006] The purpose of this invention is to provide a real-time adjustment method and system for the periodic delivery route of fresh stewed bird's nest, in order to solve the technical problem of insufficient coupling between delivery route, parcel freshness decay, temperature anomaly, order disturbance and recycling temperature feedback in the existing periodic delivery route adjustment.

[0007] To address the aforementioned technical problems, this invention provides a method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest, comprising: Collect periodic subscription orders, product batches, customer coordinates and delivery time windows, vehicle locations, temperature sequences of temperature control boxes, door opening and closing records, and residual temperature sequences of recycled temperature control boxes; The initial freshness index, pre-exponential factor, activation energy, real-time temperature and sampling interval of the product batch are input into the Arrhenius recursive relation to form the current freshness index, remaining shelf life and hard constraint boundary of the shelf life for each package. Calculate the predicted freshness index of unfinished orders along the current path timeline, and convert the predicted freshness index and the minimum acceptable freshness threshold into preservation potential energy. Perform counterfactual reasoning on the order node status offset, and generate a preservation vulnerability heatmap based on the statistical changes in delivery freshness caused by the status offset; When a new order is inserted, the address is changed, or the temperature of the temperature control box deviates from the preset temperature range, the disturbance sub-map is cut by the heat map of the disturbance location and the vulnerability of preservation, and the temperature sampling or road condition confirmation action is triggered on the vulnerable nodes in the disturbance sub-map. The perturbation subgraph, preservation potential energy, time window, travel time, traffic status, and temperature rise coefficient are input into a graph attention policy network with physical modulation. The attention weights between nodes are modulated according to the difference in preservation potential energy caused by the service order. The non-negative part of the change in preservation potential energy before and after the action is added to the policy loss, and the adjusted delivery route and information collection instructions are output.

[0008] In some specific embodiments, the formation of the current freshness index, remaining shelf life, and hard constraint boundary for each package further includes: Read the package temperature within the sampling interval according to the temperature sequence of the temperature control box; The packaging temperature, pre-finger factor, and activation energy are input into the temperature decay term, and the retention result of the freshness index at the previous moment after the effect of the temperature decay term is used as the freshness index at the next moment. The remaining shelf life is calculated by accumulating the freshness index along the unfinished delivery route at the next moment. A critical package marker is generated when the remaining shelf life is lower than the preset warning condition, and a normal package marker is retained when the remaining shelf life is not lower than the preset warning condition.

[0009] In some specific embodiments, performing counterfactual reasoning on the order node state offset further includes: Insertion location, insertion order parameters, adjacent order service order, adjacent order time window, adjacent package temperature box superimposed heat effect, and road segment traffic status are written into the variable nodes of the structural causal model; Estimate causal edges between variable nodes using historical delivery samples; Apply state parameter offset intervention to nodes with incomplete orders, compare the freshness index of neighboring orders before and after the intervention, and write the maximum or average value of the difference in freshness index into the preservation vulnerability heat map.

[0010] In some specific embodiments, segmenting the perturbation submap using a heatmap of perturbation location and preservation vulnerability further includes: Use the location of the new order insertion, the location of the address change, or the location of the package with abnormal temperature as the disturbance seed; Read order nodes affected by disturbed seeds from the heat map of preservation vulnerability; Expand the neighborhood of the affected order node according to the constraints of historical path segments, estimated travel time and delivery time window; Neighborhood edges that do not meet the delivery time window constraint or the hard freshness constraint are removed, and perturbation subgraphs are generated for the remaining order nodes and path edges.

[0011] In some specific embodiments, triggering temperature sampling or road condition confirmation actions on vulnerable nodes in the perturbation subgraph further includes: For vulnerable nodes, simulate actions such as shortening the temperature sampling interval, confirming driver road conditions, and maintaining the sampling frequency. The structural causal model is used to estimate the uncertainty change of preservation potential energy before and after the execution of each type of action; The amount of uncertainty reduction is calculated by subtracting the sensor power consumption, communication traffic, and driver confirmation delay. The action with the largest difference is written into the features of the perturbation subgraph node, and the collected return data is input into the graph attention policy network with physical modulation.

[0012] In some specific embodiments, the modulation of inter-node attention weights based on the preservation potential difference caused by the service order further includes: Calculate the preservation potential energy of the second order node when the first order node serves before the second order node, and use it as the first preservation potential energy; Calculate the preservation potential energy of the second order node when the second order node serves before the first order node, and use it as the second preservation potential energy; The difference between the first preservation potential energy and the second preservation potential energy is input into the exponential modulation term; The attention coefficients of the standard graph are modulated by an exponential modulation term and then normalized to generate freshness causal attention weights; Input the causal attention weight of preservation into the service sequence action selection.

[0013] In some specific embodiments, after outputting the adjusted delivery route and information collection instructions, the process further includes: Substituting the forward delivery temperature sequence into the preservation potential energy calculation yields the predicted delivery preservation potential energy, and substituting the residual temperature sequence of the recovery temperature control box into the Arrhenius recursive relation yields the actual preservation potential energy. Calculate the duality loss between the predicted delivery preservation potential energy and the actual preservation potential energy; When the duality loss exceeds the preset trigger threshold, the target batch activation energy correction factor is updated and corrected online using Bayesian method and then input into the next round of freshness index calculation. The target batch activation energy correction factor is retained when the duality consistency loss does not exceed the preset trigger threshold. When deploying across scenarios, the physical conservation regularization term is added to the gradient update of the meta-learning inner layer to form the initial parameters of the policy network, and the initial parameters of the policy network are updated with local periodic delivery data.

[0014] In some specific embodiments, writing the action with the largest difference into the perturbation subgraph node features further includes: Establish a candidate queue for data collection actions; The uncertain range of preservation potential energy of vulnerable nodes, the remaining power of the temperature control box, the communication delay, and the driver confirmation delay are written into the candidate action attributes respectively; Candidate actions that do not meet the power or delay conditions will be eliminated. Candidate actions are sorted according to the difference between the reduction in uncertainty and the acquisition cost, and the action with the highest ranking is used to generate an information acquisition instruction.

[0015] In some specific embodiments, updating the target batch activation energy correction factor using online Bayesian methods further includes: Extract the temperature integral and receipt timestamp after receiving from the residual temperature sequence of the recycled temperature control box; The temperature integral after receiving the goods is input into the Arrhenius recursive relation to deduce the actual freshness index. The posterior distribution of the activation energy correction factor for the target batch is updated by the difference between the predicted freshness index and the actual freshness index. When the posterior distribution variance meets the preset convergence condition, the posterior mean is input into the next round of package freshness recursion. When the posterior distribution variance does not meet the preset convergence condition, the activation energy correction factor of the target batch in the previous round is retained and the recovery temperature sequence continues to be accumulated.

[0016] Based on the same concept, the present invention also provides a real-time adjustment system for the periodic delivery route of freshly stewed bird's nest, comprising: The data aggregation module is configured to collect periodic subscription orders, product batches, customer coordinates and receiving time windows, vehicle locations, temperature sequences of temperature control boxes, door opening and closing records, and residual temperature sequences of recycled temperature control boxes. The freshness recursion module is configured to input the initial freshness index, pre-exponential factor, activation energy, real-time temperature and sampling interval of the product batch into the Arrhenius recursion relation, and form the current freshness index, remaining shelf life and shelf hard constraint boundary for each package. The preservation potential energy module is configured to calculate the predicted delivery freshness index of unfinished orders along the current path timeline, and convert the predicted delivery freshness index and the minimum acceptable freshness threshold into preservation potential energy. The causal reasoning module is configured to perform counterfactual reasoning on the order node state offset and generate a heat map of preservation vulnerability based on the statistical changes in delivery freshness caused by the state offset. The disturbance response module is configured to cut a disturbance sub-map based on the disturbance location and the preservation vulnerability heatmap when a new order is inserted, the address is changed, or the temperature of the temperature control box deviates from the preset temperature range, and trigger temperature sampling or road condition confirmation actions for vulnerable nodes in the disturbance sub-map. The path solving module is configured to input the perturbation subgraph, preservation potential energy, time window, travel time, traffic status and temperature rise coefficient into a graph attention policy network with physical modulation, modulate the attention weights between nodes according to the difference in preservation potential energy caused by the service order, and add the non-negative part of the change in preservation potential energy before and after the action to the policy loss, and output the adjusted delivery path and information collection instructions.

[0017] Compared with existing technologies, its advantages are as follows: This invention discloses a real-time adjustment method and system for the periodic delivery route of fresh stewed bird's nest. The method inputs product batch parameters, initial freshness index, real-time temperature and sampling interval into the Arrhenius recursive relation, and forms the current freshness index, remaining shelf life and hard constraint boundary of each package. The route adjustment is no longer based solely on distance and time window, but can incorporate the freshness differences formed by different packages under different temperature conditions into the delivery order adjustment.

[0018] The predicted freshness index is converted into preservation potential energy, and a preservation vulnerability heat map is formed through counterfactual reasoning. When a new order is inserted, the address is changed, or the temperature of the temperature control box deviates from the preset temperature range, the order nodes that are more easily affected can be identified around the disturbed seed, and the subsequent path adjustment is carried out with the perturbation subgraph to reduce the impact of changes in irrelevant nodes on the current delivery path.

[0019] The simulation of actions such as shortening the temperature sampling interval, confirming driver road conditions, and maintaining the collection frequency is performed on vulnerable nodes in the perturbation subgraph. The information collection action is selected based on the difference between the reduction in the uncertainty of preservation potential energy and the collection cost. Temperature sampling and road condition confirmation can serve to reduce the risk of freshness.

[0020] In the graph attention strategy network, the attention weights between nodes are modulated according to the difference in preservation potential energy caused by the service order, and the non-negative part of the change in preservation potential energy before and after the action is added to the strategy loss. The path solving process can simultaneously consider the time window, travel time, traffic conditions and freshness decay constraints, making the output path more in line with the preservation constraints in the cold chain distribution of fresh stewed bird's nest.

[0021] The predicted freshness preservation potential energy obtained from the forward delivery temperature sequence is used to perform dual consistency calculation with the actual freshness preservation potential energy obtained from the residual temperature sequence of the recovery temperature control box. When the trigger condition is met, the activation energy correction factor of the target batch is corrected. The freshness recursion in subsequent cycle delivery can absorb the recovery temperature feedback and reduce the accumulation of deviation between the predicted freshness and the actual temperature experience. Attached Figure Description

[0022] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating some specific embodiments of the method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest according to the present invention. Figure 2 This is a schematic diagram of the structure of a real-time adjustment system for the periodic delivery route of freshly stewed bird's nest according to some specific embodiments of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms "a," "the," and "the" as used in the embodiments of this application are also intended to include the plural forms, unless the context clearly indicates otherwise, and "multiple" generally includes at least two.

[0025] It should be understood that the term "and / or" used in this article 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 existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0026] It should be understood that although the terms first, second, third, etc., may be used in the embodiments of this application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, first may also be referred to as second without departing from the scope of the embodiments of this application, and similarly, second may also be referred to as first.

[0027] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0028] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0029] It should be noted that any symbols and / or numbers present in the specification that are not marked in the accompanying drawings are not reference numerals.

[0030] Reference Figure 1 A method for real-time adjustment of the delivery route for freshly stewed bird's nest, comprising: S101 collects periodic subscription orders, product batches, customer coordinates and delivery time windows, vehicle locations, temperature sequence of temperature control boxes, door opening and closing records, and residual temperature sequence of recycled temperature control boxes; S102, input the initial freshness index, pre-exponential factor, activation energy, real-time temperature and sampling interval of the product batch into the Arrhenius recursive relation, and form the current freshness index, remaining shelf life and shelf hard constraint boundary for each package; S103, calculate the predicted delivery freshness index of unfinished orders along the current path timeline, and convert the predicted delivery freshness index and the minimum acceptable freshness threshold into preservation potential energy. S104, Perform counterfactual reasoning on the order node status offset, and generate a preservation vulnerability heatmap based on the statistical quantity of delivery freshness change caused by the status offset; S105, when a new order is inserted, the address is changed, or the temperature of the temperature control box deviates from the preset temperature range, the disturbance sub-map is cut by the heat map of the disturbance location and the freshness vulnerability, and the temperature sampling or road condition confirmation action is triggered on the vulnerable nodes in the disturbance sub-map. S106 inputs the perturbation subgraph, preservation potential energy, time window, travel time, traffic status and temperature rise coefficient into the graph attention policy network with physical modulation, modulates the attention weights between nodes according to the difference in preservation potential energy caused by the service order, and adds the non-negative part of the change in preservation potential energy before and after the action to the policy loss, and outputs the adjusted delivery route and information collection instructions.

[0031] To provide a clearer explanation, the steps in the embodiments of the present invention are described in detail below: S101 collects periodic subscription orders, product batches, customer coordinates and delivery time windows, vehicle locations, temperature sequence of temperature control boxes, door opening and closing records, and residual temperature sequence of recycled temperature control boxes; Specifically, in this embodiment of the invention, periodic subscription orders, product batches, customer coordinates, and delivery time windows are collected to form order data; vehicle locations, temperature sequences of temperature control boxes, and door opening and closing records are collected to form in-transit environmental data; residual temperature sequences of recycled temperature control boxes are collected, and a forward delivery and reverse recycling node relationship is established between customer nodes with recycling needs and periodic subscription orders.

[0032] In this embodiment, order data can come from an enterprise resource planning (ERP) system and a transportation management system. A periodic subscription order includes at least an order identifier, customer identifier, delivery date, delivery time window, ordered product type, product batch, delivery address, customer coordinates, a pause delivery flag, a relocation flag, and a new order flag. A product batch includes at least the manufacturing date, initial freshness index, pre-exposure factor, activation energy, and minimum acceptable freshness threshold. Customer coordinates can be latitude and longitude coordinates or a unified map coordinate system within the delivery system; when using map coordinates, the latitude and longitude coordinates are converted to the same planar coordinate system before entering subsequent path processing to avoid vehicle location, customer coordinates, and path segment distances being on different coordinate bases.

[0033] In this embodiment, the environmental data during transit can come from an IoT platform, vehicle positioning terminal, passenger compartment temperature sensor, temperature control box temperature sensor, and door status sensor. The temperature control box temperature sequence is recorded according to package identification, and the recorded items may include package identification, temperature control box identification, sampling time, sampling temperature, sampling interval, wireless communication status, and data reception time. Door opening / closing records may include vehicle identification, door opening time, door closing time, door opening duration, and door opening position. Vehicle location may include vehicle identification, positioning time, latitude and longitude coordinates, speed, driving direction, and current path segment identification. If the temperature sequence has missing sampling points, the time difference between adjacent sampling points is first used to determine if it exceeds a preset missing condition; if it does not exceed the preset missing condition, linear interpolation or maintaining the previous sampling value is used to generate the missing temperature; if it exceeds the preset missing condition, the package is marked as having an uncertain temperature status.

[0034] In this embodiment, the residual temperature sequence of the recycled temperature control box is obtained when the customer signs for and returns the empty bottle or insulated box. Record items may include the recycled temperature control box identifier, associated order identifier, signing time, recycling start time, residual temperature sequence, and recycling vehicle identifier. A forward delivery and reverse recycling node relationship is established based on the customer node, order identifier, and recycled temperature control box identifier. This relationship provides a data source for duality consistency loss and target batch activation energy correction.

[0035] In this embodiment, order data and in-transit environment data are entered into a unified message queue or streaming data cache, and data alignment is performed according to order identifier, package identifier, vehicle identifier, and timestamp. The data alignment result can be stored as an order status record. The order status record includes at least the order node, package node, path segment, product batch, customer coordinates, delivery time window, current vehicle location, temperature sequence of the temperature control box, residual temperature sequence of the recycling temperature control box, and data availability marker. This order status record serves as the common input for freshness recursion, preservation potential calculation, counterfactual reasoning, and policy network reasoning.

[0036] In this embodiment, a periodic order baseline graph can also be established during the data acquisition phase. The nodes of the periodic order baseline graph are regular customer nodes and temporary order nodes, and the edges are historical path segments, candidate travel path segments, or expected travel path segments. For long-term subscribers, the commonly used delivery order, average service time, and travel time of each path segment are statistically analyzed based on historical fulfillment records. For new trial orders, the new order node is temporarily attached to a spatially adjacent and time-compatible candidate path segment. The periodic order baseline graph serves as the initial graph structure for path adjustments in subsequent steps, avoiding the need to rebuild the graph from a completely blank state for each disturbance.

[0037] S102, input the initial freshness index, pre-exponential factor, activation energy, real-time temperature and sampling interval of the product batch into the Arrhenius recursive relation, and form the current freshness index, remaining shelf life and shelf hard constraint boundary for each package; S102 further includes: Read the package temperature within the sampling interval according to the temperature sequence of the temperature control box; The packaging temperature, pre-finger factor, and activation energy are input into the temperature decay term, and the retention result of the freshness index at the previous moment after the effect of the temperature decay term is used as the freshness index at the next moment. The remaining shelf life is calculated by accumulating the freshness index along the unfinished delivery route at the next moment. A critical package marker is generated when the remaining shelf life is lower than the preset warning condition, and a normal package marker is retained when the remaining shelf life is not lower than the preset warning condition.

[0038] Specifically, in this embodiment of the invention, for any package At any moment The freshness index after sampling intervals can be represented by the following Arrhenius recursive relation: ; in, For the package At any moment The freshness index, with an initial value given by the product batch record; For the package The freshness index at the next sampling time; This is the pre-exponential factor, calibrated by accelerated shelf-life testing; The activation energy is determined by accelerated shelf-life testing. The molar gas constant; For the package At any moment The Kelvin temperature is obtained from the temperature sequence of the temperature control chamber; The sampling interval is denoted as .

[0039] The output of the above recursive relationship is the freshness index at the next time step, which serves as the input for calculating the remaining shelf life and the predicted delivery freshness index.

[0040] In this embodiment, when reading the package temperature, the temperature sensor data of the temperature control box bound to the package identifier is read first. When there are multiple packages in the same temperature control box, the temperature of the temperature control box can be used as the common temperature input for all packages, or the package temperature can be corrected according to the layer position of the package in the temperature control box, its proximity to the door, and the door opening and closing records. For example, when the door is open for more than 30 seconds, the temperature of the package near the door during the opening period is increased by a door opening disturbance correction amount; when the sensor sampling frequency is increased in active acquisition mode, the sampling interval is... The freshness regression is shortened, and the process is executed with a shorter time step. This process allows temperature sampling intervals and active data acquisition actions to be incorporated into the freshness regression process.

[0041] In this embodiment, the temperature decay term is Read the product batch and Read the current sampling point corresponding to and The freshness index of the previous time step, after being affected by the temperature decay term, is used as the freshness index of the next time step. When the temperature decay term leads to... When it is below 0, The temperature is truncated to 0, and an undeliverable flag is output. When the temperature data is missing and cannot be filled in, the nearest available temperature in the same path segment, the same temperature control box, or the same vehicle compartment is used as the temporary temperature input. At the same time, a temperature status uncertainty flag is output, and this flag is added to the candidate queue of data acquisition actions.

[0042] In this embodiment, the remaining shelf life is calculated by accumulating the freshness index along the unfinished delivery route to the next moment. Specifically, starting from the vehicle's current location, the estimated travel time, estimated waiting time, service time, and door opening / closing impact time are accumulated segment by segment according to the service order of unfinished orders in the current route plan. The freshness index is then recursively calculated based on the estimated temperature for each time segment. When the recursively calculated freshness index first falls below the minimum acceptable freshness threshold, the remaining time between the current location and that point in time is recorded as the remaining shelf life. If the remaining shelf life is not lower than the minimum acceptable freshness threshold until the estimated delivery time, then the remaining shelf life at least covers the estimated delivery time under the current route plan.

[0043] In this embodiment, either a critical package marker or a regular package marker is generated. The preset warning condition can be determined by the condition that the remaining shelf life is less than the sum of the estimated time required to reach the target customer and the safety buffer time. For example, the safety buffer time can be any value between 15 minutes and 60 minutes, and the specific value can be jointly determined by the product's cold chain standards, the degree of urban delivery congestion, and the company's fulfillment strategy. If the remaining shelf life of the package is lower than the preset warning condition, a critical package marker is generated; if the remaining shelf life of the package is not lower than the preset warning condition, a regular package marker is retained.

[0044] In this embodiment, batch parameter boundary processing can also be set for the Arrhenius recursive relationship. If the pre-exponential factor or activation energy of a certain product batch is not yet read, the default parameters of the same product type, the same process batch, or the same shelf life grade are read, and the package is marked as a batch parameter pending correction state. When the loss of duality triggers an online Bayesian update, the activation energy correction factor of the target batch is corrected using the residual temperature sequence of the recovery temperature control box. This processing enables freshness recursion to be executed accurately when batch parameters are complete, and to maintain the path adjustment process through traceable alternative parameters when parameters are temporarily missing.

[0045] S103, calculate the predicted delivery freshness index of unfinished orders along the current path timeline, and convert the predicted delivery freshness index and the minimum acceptable freshness threshold into preservation potential energy. Specifically, in this embodiment of the invention, the predicted delivery freshness index of unfinished orders is calculated along the current path timeline; the predicted delivery freshness index and the minimum acceptable freshness threshold are converted into preservation potential energy; and the preservation potential energy is associated with critical package markers, ordinary package markers and the current path plan to form the physical state input of unfinished order nodes.

[0046] In this embodiment, the current freshness index and remaining shelf life are read, along with the current path timeline. The current path timeline may include the vehicle's current location, the service order of incomplete orders, the estimated travel time for each path segment, the estimated arrival time for each order, the estimated waiting time, the estimated delivery service time, and the door opening / closing time. Starting from the current moment, the freshness index of the package bound to each incomplete order is recursively calculated along the path timeline to obtain the predicted delivery freshness index. If an order contains multiple packages, the minimum value among the predicted delivery freshness indices of the multiple packages can be used as the predicted delivery freshness index for the order. Alternatively, the predicted delivery freshness indices of the multiple packages can be weighted according to the product risk level to obtain the predicted delivery freshness index for the order.

[0047] In this embodiment, the following relationship is used to calculate the order. Preservation potential: ; in, For orders The preservation potential energy; For orders The predicted delivery freshness index under the current route plan is calculated along the route timeline. This is the minimum acceptable freshness threshold.

[0048] like Greater than ,but Greater than 1, A value less than 0 indicates that the order is in a safe and fresh condition under the current route plan; like equal ,but A value of 0 indicates that the order is at the critical point of preservation. like Less than ,but A value greater than 0 indicates that the order may not meet the freshness standard under the current route plan.

[0049] This preservation potential is written into the order node features and used as input for subsequent steps.

[0050] In this embodiment, the preservation potential is associated with critical package markers, ordinary package markers, and the current route plan. The association method can be as follows: The order node record is written with the order identifier, package identifier, current service order, predicted delivery time, predicted delivery freshness index, minimum acceptable freshness threshold, preservation potential, package marker, remaining time window, and route segment identifier. For order nodes with a preservation potential greater than 0, the node is added to the set of nodes to be adjusted; for order nodes with a preservation potential close to 0 and remaining shelf life lower than a preset warning condition, the node is added to the set of vulnerable node candidates; for order nodes with a preservation potential less than 0 and sufficient remaining shelf life, the ordinary route candidate status of the node is retained.

[0051] In this embodiment, if the minimum acceptable freshness threshold is not directly given in the current product batch, the enterprise's internal control threshold for the same product type can be read, or the minimum acceptable freshness ratio in the product shelf-life rules can be read. For example, in a complete and specific example, the minimum acceptable freshness threshold can be 80, and the initial freshness index can be 100. The above values ​​are only used to illustrate the calculation process and do not represent actual product parameters or optimal parameters. If the predicted delivery freshness index is missing, the order node is marked as having an uncertain prediction status, and the node is input into the active information collection action selection process.

[0052] In this embodiment, the potential energy can be set without adding constraints. For any candidate path action... Calculate the difference between the preservation potential energy after the action and the preservation potential energy before the action. If at least one incomplete order node A value greater than 0 indicates that the action of the candidate path increases the risk of the corresponding node remaining unfilled; this difference is entered into the physical conservation regularization term. If all unfinished order nodes... If all values ​​are less than or equal to 0, it indicates that the candidate path action satisfies the physical constraints at the preservation potential energy level. In this way, the preservation potential energy serves not only as a state variable but also as a constraint basis for path adjustment actions.

[0053] S104, Perform counterfactual reasoning on the order node status offset, and generate a preservation vulnerability heatmap based on the statistical quantity of delivery freshness change caused by the status offset; S104 further includes: Insertion location, insertion order parameters, adjacent order service order, adjacent order time window, adjacent package temperature box superimposed heat effect, and road segment traffic status are written into the variable nodes of the structural causal model; Estimate causal edges between variable nodes using historical delivery samples; Apply state parameter offset intervention to nodes with incomplete orders, compare the freshness index of neighboring orders before and after the intervention, and write the maximum or average value of the difference in freshness index into the preservation vulnerability heat map.

[0054] Use the location of the new order insertion, the location of the address change, or the location of the package with abnormal temperature as the disturbance seed; Further: Read order nodes affected by disturbed seeds from the heat map of preservation vulnerability; Expand the neighborhood of the affected order node according to the constraints of historical path segments, estimated travel time and delivery time window; Neighborhood edges that do not meet the delivery time window constraint or the hard freshness constraint are removed, and perturbation subgraphs are generated for the remaining order nodes and path edges.

[0055] Specifically, in this embodiment of the invention, the variable nodes of the structural causal model may include insertion position, insertion order parameters, adjacent order service order, adjacent order time window, superimposed thermal effect of adjacent package temperature control boxes, road traffic conditions, vehicle current location, candidate route segment travel time, door opening / closing status, and current preservation potential. Insertion order parameters may include new order customer coordinates, product batch, delivery time window, estimated service duration, and initial freshness index of the package. The superimposed thermal effect of adjacent package temperature control boxes can be formed based on the number of packages in the same compartment, temperature control box door opening / closing records, and temperature differences at multiple points in the compartment, used to describe the differences in temperature distribution among different packages in the same compartment. The output variable is the delivery freshness index or the change in delivery freshness index of adjacent orders.

[0056] In this embodiment, causal edges between variable nodes are estimated using historical delivery samples. Historical delivery samples may include the actual service sequence of completed orders, actual travel time, actual delivery time, temperature sequence, freshness index at receipt, door opening / closing records, road traffic conditions, dynamic event types, and final route adjustment actions. Causal edges are prohibited based on business constraints, such as future delivery freshness not being allowed to negatively impact historical traffic conditions. Causal edges between variables are estimated through structural learning or constrained regression modeling. Model training can be completed offline, with the training objective being to enable the structural causal model to estimate changes in the delivery freshness index given intervention variables. After training, the structural causal model parameters are loaded into an online inference service.

[0057] In this embodiment, a state parameter offset intervention is applied to nodes with incomplete orders. State parameter offsets may include changes in insertion position, earlier or later service order, increased estimated travel time, tighter time windows, increased temperature, increased door opening / closing time, and changes in traffic congestion. At least one state parameter offset is applied to each node with incomplete orders, and the difference in freshness index between neighboring orders before and after the intervention is calculated. If the state parameter offset of a node affects multiple neighboring orders, the maximum or average of the freshness index differences can be taken as the vulnerability score of that node. The vulnerability score is written into a freshness vulnerability heatmap; high-value nodes in the heatmap indicate that changes in the state of that node are more likely to cause the spread of freshness risk to surrounding orders.

[0058] In this embodiment, a perturbation subgraph is generated when actual dynamic events occur. Dynamic events can include new order insertion, address change, temperature deviation of the temperature control box from the preset temperature range, vehicle location deviation, road congestion, or customer suspension of delivery. The location of the event is used as a perturbation seed, and the order nodes affected by the perturbation seed are read from the freshness vulnerability heatmap. During neighborhood expansion, the expansion proceeds outward along historical path segments and current candidate path segments, while simultaneously checking the estimated travel time and delivery time window constraints. If a neighborhood edge causes the estimated arrival time of the order to exceed the delivery time window, or causes the predicted freshness index of the bound package to fall below the minimum acceptable freshness threshold, the neighborhood edge is removed. The retained order nodes and path edges constitute the perturbation subgraph. The number of nodes in this perturbation subgraph is usually less than the total number of order nodes in the entire network, so the subsequent policy network can perform real-time path adjustments on a smaller graph structure.

[0059] In this embodiment, an update cycle and triggered updates can also be set for the preservation vulnerability heatmap. For example, batch counterfactual reasoning can be performed on all incomplete orders across the network every 5 minutes, and local counterfactual reasoning can be performed immediately on the neighborhood of the perturbation seed when a new order is inserted, the address changes, or the temperature of the temperature-controlled box deviates from the preset temperature range. If the structural causal model cannot output the vulnerability score of a certain node temporarily, the preservation potential energy, remaining time window, and density of neighboring orders of that node can be read to generate a temporary vulnerability score, which is then marked as a low-confidence state. Nodes in the low-confidence state are given priority in the information collection action selection process.

[0060] S105, when a new order is inserted, the address is changed, or the temperature of the temperature control box deviates from the preset temperature range, the disturbance sub-map is cut by the heat map of the disturbance location and the freshness vulnerability, and the temperature sampling or road condition confirmation action is triggered on the vulnerable nodes in the disturbance sub-map. S105 further includes: For vulnerable nodes, simulate actions such as shortening the temperature sampling interval, confirming driver road conditions, and maintaining the sampling frequency. The structural causal model is used to estimate the uncertainty change of preservation potential energy before and after the execution of each type of action; The amount of uncertainty reduction is calculated by subtracting the sensor power consumption, communication traffic, and driver confirmation delay. The action with the largest difference is written into the features of the perturbation subgraph node, and the collected return data is input into the graph attention policy network with physical modulation.

[0061] Furthermore, the action with the largest difference is written into the perturbation subgraph node features, including: Establish a candidate queue for data collection actions; The uncertain range of preservation potential energy of vulnerable nodes, the remaining power of the temperature control box, the communication delay, and the driver confirmation delay are written into the candidate action attributes respectively; Candidate actions that do not meet the power or delay conditions will be eliminated. Candidate actions are sorted according to the difference between the reduction in uncertainty and the acquisition cost, and the action with the highest ranking is used to generate an information acquisition instruction.

[0062] Specifically, in this embodiment of the invention, vulnerable nodes are derived from the generated perturbation subgraph and the preservation vulnerability heatmap. A candidate set of acquisition actions is established for each vulnerable node. The candidate set of acquisition actions may include actions such as shortening the temperature sampling interval, confirming driver traffic conditions, and maintaining the acquisition frequency. The action of shortening the temperature sampling interval refers to switching the temperature sensor of the temperature control box from the default sampling interval to a shorter sampling interval, for example, from 5 minutes to 1 minute; the action of confirming driver traffic conditions refers to sending a traffic condition confirmation request for a specified route segment to the driver's terminal; the action of maintaining the acquisition frequency refers to maintaining the current sampling method without increasing the acquisition cost. Each candidate action is bound to the vulnerable node, the applicable package, the target vehicle, the expected execution time, and the execution result write-back location.

[0063] In this embodiment, a structural causal model is used to estimate the change in preservation potential energy uncertainty before and after each type of action. Specifically, using the current perturbation subgraph features as the pre-collection state, the preservation potential energy distribution or preservation potential energy interval of vulnerable nodes is estimated; the data improvement effects generated by candidate collection actions are written into the structural causal model, such as shortening the temperature sampling interval to reduce temperature estimation error, and driver road condition confirmation to reduce uncertainty in path segment travel time; the preservation potential energy distribution or preservation potential energy interval after collection is estimated. The preservation potential energy uncertainty before and after collection can be represented by interval width, variance, or quantile difference. The amount of uncertainty reduction is the uncertainty before collection minus the uncertainty after collection, and this reduction is taken as the collection gain.

[0064] In this embodiment, the uncertainty reduction is calculated by subtracting the difference between sensor power consumption, communication traffic, and driver confirmation delay. For example, the data acquisition action... The action score can be expressed as: ; in, For the collection action Action scoring; Uncertainty regarding the preservation potential before collection; The uncertainty of the preservation potential after collection; For the collection action The corresponding sensor power consumption cost; For the collection action The corresponding communication traffic cost; For the collection action The corresponding driver confirmation delay cost.

[0065] Each cost item can be normalized to the same dimension, and the normalization weights can be provided as exemplary values ​​in the complete and specific examples. If a data acquisition action does not meet the power consumption condition, communication condition, or latency condition, the candidate action will be eliminated.

[0066] In this embodiment, the action with the largest difference is written into the perturbation subgraph node features, and the collected return data is input into a graph attention policy network with physical modulation. When an action is written into the perturbation subgraph node features, the action type, target package, target vehicle, estimated start time, estimated end time, action score, and action execution status are recorded. The collected return data can be a new temperature sampling point, the driver's confirmed remaining path congestion status, temporary road closure status, or an estimated arrival time correction value. The collected return data updates the temperature sequence and preservation potential on the one hand, and updates the perturbation subgraph node features on the other, allowing the policy network to use the collected state when adjusting the path.

[0067] In this embodiment, the candidate queue for data collection actions can be indexed according to node identifiers and action types. The uncertain range of preservation potential energy is output from the above steps, the remaining power of the temperature control box is read by the IoT platform, the communication delay is estimated from the most recent communication record, and the driver confirmation delay is estimated from the driver terminal's historical response record or current online status. If all candidate actions are eliminated, a data collection unavailable marker is generated and written into the node features of the perturbation subgraph, making the path solving process more inclined to select path actions that do not depend on newly collected data.

[0068] In this embodiment, detectable records can also be written into the information collection instructions. These detectable records include the candidate queue identifier for the collection action, the identifier of vulnerable nodes, the uncertain range of preservation potential energy, the remaining power of the temperature control box, communication latency, driver confirmation delay, action score, reason for removal, sorting result, and the final issued action. These detectable records can be stored in the route adjustment log for subsequent operational verification and quality traceability.

[0069] S106 inputs the perturbation subgraph, preservation potential energy, time window, travel time, traffic status and temperature rise coefficient into the graph attention policy network with physical modulation, modulates the attention weights between nodes according to the difference in preservation potential energy caused by the service order, and adds the non-negative part of the change in preservation potential energy before and after the action to the policy loss, and outputs the adjusted delivery route and information collection instructions.

[0070] S106 further includes: Calculate the preservation potential energy of the second order node when the first order node serves before the second order node, and use it as the first preservation potential energy; Calculate the preservation potential energy of the second order node when the second order node serves before the first order node, and use it as the second preservation potential energy; The difference between the first preservation potential energy and the second preservation potential energy is input into the exponential modulation term; The attention coefficients of the standard graph are modulated by an exponential modulation term and then normalized to generate freshness causal attention weights; Input the causal attention weight of preservation into the service sequence action selection.

[0071] Furthermore, after outputting the adjusted delivery route and information collection instructions, it includes: Substituting the forward delivery temperature sequence into the preservation potential energy calculation yields the predicted delivery preservation potential energy, and substituting the residual temperature sequence of the recovery temperature control box into the Arrhenius recursive relation yields the actual preservation potential energy. Calculate the duality loss between the predicted delivery preservation potential energy and the actual preservation potential energy; When the duality loss exceeds the preset trigger threshold, the target batch activation energy correction factor is updated and corrected online using Bayesian method and then input into the next round of freshness index calculation. The target batch activation energy correction factor is retained when the duality consistency loss does not exceed the preset trigger threshold. When deploying across scenarios, the physical conservation regularization term is added to the gradient update of the meta-learning inner layer to form the initial parameters of the policy network, and the initial parameters of the policy network are updated with local periodic delivery data.

[0072] Furthermore, the target batch activation energy correction factor is updated and corrected online using Bayesian methods, including: Extract the temperature integral and receipt timestamp after receiving from the residual temperature sequence of the recycled temperature control box; The temperature integral after receiving the goods is input into the Arrhenius recursive relation to deduce the actual freshness index. The posterior distribution of the activation energy correction factor for the target batch is updated by the difference between the predicted freshness index and the actual freshness index. When the posterior distribution variance meets the preset convergence condition, the posterior mean is input into the next round of package freshness recursion. When the posterior distribution variance does not meet the preset convergence condition, the activation energy correction factor of the target batch in the previous round is retained and the recovery temperature sequence continues to be accumulated.

[0073] Specifically, in this embodiment of the invention, the obtained perturbation subgraph is converted into a graph structure that can be processed by the policy network. Nodes in the graph structure are nodes with incomplete orders, critical package nodes, or customer nodes with recycling needs; edges represent candidate service sequences, candidate path segments, or vehicle reachability relationships. Node features may include freshness potential, start time of the delivery time window, end time of the delivery time window, predicted delivery time, estimated travel time, vulnerability tag, critical package marker, collection action marker, and recycling node marker. Edge features may include real-time traffic conditions, temperature rise coefficient, historical average travel time, expected door opening / closing impact, and path segment availability marker. The temperature rise coefficient can be obtained from historical temperature sequences and path segment travel time statistics, used to estimate the recursive impact of package temperature rise per unit travel time on freshness.

[0074] In this embodiment, a causal attention mechanism for food preservation is constructed. For the first order node... Second order node Calculate the first order node Prior to the second order node Second order node during service The preservation potential energy is used as the first preservation potential energy; the second order node is calculated. Prior to the first order node Second order node during service The preservation potential energy is taken as the second preservation potential energy; the difference between the first and second preservation potential energies is denoted as... Standard graph attention coefficient Can be generated by nodes and nodes The feature vector is calculated after a linear transformation. The preservation causal attention weights after introducing physical modulation can be expressed as: ; in, For nodes To the node The causal attention weight of preservation; The standard graph attention coefficient; This is the difference between the first preservation potential energy and the second preservation potential energy; For temperature parameters, a value of 0.3 to 1.0 is preferred.

[0075] If a certain service sequence leads to the second order node The preservation potential increases, then If the value is positive, the exponential modulation term decreases, and the attention weight corresponding to that service order is reduced; if a service order does not increase the preservation potential, then that service order is retained in the attention calculation. The preservation causal attention weight is incorporated into the service order action selection.

[0076] In this embodiment, the non-negative portion of the change in preservation potential energy before and after the action is added to the strategy loss. Candidate actions may include exchanging service order, inserting a new order node between specified locations, initiating backup vehicle transfer, shortening the temperature sampling interval, and requesting the driver to confirm road conditions. For candidate actions... and the set of nodes in the perturbation subgraph The physical conservation regularity term can be expressed as: ; in, This is a physical conservation regularity term; For the set of nodes in the perturbation subgraph; For action Resulting nodes The change in preservation potential energy is defined as the preservation potential energy after the action minus the preservation potential energy before the action.

[0077] When an action increases the preservation potential energy of a node, that node incurs a positive penalty; when an action keeps the preservation potential energy unchanged or decreases it, the node's contribution to the physical conservation regularization term is zero. The total loss of the policy network can be expressed as: ; in, The total loss of the policy network; The policy gradient loss; The weight of the physical regularization term is preferably between 0.5 and 2.0; The weight of the entropy regularization term is preferably between 0.005 and 0.02; Let be the policy entropy.

[0078] During training, the strategy network parameters are updated using historical order flows and simulated perturbation samples. During inference, a single forward propagation is performed on the perturbation subgraph, outputting the adjusted delivery route and information collection instructions. The output route includes service order, vehicle allocation, estimated arrival time, estimated delivery freshness index, and the information collection actions to be performed.

[0079] In this embodiment, a dual verification of forward delivery and reverse recycling is implemented. The predicted delivery preservation potential energy is obtained by substituting the forward delivery temperature sequence into the equation. Substituting the residual temperature sequence of the recovery temperature control box into the Arrhenius recursive relation yields the actual preservation potential energy. The duality consistency loss can be expressed as: ; in, This represents the loss of duality consistency. This is a collection of orders with recycling records. The number of orders with a recycling record; This represents the delivery and preservation potential predicted by the positive delivery network. This refers to the actual preservation potential energy derived from the temperature data of the temperature control box.

[0080] when When the preset trigger threshold is exceeded, the activation energy correction factor of the target batch is updated and corrected online using Bayesian methods; when If the preset trigger threshold is not exceeded, the activation energy correction factor for the target batch is retained. The preset trigger threshold can preferably be any value between 0.1 and 0.3, and can be configured according to the product type and the stability of the recycling data.

[0081] In this embodiment, the temperature integral after receipt can be obtained by summing the residual temperature sequence of the recycled temperature control box according to the sampling interval. The receipt timestamp is used to determine the end point of the forward delivery and the start point of the recycled temperature sequence. The prior of the posterior distribution can be given by the factory inspection data or historical recycled data of the same batch. If the recycled temperature sequence is missing or the recycled temperature control box identifier cannot match the order, the recycled data is marked as unusable.

[0082] In this embodiment, meta-learning cold start is implemented using physical prior embedding. Sampling training tasks are configured according to city size, season type, cyclical order density, and cold chain equipment. For each task... The inner gradient update can be represented as: ; Meta-update can be represented as: ; in, These are the current model parameters; In the mission The model parameters after the inner layer update; The inner learning rate; For parameters The gradient operator; For the task Policy gradient loss; The weight of the physical regularization term; This is a physical conservation regularity term; These are the updated model parameters; The outermost learning rate; To sum over a batch of sampled tasks; To use the updated parameters in the inner layer in the task The policy gradient loss is calculated above.

[0083] When deployed to a new city or new season, load The initial parameters of the strategy network are used as inputs and updated with local periodic delivery data. Local data may include local orders, vehicle routes, temperature sequences, delivery records, and residual temperature sequences of recycled temperature-controlled boxes. This meta-learning process serves as the implementation support for the strategy network.

[0084] In this embodiment, the training, inference, and evaluation processes of the policy network are as follows. Training data comes from historical order flows, historical vehicle locations, historical temperature sequences of temperature-controlled boxes, historical dynamic events, historical route adjustment actions, delivery records, and residual temperature sequences of recycled temperature-controlled boxes. The training input is a perturbation subgraph and its node and edge features; the training output is path actions, information collection actions, and service sequence actions. Training objectives include policy gradient loss, physical conservation regularization, entropy regularization, and duality consistency loss. The training stopping condition can be that the path action loss on the validation set changes less than a preset training convergence threshold within five consecutive training rounds, or that a preset maximum number of training rounds is reached. The inference process starts from the output perturbation subgraph, sequentially calculating the preservation causal attention weight, candidate action score, physical conservation regularization, and action probability, ultimately outputting the adjusted delivery path and information collection instructions. The evaluation process can use historical replay orders and simulated perturbation samples. The evaluation results are used to select whether to save the parameter version, whether to trigger the activation energy correction factor update, and whether to revert to the previous parameter version; the evaluation process does not output true performance conclusions.

[0085] In this embodiment, an operation log can also be output. The operation log includes at least the disturbance event identifier, the set of nodes in the disturbance subgraph, the node preservation potential energy, the node vulnerability score, the candidate queue of collection actions, the summary of preservation causal attention weights, the candidate path actions, the physical conservation regularization term, the duality consistency loss, the target batch activation energy correction factor version, the final delivery route, and the information collection instructions. The operation log can be written to a database or object storage and read according to the order identifier, vehicle identifier, and timestamp index. This operation log is used for implementation detection, result verification, and operation status verification, without changing the main path adjustment process.

[0086] The following describes this embodiment in conjunction with application scenarios: For example, a vehicle currently carries three orders for freshly stewed bird's nest. Orders A, B, and C have customer coordinates, delivery time windows, and product batch parameters, respectively. The initial freshness index of package a bound to order A is collected as 100, the minimum acceptable freshness threshold is 80, the most recently sampled temperature of the temperature control chamber is 278K, the sampling interval is 0.083h, and the pre-exponential factor and activation energy are from the batch record. Substituting the above data into the Arrhenius recursive relation, the freshness index at the next sampling time is obtained, and this is recursively applied along the current path timeline to the expected delivery time of order A, forming the predicted delivery freshness index of order A. If the exemplary predicted delivery freshness index is 84, then the preservation potential energy is calculated. If this value is less than 0, order A is in a safe state; if the predicted delivery freshness index of order B is 79, then... If this value is greater than 0, order B will be added to the set of nodes to be adjusted.

[0087] In this example, the insertion position of new order D is detected to be between orders B and C. This insertion position is used as a perturbation seed, and the preservation vulnerability heatmap is read. If orders B and C have high vulnerability scores in the heatmap, the neighborhood is expanded along historical path segments and delivery time window constraints. Path edges that cause order B to exceed the delivery time window are removed, generating a perturbation subgraph containing orders B, C, and new order D. Temperature sampling interval shortening, driver road condition confirmation, and sampling frequency maintenance actions are simulated for order B. If the temperature sampling interval shortening action reduces the uncertainty interval of preservation potential energy from 0.12 to 0.05, the normalized power consumption cost is 0.02, the communication cost is 0.01, and the latency cost is 0, then the action score is 0.04. If the driver road condition confirmation action reduces the uncertainty interval from 0.12 to 0.08, and the normalized latency cost is 0.02, then the action score is 0.02. The temperature sampling interval shortening action is selected, and this action is written into the perturbation subgraph node features.

[0088] In this example, the perturbation subgraph is input into a graph attention policy network with physical modulation. For orders B and C, the preservation potential energy of order C is calculated when order B is served first and then order C, as the first preservation potential energy; and the preservation potential energy of order C is calculated when order C is served first and then order B, as the second preservation potential energy; the difference between the first preservation potential energy and the second preservation potential energy is denoted as... .like It is 0.2. If we take 0.5, then the exponential modulation term is: The attention coefficients of the standard graph are modulated by an exponential modulation term and then normalized to generate preservation causal attention weights. Subsequently, the preservation potential energy after each action is calculated for candidate actions such as service order exchange, inserting a new order D, and initiating backup vehicle transfer, and then accumulated. A physical conservation regularization term is formed. If a candidate action causes an increase in the preservation potential energy of order B, this candidate action is penalized in the strategy loss; if another candidate action prevents an increase in the preservation potential energy of both order B and order C, this candidate action is taken as the priority path action, and the adjusted delivery route and information collection instructions are output. After receipt, if there is a residual temperature sequence in the recycling temperature control box of order B, this temperature sequence is substituted into the Arrhenius recursion to deduce the actual preservation potential energy, and the dual consistency loss is calculated with the predicted delivery preservation potential energy. If the dual consistency loss exceeds 0.2, the target batch activation energy correction factor is updated; if it does not exceed 0.2, the current correction factor is retained.

[0089] For the purpose of simplicity, the method steps disclosed in the above embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0090] like Figure 2 As shown, the present invention also provides a real-time adjustment system for the periodic delivery route of freshly stewed bird's nest, comprising: The data aggregation module 201 is configured to collect periodic subscription orders, product batches, customer coordinates and receiving time windows, vehicle locations, temperature sequences of temperature control boxes, door opening and closing records, and residual temperature sequences of recycled temperature control boxes. The freshness recursion module 202 is configured to input the initial freshness index, pre-exponential factor, activation energy, real-time temperature and sampling interval of the product batch into the Arrhenius recursion relation, and form the current freshness index, remaining shelf life and shelf hard constraint boundary for each package. The preservation potential energy module 203 is configured to calculate the predicted delivery freshness index of unfinished orders along the current path timeline, and convert the predicted delivery freshness index and the minimum acceptable freshness threshold into preservation potential energy. Causal reasoning module 204 is configured to perform counterfactual reasoning on the order node state offset and generate a preservation vulnerability heat map based on the statistical quantity of delivery freshness change caused by the state offset. The disturbance response module 205 is configured to cut a disturbance sub-map based on the disturbance location and the heat map of preservation vulnerability when a new order is inserted, the address is changed, or the temperature of the temperature control box deviates from the preset temperature range, and to trigger temperature sampling or road condition confirmation actions on the vulnerable nodes in the disturbance sub-map. The path solving module 206 is configured to input the perturbation subgraph, preservation potential energy, time window, travel time, traffic status and temperature rise coefficient into a graph attention policy network with physical modulation, modulate the attention weights between nodes according to the difference in preservation potential energy caused by the service order, and add the non-negative part of the change in preservation potential energy before and after the action to the policy loss, and output the adjusted delivery path and information collection instructions.

[0091] It is worth noting that although only some basic functional modules are disclosed in the embodiments of this invention, it does not mean that the composition of this system is limited to the above-mentioned basic functional modules. On the contrary, based on the above-mentioned basic functional modules, those skilled in the art can arbitrarily add one or more functional modules in combination with existing technology to form an infinite number of embodiments or technical solutions. That is to say, this system is open rather than closed. The fact that this embodiment only discloses a few basic functional modules should not be considered as the scope of protection of this invention being limited to the disclosed basic functional modules. At the same time, for the convenience of description, the above devices are described separately according to their functions as various units and modules. Of course, in implementing this invention, the functions of each unit and module can be implemented in one or more software and / or hardware.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest, characterized in that, include: Collect periodic subscription orders, product batches, customer coordinates and delivery time windows, vehicle locations, temperature sequences of temperature control boxes, door opening and closing records, and residual temperature sequences of recycled temperature control boxes; The initial freshness index, pre-exponential factor, activation energy, real-time temperature and sampling interval of the product batch are input into the Arrhenius recursive relation to form the current freshness index, remaining shelf life and hard constraint boundary of the shelf life for each package. Calculate the predicted delivery freshness index of uncompleted orders along the current path timeline, and convert the predicted delivery freshness index and the minimum acceptable freshness threshold into preservation potential energy. Perform counterfactual reasoning on the order node status offset, and generate a preservation vulnerability heatmap based on the statistical changes in delivery freshness caused by the status offset; When a new order is inserted, the address is changed, or the temperature of the temperature control box deviates from the preset temperature range, the disturbance sub-map is cut by the heat map of the disturbance location and the vulnerability of preservation, and the temperature sampling or road condition confirmation action is triggered on the vulnerable nodes in the disturbance sub-map. The perturbation subgraph, preservation potential energy, time window, travel time, traffic status, and temperature rise coefficient are input into a graph attention policy network with physical modulation. The attention weights between nodes are modulated according to the difference in preservation potential energy caused by the service order. The non-negative part of the change in preservation potential energy before and after the action is added to the policy loss, and the adjusted delivery route and information collection instructions are output.

2. The method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest according to claim 1, characterized in that, For each package, the current freshness index, remaining shelf life, and hard constraints on freshness are determined, further including: Read the package temperature within the sampling interval according to the temperature sequence of the temperature control box; The packaging temperature, pre-finger factor, and activation energy are input into the temperature decay term, and the retention result of the freshness index at the previous moment after the effect of the temperature decay term is used as the freshness index at the next moment. The remaining shelf life is calculated by accumulating the freshness index along the unfinished delivery route at the next moment. A critical package marker is generated when the remaining shelf life is lower than the preset warning condition, and a normal package marker is retained when the remaining shelf life is not lower than the preset warning condition.

3. The method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest according to claim 1, characterized in that, Performing counterfactual reasoning on the order node state offset further includes: Insertion location, insertion order parameters, adjacent order service order, adjacent order time window, adjacent package temperature box superimposed heat effect, and road segment traffic status are written into the variable nodes of the structural causal model; Estimate causal edges between variable nodes using historical delivery samples; Apply state parameter offset intervention to nodes with incomplete orders, compare the freshness index of neighboring orders before and after the intervention, and write the maximum or average value of the difference in freshness index into the preservation vulnerability heat map.

4. The method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest according to claim 1, characterized in that, The perturbation submap is further segmented using a heatmap of perturbation location and preservation vulnerability, including: Use the location of the new order insertion, the location of the address change, or the location of the package with abnormal temperature as the disturbance seed; Read order nodes affected by disturbed seeds from the heat map of preservation vulnerability; Expand the neighborhood of the affected order node according to the constraints of historical path segments, estimated travel time and delivery time window; Neighborhood edges that do not meet the delivery time window constraint or the hard freshness constraint are removed, and perturbation subgraphs are generated for the remaining order nodes and path edges.

5. The method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest according to claim 1, characterized in that, Triggering temperature sampling or road condition confirmation actions for vulnerable nodes in the perturbation subgraph, further including: For vulnerable nodes, simulate actions such as shortening the temperature sampling interval, confirming driver road conditions, and maintaining the sampling frequency. The structural causal model is used to estimate the uncertainty change of preservation potential energy before and after the execution of each type of action; The amount of uncertainty reduction is calculated by subtracting the sensor power consumption, communication traffic, and driver confirmation delay. The action with the largest difference is written into the features of the perturbation subgraph node, and the collected return data is input into the graph attention policy network with physical modulation.

6. The method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest according to claim 1, characterized in that, Modulating the attention weights between nodes based on the preservation potential difference caused by the service order, further including: Calculate the preservation potential energy of the second order node when the first order node serves before the second order node, and use it as the first preservation potential energy; Calculate the preservation potential energy of the second order node when the second order node serves before the first order node, and use it as the second preservation potential energy; The difference between the first preservation potential energy and the second preservation potential energy is input into the exponential modulation term; The attention coefficients of the standard graph are modulated by an exponential modulation term and then normalized to generate freshness causal attention weights; Input the causal attention weight of preservation into the service sequence action selection.

7. The method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest according to claim 1, characterized in that, After outputting the adjusted delivery route and information collection instructions, the following further steps are included: Substituting the forward delivery temperature sequence into the preservation potential energy calculation yields the predicted delivery preservation potential energy, and substituting the residual temperature sequence of the recovery temperature control box into the Arrhenius recursive relation yields the actual preservation potential energy. Calculate the duality loss between the predicted delivery preservation potential energy and the actual preservation potential energy; When the duality loss exceeds the preset trigger threshold, the target batch activation energy correction factor is updated and corrected online using Bayesian method and then input into the next round of freshness index calculation. The target batch activation energy correction factor is retained when the duality consistency loss does not exceed the preset trigger threshold. When deploying across scenarios, the physical conservation regularization term is added to the gradient update of the meta-learning inner layer to form the initial parameters of the policy network, and the initial parameters of the policy network are updated with local periodic delivery data.

8. The method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest according to claim 5, characterized in that, The action with the largest difference is written into the features of the perturbation subgraph node, further including: Establish a candidate queue for data collection actions; The uncertain range of preservation potential energy of vulnerable nodes, the remaining power of the temperature control box, the communication delay, and the driver confirmation delay are written into the candidate action attributes respectively; Candidate actions that do not meet the power or delay conditions will be eliminated. Candidate actions are sorted according to the difference between the reduction in uncertainty and the acquisition cost, and the action with the highest ranking is used to generate an information acquisition instruction.

9. The method for real-time adjustment of the periodic delivery route of freshly stewed bird's nest according to claim 7, characterized in that, The online Bayesian update and correction of the target batch activation energy correction factor further includes: Extract the temperature integral and receipt timestamp after receiving from the residual temperature sequence of the recycled temperature control box; The temperature integral after receiving the goods is input into the Arrhenius recursive relation to deduce the actual freshness index. The posterior distribution of the activation energy correction factor for the target batch is updated by the difference between the predicted freshness index and the actual freshness index. When the posterior distribution variance meets the preset convergence condition, the posterior mean is input into the next round of package freshness recursion. When the posterior distribution variance does not meet the preset convergence condition, the activation energy correction factor of the target batch in the previous round is retained and the recovery temperature sequence continues to be accumulated.

10. A real-time adjustment system for the periodic delivery route of freshly stewed bird's nest, characterized in that, include: The data aggregation module is configured to collect periodic subscription orders, product batches, customer coordinates and receiving time windows, vehicle locations, temperature sequences of temperature control boxes, door opening and closing records, and residual temperature sequences of recycled temperature control boxes. The freshness recursion module is configured to input the initial freshness index, pre-exponential factor, activation energy, real-time temperature and sampling interval of the product batch into the Arrhenius recursion relation, and form the current freshness index, remaining shelf life and shelf hard constraint boundary for each package. The preservation potential energy module is configured to calculate the predicted delivery freshness index of unfinished orders along the current path timeline, and convert the predicted delivery freshness index and the minimum acceptable freshness threshold into preservation potential energy. The causal reasoning module is configured to perform counterfactual reasoning on the order node state offset and generate a heat map of preservation vulnerability based on the statistical changes in delivery freshness caused by the state offset. The disturbance response module is configured to cut a disturbance sub-map based on the disturbance location and the preservation vulnerability heatmap when a new order is inserted, the address is changed, or the temperature of the temperature control box deviates from the preset temperature range, and trigger temperature sampling or road condition confirmation actions for vulnerable nodes in the disturbance sub-map. The path solving module is configured to input the perturbation subgraph, preservation potential energy, time window, travel time, traffic status and temperature rise coefficient into a graph attention policy network with physical modulation, modulate the attention weights between nodes according to the difference in preservation potential energy caused by the service order, and add the non-negative part of the change in preservation potential energy before and after the action to the policy loss, and output the adjusted delivery path and information collection instructions.

Images