Intelligent identification and sorting method and system for express terminal based on agent cooperation

Abstract

The application discloses an express terminal intelligent identification and sorting method and system based on agent cooperation, relates to the field of logistics automation technology, and comprises the following steps: inputting package multi-modal data collected by an intelligent identification robot into a preset agent, analyzing and converting the package multi-modal data into standardized package data; comparing the standardized data with abnormal judgment rules by the preset agent, adding an abnormal label to a package triggering a rule and distributing the package to an abnormal processing queue; for a package not triggering a rule, calling an external interface to obtain timing idle data, generating reservation time data and writing the reservation time data into the package data; planning a flow path according to address information, reservation time and a sorting line load state, and generating a sorting instruction; updating a state and associating a distribution vehicle after sorting is completed; pushing an event to a digital twin system and writing the event into a timing database when the state is changed, and returning state data in response to a query request. The application realizes the intelligentization, precision and whole-process cooperation of express terminal sorting.

Description

Technical Field

[0001] This invention relates to the fields of logistics automation and artificial intelligence technology, and more specifically, to a smart identification and sorting method and system for express delivery terminals based on intelligent agent collaboration. Background Technology

[0002] In recent years, the continued boom in e-commerce has driven the rapid growth of express delivery volume, especially in last-mile delivery services such as community stations and campus express service centers, where the daily parcel volume typically reaches hundreds or even thousands. Last-mile sorting, as a crucial node connecting transportation and delivery, directly determines user experience and operating costs through its efficiency and accuracy. However, existing last-mile sorting technologies still have many shortcomings and urgently need improvement. Currently, last-mile sorting largely relies on manual operation, requiring staff to spend long hours scanning waybills and sorting parcels, resulting in high labor intensity, low efficiency, and a significant increase in sorting error rates with longer working hours. During peak periods, staff shortages and parcel backlogs are particularly prominent. Meanwhile, express waybills often suffer from wrinkles, damage, random barcode orientation, and inconsistent placement. Traditional handheld scanning devices require manual alignment, severely impacting sorting speed, while fixed single-sided scanning devices cannot solve the problem of reading barcodes from multiple sides and in any orientation.

[0003] In terms of space, last-mile delivery stations have limited space, and traditional horizontal sorting lines occupy a large area, making it difficult to deploy automated equipment. Abnormal items (such as incorrect addresses, empty packages, and damaged items) are often only discovered in later stages, leading to high reverse logistics costs and numerous customer complaints. Existing systems lack source identification and isolation mechanisms. Regarding information management, the sorting, temporary storage, and delivery stages are fragmented, lacking a unified scheduling center and creating information silos. Users cannot know the accurate delivery time and often need to repeatedly check. The sorting system lacks integration with external systems such as delivery vehicles, parcel lockers, and user schedules, making it difficult to implement proactive appointment services. Furthermore, existing equipment mostly uses a closed architecture, resulting in poor compatibility and high costs for system expansion and upgrades.

[0004] In summary, existing last-mile delivery sorting technologies still have significant shortcomings in terms of identification reliability, space utilization efficiency, intelligent scheduling throughout the entire process, and the ability to handle abnormal shipments. Therefore, there is an urgent need for a more efficient, accurate, intelligent express delivery sorting solution that can adapt to the complex environment of last-mile delivery. This solution should introduce an intelligent agent collaboration mechanism to achieve a closed-loop and dynamic decision-making process across the entire chain, from identification and sorting to delivery. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes an intelligent identification and sorting method and system for express delivery terminals based on intelligent agent collaboration. This system solves the problems of low sorting efficiency, difficulty in information collection, low space utilization, and lack of intelligent management throughout the entire process, thereby achieving a closed-loop system and intelligent decision-making across the entire chain from sorting to delivery.

[0006] The first aspect of this invention provides an intelligent identification and sorting method for express delivery terminals based on agent collaboration, comprising the following steps: The multimodal data of packages collected by the intelligent recognition robot is input into the preset intelligent agent, which parses the data formats of waybills from different express companies and converts the parsing results into standardized package data. The preset intelligent agent compares the standardized package data with the preset anomaly judgment rules, adds anomaly tags to packages that trigger the rules, and assigns them to the anomaly processing queue. For packages that do not trigger the exception judgment rules, the preset intelligent agent calls an external interface to obtain time-series idle data based on the recipient identifier, generates appointment time data, and writes it into the package data; The preset intelligent agent plans the flow path and generates sorting instructions based on the address information and appointment time data in the package data and the sorting line load status data. After receiving the sorting completion event signal, the preset intelligent agent updates the package data status to sorted, associates the package data with the delivery vehicle identifier, and sends it to the delivery vehicle system. The preset intelligent agent generates a state change event when the package state changes, pushes it to the digital twin system interface and writes it into the time series database, and responds to query requests by retrieving state data from the time series database and returning it.

[0007] In this solution, the preset intelligent agent converts the data into standardized package data, specifically including: Extract the waybill image, barcode information, weight data, volume data, and damage detection results collected by the intelligent recognition robot as multimodal data, and import them into the preset intelligent agent; Using the moment when the package passes through the center line of the identification area as the reference point, an adaptive time window is dynamically calculated based on the conveyor belt speed. Data points whose timestamps fall within the current window are extracted from the buffer queues of each sensor. For sensors with no data points within the window, nearest neighbor interpolation estimation is performed, and interpolation markers are added to the standardized data. The imported form image is analyzed to identify each text block and convert it into a feature vector. The feature vector includes the distribution of text character types, text length, normalized position coordinates and surrounding visual features. All text blocks are constructed as a graph structure and input into a graph neural network. The semantic category probability distribution corresponding to each text block is output through a message passing mechanism. The text block with the highest probability is selected as the value of the corresponding field, and the probability value is output as the field confidence. Values ​​of the same field from different sources are compared. If inconsistencies are found, a weighted voting decision is made for discrete fields, and a Bayesian inference fusion is performed for continuous fields. The basic credibility weights of each data source are dynamically adjusted based on the accuracy of historical data.

[0008] In this scheme, the preset intelligent agent compares standardized package data with preset anomaly judgment rules, adds anomaly tags to packages that trigger the rules, and assigns them to the anomaly handling queue, specifically including: Extract package feature vectors from standardized package data and import them into the preset intelligent agent. The package feature vectors include data completeness, data source type, physical attribute availability, and courier company identification. The preset intelligent agent filters rules based on the package feature vector and the preconditions of each rule in the rule base, and calculates the weighted sum of the activation weight and computation cost of each rule. The agent generates a rule execution sequence from high to low according to the weighted sum result, where the activation weight is a positive correlation factor and the computation cost is a negative correlation factor. The computation cost is mapped to a dimensionless cost score through a conversion coefficient. If the number of triggered anomalies reaches the preset limit or the package is determined to be of the highest anomaly level during the rule execution process, the execution of subsequent rules will be terminated. The preset intelligent agent calls the pre-trained variational autoencoder to calculate the reconstruction error of the package feature vector, searches the boundary region of the normal package feature distribution through the adversarial training mechanism of the generative adversarial network, and performs density peak clustering on the boundary samples to generate a set of abnormal sensitive points. Calculate the Mahalanobis distance from the package feature vector to the nearest anomaly sensitive point, and combine the reconstruction error with the reciprocal of the Mahalanobis distance to synthesize an anomaly score. Dynamically calculate the anomaly score threshold based on the distribution statistics of all package anomaly scores within a preset time period, and determine packages with anomaly scores greater than the anomaly score threshold as anomalies. The judgment results of each rule in the rule execution sequence are fused with the anomaly score to generate a comprehensive anomaly judgment result, and the anomaly inference chain is recorded. Based on the comprehensive anomaly judgment result, the corresponding anomaly tag is added to the package, the abnormal package is assigned to the corresponding processing queue, and the anomaly inference chain is attached to the package data.

[0009] In this solution, for packages that do not trigger the anomaly detection rules, the preset intelligent agent calls an external interface to obtain time-series idle data based on the recipient identifier, generates appointment time data, and writes it into the package data. Specifically, this includes: Extract the recipient's historical pickup records, package query logs, and static attribute data, import them into a time-series graph attention network, and generate user behavior embedding vectors. The user behavior embedding vector, the time-series idle data obtained from the external interface, the current delivery resource status, and the time distribution of reserved packages are used as the state space and imported into a multi-objective deep reinforcement learning scheduler. In the multi-objective deep reinforcement learning scheduler, a multi-objective proximal policy optimization algorithm is adopted to output the selection probability distribution of each candidate time window, and the time window with the highest probability is selected as the initial reservation time. Conflict detection is performed on the initial reservation time, and a conflict graph is constructed with allocated reservations as nodes and resource capacity conflicts or user time conflicts as edges. If a conflict is detected, a constraint satisfaction search is performed until the conflict is eliminated. The reservation confidence score is calculated based on historical punctuality rate, real-time traffic congestion index, weather impact coefficient, vehicle status, and current vehicle parcel density. The corresponding reservation time format is output according to the threshold range of the confidence score, and the final reservation time data and confidence score are written into the parcel data.

[0010] In this solution, the preset intelligent agent plans the flow path based on the address information and appointment time data in the package data, combined with the sorting line load status data, specifically including: A digital twin model of the sorting line is constructed, and the dynamic attributes of the sorting line are extracted. The dynamic attributes include the current load rate, queue length, overflow flag and health status. The state of each dynamic attribute is estimated and smoothed real-time state data is output. Each package is modeled as an independent package agent, and each package agent carries priority weight, time constraints, and physical constraints. The pre-set agents periodically publish a list of available resource units. Each package agent calculates the bid amount for the resource unit based on its own attributes and submits a bid. The package agent with the highest bid wins the right to use the resource unit. Before executing the bid, the package agent performs a finite-time Pareto optimal path search, outputs a set of candidate paths, and selects the execution path from the candidate paths based on the remaining virtual budget.

[0011] In this solution, the generation of sorting instructions specifically includes: Establish queuing theory models for each key node to predict the queue length and congestion probability for multiple future time windows. When the predicted waiting time of any node exceeds the time constraint margin of the package or the congestion probability exceeds the set threshold, a forward-looking rerouting is triggered to generate an alternative path and perform group rerouting coordination for multiple packages that are rerouting simultaneously. Real-time monitoring of instruction queue length, average instruction processing latency, and PLC occupancy rate; load levels are classified according to comprehensive load score, including low load level, medium load level, and high load level. Generate fine-grained instructions at low load levels, with each package containing independent execution parameters; Under medium load levels, a package clustering algorithm is executed to merge packages whose path and time similarity meet preset requirements into batch instructions. Under high load levels, it matches pre-generated path templates to generate coarse-grained instructions that include path template identifiers and a package list.

[0012] In this solution, after receiving the sorting completion event signal, the preset intelligent agent updates the package data status to "sorted" and associates the package data with the delivery vehicle identifier before sending it to the delivery vehicle system. Specifically, this includes: After the package enters the sorting process, the preset intelligent agent queries the delivery vehicle responsible for the corresponding area based on the package's scheduled time data and target area identifier, reserves a matching compartment on the vehicle, and writes the reserved information into the package data. Obtain the planned delivery routes of delivery vehicles, assign a sequence index to each delivery point, sort the packages to be loaded in ascending order according to the delivery sequence index, generate a loading sequence, and recalculate the sequence index and update the loading sequence when the delivery route changes. After the package falls into the compartment, the sorting end generates a falling signal and sends it to the vehicle end. After receiving the signal, the vehicle end locks the compartment and generates a locking confirmation and sends it to the system end. After receiving the confirmation, the system end updates the package status to loaded and generates a record receipt and sends it to both the sorting end and the vehicle end. The confirmed loading plan and loading sequence will be sent to the delivery vehicle's onboard system.

[0013] In this solution, the preset intelligent agent generates a state change event when the package state changes, pushes it to the digital twin system interface, and writes it into the time-series database, specifically including: Each state change event is written to both hot storage and cold storage. The hot storage is used to store events within the most recent preset time period, and the cold storage is used to permanently store all historical events. Perform summary aggregation on consecutive status change events of the same package within a preset time window, merging multiple fine-grained events into a summary event; Based on the event type, query the preset push conditions, combine them with the user preference configuration to determine whether to push the status change event, send the event to the digital twin system interface according to the push strategy, and write it into the time series database.

[0014] A second aspect of the present invention provides an intelligent identification and sorting system for express delivery terminals based on intelligent agent collaboration, the system comprising: The intelligent recognition module is used to collect multimodal data of the package; A pre-defined intelligent agent is used to receive and standardize the multimodal data, compare the standardized data with pre-defined anomaly judgment rules, add anomaly tags to packages that trigger the rules and allocate them to the anomaly processing queue; for packages that do not trigger the rules, it calls an external interface to obtain time-series idle data and generate appointment time data based on the recipient identifier; based on the package's address information and appointment time data, it plans the flow path and generates sorting instructions in conjunction with the sorting line load status data; after receiving a sorting completion event signal, it updates the package data status to "sorted" and associates the package data with the delivery vehicle identifier before sending it to the delivery vehicle system; when the package status changes, it generates a status change event, pushes it to the digital twin system interface and writes it into the time-series database; The sorting control system is used to receive the sorting instructions, convert the sorting instructions into electromechanical control signals, and provide real-time feedback on the operating status of the physical equipment. The multi-layer automated sorting and conveying system uses lifting or lowering mechanisms to transfer packages between layers. Each layer is equipped with multiple sorting slots, at least one of which is directly connected to a smart express cabinet. The delivery collaboration system is used to realize two-way data exchange between the preset intelligent agent and the on-board system of the delivery vehicle, including the distribution of loading list and route planning data, as well as the receipt of status information transmitted back by the vehicle. The fulfillment platform is used to encapsulate the connection capabilities with external data sources, obtain time-series idle data based on the recipient identifier, calculate and return the scheduled time window; A digital twin and visualization system is used to receive status change events and visualize the operational status in real time, including data dashboards and logistics maps. The data storage system is used to persistently store state change events and package status data, including time-series databases, event log storage, hot storage, and cold storage.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes a pre-set intelligent agent as a unified decision-making center, achieving a closed-loop and dynamic scheduling across the entire chain from identification and sorting to delivery. This solves the problems of traditional last-mile sorting, which relies on manual operation, is inefficient, and prone to errors. The pre-set intelligent agent introduces a multi-dimensional anomaly detection mechanism that automatically identifies and accurately diverts abnormal items such as missing addresses, abnormal weights, and damage at the sorting source, effectively preventing abnormal items from entering the normal delivery process and reducing reverse logistics costs and user complaint rates. Through real-time interaction between the intelligent agent and an external scheduling system, it automatically calculates and schedules the optimal delivery time, deeply coupling sorting actions with delivery services. This transforms the process from passive waiting to proactive scheduling, significantly improving the first-time delivery success rate and user fulfillment experience. In terms of route planning, the intelligent agent dynamically generates sorting instructions based on the real-time load status of the sorting line. Combined with a multi-layered, three-dimensional sorting system, this achieves efficient package flow within a limited space, greatly improving the sorting capacity per unit area. In the loading coordination stage, real-time communication between the intelligent agent and delivery vehicles enables automatic synchronization of package lists, scheduled times, and route planning, forming a seamless connection between sorting and delivery.

[0016] The digital twin dashboard displays real-time package flow, vehicle location, and operational data, providing users with transparent and accurate logistics information query services, while offering managers a global visual monitoring tool. End-to-end event tracing and status change management ensure complete traceability of package status, providing data support for dispute resolution and operational optimization. Furthermore, the system is compatible with hardware devices from different manufacturers through a unified device access gateway, exhibiting excellent scalability and maintainability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0018] Figure 1 A flowchart of an intelligent identification and sorting method for last-mile delivery based on agent collaboration is shown. Figure 2 The flowchart illustrates the comparison of standardized package data with preset anomaly detection rules. Figure 3 The flowchart illustrates the process of calling an external interface to obtain time-series idle data and generate appointment time data. Figure 4 A block diagram of an intelligent identification and sorting system for last-mile delivery based on agent collaboration is shown. Detailed Implementation

[0019] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0021] like Figure 1 As shown, this embodiment provides an intelligent identification and sorting method for express delivery terminals based on agent collaboration, including: The multimodal data of packages collected by the intelligent recognition robot is input into the preset intelligent agent, which parses the data formats of waybills from different express companies and converts the parsing results into standardized package data. The preset intelligent agent compares the standardized package data with the preset anomaly judgment rules, adds anomaly tags to packages that trigger the rules, and assigns them to the anomaly processing queue. For packages that do not trigger the exception judgment rules, the preset intelligent agent calls an external interface to obtain time-series idle data based on the recipient identifier, generates appointment time data, and writes it into the package data; The preset intelligent agent plans the flow path and generates sorting instructions based on the address information and appointment time data in the package data and the sorting line load status data. After receiving the sorting completion event signal, the preset intelligent agent updates the package data status to sorted, associates the package data with the delivery vehicle identifier, and sends it to the delivery vehicle system. The preset intelligent agent generates a state change event when the package state changes, pushes it to the digital twin system interface and writes it into the time series database, and responds to query requests by retrieving state data from the time series database and returning it.

[0022] It should be noted that once a package enters the sorting process, the intelligent recognition robot initiates a multi-dimensional scan. The intelligent recognition robot integrates a high-definition industrial camera, weight sensor, and 3D vision sensor to collect image data from the waybill, obtain the actual weight data of the package, and measure the package's dimensions and surface flatness for damage detection. The waybill images, barcode information, weight data, volume data, and damage detection results collected by the intelligent recognition robot are extracted as multimodal data and imported into the preset intelligent agent.

[0023] Accurate alignment of multi-sensor data eliminates data mismatch issues caused by differences in equipment physical characteristics. Using the moment the package passes the center line of the identification area as a reference point, an adaptive time window is dynamically calculated based on the conveyor belt speed. Data points whose timestamps fall within the current window are extracted from the buffer queues of each sensor. For sensors without data points within the window, nearest neighbor interpolation estimation is performed, and interpolation markers are added to the standardized data. Layout analysis is performed on the imported waybill image to identify each text block and convert it into a feature vector. The feature vector includes the text character type distribution, text length, normalized position coordinates, and surrounding visual features. All text blocks are constructed as a graph structure and input into a graph neural network. Each text block is a node in the graph, and the spatial adjacency relationships between nodes form edges. This graph structure is fed into a pre-trained graph neural network. The graph neural network aggregates the features of neighboring nodes through a message passing mechanism, outputting a semantic category probability distribution for each node. Possible semantic categories include: recipient name, recipient phone number, recipient address, tracking number, sender information, etc. The text block with the highest probability is selected as the value of the corresponding field, and the probability value is output as the field confidence score. Values ​​of the same field from different channels are compared. If inconsistencies exist, a weighted voting decision is performed on discrete fields. Each candidate value receives the sum of the product of its source data source's basic confidence weight and the field confidence score. The candidate value with the highest score is selected as the final value. For continuous fields, Bayesian inference fusion is performed. The measurements of each sensor are treated as independent observations of the true value. The agent calculates the value with the highest posterior probability as the fused output, using the historical accuracy (variance) of each sensor as the conditional probability. For example, if the historical accuracy of the weight sensor is high but the volume estimation accuracy is low, the final weight value will be more biased towards the weight sensor data, but the volume sensor data will play a corrective role. The basic reliability weights of each data source are dynamically adjusted based on the accuracy of historical data. Conflict events and their adjudication results are recorded in a conflict log for periodic adjustment of the data source weights.

[0024] It should be noted that the preset intelligent agent compares standardized package data with preset anomaly judgment rules, adds anomaly tags to packages that trigger the rules, and assigns them to the anomaly processing queue. For example... Figure 2As shown, package feature vectors are extracted from standardized package data and imported into the preset agent. These package feature vectors include data completeness, data source type, physical attribute availability, and courier company identification. The preset agent maintains a rule base, where each rule includes not only the decision logic but also the data conditions required for rule execution, the applicability of the rule to the current package feature vector, and the computational resource consumption required to execute the rule. The preset agent filters rules based on the package feature vectors and the preconditions of each rule in the rule base, and calculates a weighted sum of the activation weight and computational cost of each rule. , For the activation weight of the rule, For the converted cost score, To activate the weighting coefficients, The cost penalty coefficient is used to generate a rule execution sequence from high to low based on the weighted summation result. The activation weight is a positive correlation factor, and the computational cost is a negative correlation factor. The computational cost is mapped to a dimensionless cost score through a conversion coefficient, which linearly maps the computational cost to the same value range as the activation weight. Let the computational cost of a certain rule be... Define the transformed cost score as ; The minimum computational cost for all rules in the rule base. The maximum computational cost of all rules in the rule base. The maximum value for the cost score is set. If the number of triggered anomalies reaches the preset limit or the package is determined to be of the highest anomaly level during rule execution, the execution of subsequent rules will be terminated.

[0025] Parcel data marked as normal and confirmed to be without anomalies within the past 30 days are extracted from the historical sorting database. A variational autoencoder (VAE) is trained on these parcels. After training, a pre-defined agent constructs a generative adversarial network (GAN) to search for suspicious regions near the normal distribution boundary. The generator is constrained to generate samples only in low-density regions of the normal distribution. Specifically, the generator receives a random noise vector and boundary exploration coefficients, and outputs a synthesized feature vector. During training, the generator is optimized to maximize the discriminator's output, but is simultaneously constrained by a distance penalty term, ensuring that generated samples are far from the high-density regions reconstructed by the VAE. The discriminator receives real normal samples and boundary samples synthesized by the generator, and its output represents the probability that the input sample belongs to the real normal distribution. The discriminator is trained to distinguish between real samples and boundary samples as much as possible. The pre-set intelligent agent invokes a pre-trained variational autoencoder to calculate the reconstruction error of the package feature vector. It then searches for the boundary region of the normal package feature distribution using the adversarial training mechanism of a generative adversarial network, and performs density peak clustering on the boundary samples to generate a set of anomaly sensitive points. The agent calculates the Mahalanobis distance from the package feature vector to the nearest anomaly sensitive point, and combines the reconstruction error with the inverse of the Mahalanobis distance to synthesize an anomaly score. Based on the distribution statistics of all package anomaly scores within a preset time period, an anomaly score threshold is dynamically calculated, and packages with anomaly scores greater than the threshold are judged as anomalies. The judgment results of each rule in the rule execution sequence are fused with the anomaly score to generate a comprehensive anomaly judgment result, and an anomaly inference chain is recorded. This inference chain includes a trigger rule identifier, an input data snapshot, judgment criteria, and a timestamp. Based on the comprehensive anomaly judgment result, a corresponding anomaly tag is added to the package, the abnormal package is assigned to the corresponding processing queue, and the anomaly inference chain is appended to the package data.

[0026] It should be noted that for packages that do not trigger the anomaly detection rules, the preset intelligent agent calls an external interface to obtain time-series idle data based on the recipient's identifier, generates appointment time data, and writes it into the package data. For example... Figure 3As shown, historical pickup records, package query logs, and static attribute data of recipients are extracted to model each user's historical behavior sequence as a time series graph. Nodes in the graph represent user states in different time slices, and edges represent transition relationships between time slices. Each node contains behavioral features within a time slice. These features are imported into a time series graph attention network. Temporal attention assigns different attention weights to behaviors in different time slices, spatial attention aggregates the behavior of people in the user's region, and multi-head attention learns different patterns of user behavior in parallel to generate user behavior embedding vectors. The user behavior embedding vectors, time-series idle data obtained from external interfaces, current delivery resource status, and the time distribution of reserved packages are used as the state space and imported into a multi-objective deep reinforcement learning scheduler. In this scheduler, a multi-objective proximal policy optimization algorithm is used to output the selection probability distribution of each candidate time window, and the time window with the highest probability is selected as the initial reservation time. The scheduler's reward function simultaneously optimizes user convenience rewards and delivery efficiency rewards. , , in Based on the user's historical appointment acceptance rate, This is a weighting coefficient for delivery efficiency.

[0027] The pre-defined agent is trained using a multi-objective proximal policy optimization algorithm. A policy network and a value network are maintained. The policy network outputs the probability of choosing each action; the value network outputs the expected cumulative reward vector for each state. During training, trajectories are sampled from the environment, and the advantage function (the deviation between the actual reward and the expected reward) is calculated for each time step. The loss function of the policy network consists of three terms: policy gradient loss, value loss, and entropy loss. Multi-objective optimization is achieved by vectorizing the value network output. The advantage function in the policy gradient loss is calculated as a weighted sum of the three objectives, with the weights dynamically adjusted based on the current Pareto front.

[0028] Conflict detection is performed on the initial reservation times. A conflict graph is constructed with allocated reservations as nodes and resource capacity conflicts or user time conflicts as edges. If a conflict is detected, a constraint satisfaction search is performed, which includes at least one operation among local adjustment, time window swapping, and cascading adjustment, until the conflict is resolved or the search depth reaches its limit. The new package is assigned to the highest priority time window. The conflict graph is checked for conflict edges after allocation. If no conflict exists, the allocation is accepted. If a conflict exists, all nodes within the conflict area are extracted. For each conflict node, it is checked whether the user with the associated reservation has other available time windows. If so, an adjustment candidate set is generated. The conflict node is attempted to be reallocated. If the conflict is resolved after reallocation, the adjustment is performed. If a single node adjustment fails, a two-node swap is attempted: the time window of conflict node A is swapped with the time window of conflict node B, and the conflict is checked. If the swap is successful, the swap is performed. If all of the above fail, the conflict area is expanded by one level, and the process is repeated recursively.

[0029] The reservation confidence score is calculated based on historical on-time performance, real-time traffic congestion index, weather impact coefficient, vehicle status, and current vehicle package density. The corresponding reservation time format is output based on the threshold range of the confidence score, and the final reservation time data and confidence score are written into the package data. Delivery records for the same time period and area within the past 30 days are queried from the historical database to calculate the historical on-time performance. The congestion index (0-1, higher for more congested) of the current route is obtained using the real-time traffic API. Precipitation probability and amount are obtained from the weather API and converted into impact coefficients (1 for no rain, 0.9 for light rain, 0.7 for moderate rain, and 0.5 for heavy rain). The current vehicle's health score and remaining battery range are obtained from the vehicle management system. These four factors are used as multipliers for the base confidence score. Furthermore, package density is considered: for every 80% increase in the number of packages allocated to the current vehicle beyond its capacity, the confidence score is multiplied by 0.95. The final confidence score is the base value (0.95) multiplied by each multiplier. The corresponding output format is selected based on the confidence score's range.

[0030] It should be noted that the preset intelligent agent plans the flow path based on the address information and appointment time data in the package data, combined with the sorting line load status data. A digital twin model of the sorting line is constructed. Photoelectric sensors, pressure sensors, encoders, etc. deployed on the sorting line report data at high frequency, which is preprocessed and mapped to corresponding entities. The programmable logic controller periodically reports the status of each actuator. The top-mounted camera uses image recognition algorithms to detect the package density and congestion level on the conveyor belt in real time. Dynamic attributes of the sorting line are extracted, including the current load rate, queue length, overflow flag, and health status. State estimation is performed on each dynamic attribute. Due to the delay and noise in sensor reporting, a Kalman filter is used to estimate the state of each dynamic attribute, outputting smoothed real-time state data. The prediction step of the Kalman filter uses a system dynamics model (e.g., the rate of change of load rate equals the input rate minus the output rate) to predict the current state. The update step fuses sensor observations, calculates the Kalman gain, and outputs the optimal estimate. The Kalman gain is dynamically calculated based on the observation noise covariance and the prediction covariance; sensors with high observation noise receive lower gains. The state estimate output by the filter is smoother and more accurate than the original observations.

[0031] Each package is modeled as an independent package agent, and each package agent carries priority weights. (Calculated based on factors such as the urgency of the package's reservation time, user level, and whether it is an express delivery, etc.) Time constraints (Completion of sorting deadline) and physical constraints (the size and weight of the package limit the type of conveyor belt it can use) attributes; pre-set agents periodically publish a list of available resource units, each package agent calculates the bidding amount for the resource unit based on its own attributes and submits a bid, the package agent with the highest bid wins the right to use the resource unit; first, calculate the time savings that the package agent can bring by using the resource unit. Time savings are calculated by comparing the expected remaining path time difference between using resource units and not using resource units. The baseline bid amount is then calculated. To avoid multiple agents offering the same price, a bidding coefficient is introduced. From uniform distribution Random sampling was conducted. Final bid amount. The agent also needs to check the remaining virtual budget. If the final bid amount exceeds the remaining virtual budget, the bid amount is adjusted to match the remaining virtual budget. Before executing the bid, the agent performs a finite-time Pareto optimal path search using a variant of A*. The evaluation function considers both time and budget costs, outputting multiple candidate paths on the Pareto front. Based on the current remaining budget and urgency, the agent selects a path from the Pareto front to execute.

[0032] A queuing theory model is established for each key node, with each node modeled as an M / M / 1 queue (arrival process follows a Poisson distribution, service time follows an exponential distribution, single server). The queue length and congestion probability are predicted for each node in the next 5, 10, and 20 seconds. ,in As the current captain, For the current arrival rate, For service rate, To predict the time window, predict the waiting time. The probability of congestion is expressed as If the prediction queue length of any node in any prediction window is... Exceeding the capacity threshold If a node is identified as a potential congestion node, it is then traced upstream to find all packages that will pass through that node within the predicted time window and added to the affected list. If the affected list is not empty, a rerouting flag is triggered, and the list is output.

[0033] When the predicted waiting time of any node exceeds the time constraint margin for the package or the congestion probability exceeds a set threshold, a forward-looking rerouting is triggered, generating a dynamic programming solution for an alternative path. This dynamic programming uses backward induction to solve the problem, searching backward from the target grid to the current position and calculating the optimal cost for each state. The solution process is completed within 10 milliseconds, meeting real-time requirements. For multiple packages rerouting simultaneously, group rerouting coordination is performed. This coordination groups the packages, guiding packages within the same group to the same alternative path, avoiding resource fragmentation caused by scattered rerouting. Grouping is based on the proximity of the packages' current locations and the similarity of their target areas.

[0034] The system monitors the instruction queue length, average instruction processing latency, and PLC occupancy rate in real time, calculates a weighted comprehensive load score, and classifies load levels into low, medium, and high load levels. At the low load level, fine-grained instructions are generated, with each package containing independent execution parameters. At the medium load level, a package clustering algorithm is executed to merge packages with path and time similarity meeting preset requirements into batch instructions. At the high load level, pre-generated path templates are matched to generate coarse-grained instructions containing path template identifiers and a package list. The generated sorting instructions are then sent to the sorting control system. If the current load level differs from the level at the time of the last decision, a granularity switch is triggered. The switch is gradual to avoid instability caused by sudden changes.

[0035] It should be noted that after receiving the sorting completion event signal, the preset intelligent agent updates the package data status to "sorted" and associates the package data with the delivery vehicle identifier before sending it to the delivery vehicle system. After the package enters the sorting process, the preset intelligent agent queries the delivery vehicles responsible for the corresponding area based on the package's scheduled time data and target area identifier, reserves a matching compartment on the vehicle, and writes the reservation information into the package data. After sorting, the package directly falls into the reserved compartment. Specifically, it queries the vehicle list for the responsible area and selects the vehicle with the largest remaining capacity and whose estimated arrival time matches the package's scheduled time. On the selected vehicle, it traverses the compartment layout to find a compartment that meets the following conditions: the compartment size is larger than the package volume; the compartment is currently unreserved or the reservation time does not conflict with the package's scheduled time; the compartment's position on the vehicle matches the delivery order. If multiple compartments meet the conditions, the one with the best position is selected. If no compartment meets the conditions, it triggers cascading reservation of adjacent compartments, shifting packages in conflicting compartments to the next compartment, freeing up the current compartment.

[0036] The system retrieves the planned delivery routes for delivery vehicles, assigns a sequence index to each delivery point, and sorts the packages to be loaded in ascending order according to their delivery sequence index, generating a loading sequence. When the delivery route changes, the sequence index is recalculated and the loading order is updated. Specifically, the system iterates through the list of packages to be loaded, determines the delivery point based on the package's destination address, and obtains the corresponding sequence index. All packages are sorted in ascending order according to their sequence index. For packages with the same sequence index, they are sorted in descending order by package volume, with larger items placed at the bottom. The sorted sequence of package identifiers is output as the loading order. If the delivery point sequence changes, the above sorting is re-executed, and the number of packages that need to be moved is calculated as an adjustment difference. If the difference is less than a set threshold, adjustments are made through local swaps; if the difference is large, a complete sequence is regenerated. After a package falls into the compartment, the sorting end generates a falling signal and sends it to the vehicle end. Upon receiving the signal, the vehicle end locks the compartment and generates a lock confirmation, which is sent to the system end. Upon receiving the confirmation, the system end updates the package status to "loaded" and generates a record receipt, which is sent to both the sorting end and the vehicle end. If any step times out, a retry mechanism or a degraded mode is triggered. The confirmed loading plan and loading sequence are then sent to the delivery vehicle system.

[0037] In another embodiment of the invention, a pre-defined intelligent agent places sorted packages into a loading buffer pool, generates a query hash based on the package's address information and appointment time, and queries the spatiotemporal hash signature of matching vehicles in a cuckoo filter to obtain a candidate vehicle list. The spatiotemporal hash signature is constructed by hash mapping the vehicle's spatial component (the area code the vehicle is responsible for delivering to), temporal component (the departure time window and estimated return time of the vehicle's current shift), and capacity component (the quantized code of the vehicle's remaining volume and remaining load). After precise verification of the candidate vehicles, the optimal matching vehicle is selected. The pre-defined intelligent agent maintains a cuckoo filter, storing the spatiotemporal hash signatures of all available vehicles. During the matching process, a query hash for the package is generated; the query hash is searched in the cuckoo filter. If it does not exist, it means no vehicle is a perfect match, and the fuzzy matching process begins; if it exists, a candidate vehicle list is obtained; the candidate vehicles are precisely verified, checking capacity constraints and appointment time constraints, and the optimal match is selected. When a perfect match fails, the pre-defined intelligent agent performs a Hamming distance search to find the vehicle hash with the smallest difference from the query hash. The pre-defined agent maintains a multi-level hash index, storing vehicles hierarchically according to the prefix length of the hash values. During fuzzy matching, the prefix length is gradually shortened during the search to find the vehicle with the smallest Hamming distance.

[0038] It should be noted that the preset intelligent agent generates a status change event when the package status changes, pushes it to the digital twin system interface, and writes it to the time-series database. Each status change event is simultaneously written to hot storage and cold storage to balance query performance and storage cost; the hot storage is used to store events within the last 24 hours, and the cold storage is used to permanently store all historical events. For consecutive status change events generated within a preset time window for the same package, a summary aggregation is performed, merging multiple fine-grained events into a summary event; both hot storage and push operations use the aggregated summary event, while the original events are retained in cold storage. When the aggregated event is pushed to the user and the digital twin system, only the final status is displayed. Based on the event type, preset push conditions are queried, and user preference configurations are considered to determine whether to push the status change event. For events determined to be pushable, the user's recent push records for the same package are checked. If it is not a critical event and there are recent push records, the current event is added to a batch queue, and merged and pushed after the queue is full or times out. The event is then sent to the digital twin system interface according to the push strategy and written to the time-series database.

[0039] like Figure 4 As shown, the second embodiment of the present invention provides an intelligent identification and sorting system for express delivery terminals based on intelligent agent collaboration. The system includes: The intelligent identification module is used to collect multimodal data of the package, including the waybill image, barcode information, weight data, volume data, and damage detection results; the intelligent identification module includes a 360-degree cross-scanning robot, a weight sensor, a 3D vision sensor, and a unified device access gateway. A pre-defined intelligent agent, connected to the intelligent recognition module, is used to receive and standardize the multimodal data, compare the standardized data with pre-defined anomaly judgment rules, add anomaly tags to packages that trigger the rules and allocate them to the anomaly processing queue; for packages that do not trigger the rules, it calls an external interface to obtain time-series idle data and generates appointment time data based on the recipient identifier; based on the package's address information and appointment time data, combined with the sorting line load status data, it plans the flow path and generates sorting instructions; upon receiving a sorting completion event signal, it updates the package data status to "sorted," associates the package data with the delivery vehicle identifier, and sends it to the delivery vehicle system; when the package status changes, it generates a status change event, pushes it to the digital twin system interface, and writes it into the time-series database; The sorting control system is connected to the preset intelligent agent to receive the sorting instructions, convert the sorting instructions into electromechanical control signals, and provide real-time feedback on the operating status of the physical equipment. A multi-layer automated sorting and conveying system is connected to the sorting control system. It realizes the flow of packages between layers through lifting or lowering mechanisms. Each layer is equipped with multiple sorting slots, at least one of which is directly connected to a smart express cabinet. The delivery coordination system is connected to the preset intelligent agent and is used to realize two-way data exchange between the preset intelligent agent and the delivery vehicle's on-board system, including issuing loading lists and route planning data, and receiving status information transmitted back by the vehicle. The fulfillment platform is connected to the preset intelligent agent and is used to encapsulate the connection capability with external data sources, obtain time-series idle data based on the recipient identifier, calculate and return the reservation time window; A digital twin and visualization system, connected to the pre-defined intelligent agent, is used to receive status change events and present the operational status in real time. The operational status includes a data dashboard and a logistics map. The data dashboard updates the estimated arrival volume, sorted volume, and current pickup queue time in real time. The logistics map module dynamically displays the parcel flow, delivery vehicle location, and the real-time queryable status of each parcel on a campus or community map. Additionally, a mobile query interface responds to users' parcel status query requests, returning refined status information. The data storage system, connected to the preset intelligent agent, is used to persistently store state change events and package status data, including time-series database, event log storage, hot storage, and cold storage.

[0040] The third embodiment of the present invention provides a computer-readable storage medium, which includes a program for a smart identification and sorting method for express delivery terminals based on agent collaboration. When the program for a smart identification and sorting method for express delivery terminals based on agent collaboration is executed by a processor, it implements the steps of the smart identification and sorting method for express delivery terminals based on agent collaboration.

[0041] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple modules or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be indirect coupling or communication connection through some interfaces, devices, or modules, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional modules can be integrated into one processing module, or each module can be a separate module, or two or more modules can be integrated into one module; the integrated modules can be implemented in hardware or in the form of hardware plus software functional modules.

[0042] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0043] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart identification and sorting method for express delivery terminals based on agent collaboration, characterized in that, Includes the following steps: The multimodal data of packages collected by the intelligent recognition robot is input into the preset intelligent agent, which parses the data formats of waybills from different express companies and converts the parsing results into standardized package data. The preset intelligent agent compares the standardized package data with the preset anomaly judgment rules, adds anomaly tags to packages that trigger the rules, and assigns them to the anomaly processing queue. For packages that do not trigger the exception judgment rules, the preset intelligent agent calls an external interface to obtain time-series idle data based on the recipient identifier, generates appointment time data, and writes it into the package data; The preset intelligent agent plans the flow path and generates sorting instructions based on the address information and appointment time data in the package data and the sorting line load status data. After receiving the sorting completion event signal, the preset intelligent agent updates the package data status to sorted, associates the package data with the delivery vehicle identifier, and sends it to the delivery vehicle system. The preset intelligent agent generates a state change event when the package state changes, pushes it to the digital twin system interface and writes it into the time series database, and responds to query requests by retrieving state data from the time series database and returning it.

2. The intelligent identification and sorting method for express delivery terminals based on intelligent agent collaboration according to claim 1, characterized in that, The preset intelligent agent converts the data into standardized package data, specifically including: Extract the waybill image, barcode information, weight data, volume data and damage detection results collected by the intelligent recognition robot as multimodal data, and import them into the preset intelligent agent; Using the moment when the package passes through the center line of the identification area as the reference point, an adaptive time window is dynamically calculated based on the conveyor belt speed. Data points whose timestamps fall within the current window are extracted from the buffer queues of each sensor. For sensors with no data points within the window, nearest neighbor interpolation estimation is performed, and interpolation markers are added to the standardized data. The imported form image is analyzed to identify each text block and convert it into a feature vector. The feature vector includes the distribution of text character types, text length, normalized position coordinates and surrounding visual features. All text blocks are constructed as a graph structure and input into a graph neural network. The semantic category probability distribution corresponding to each text block is output through a message passing mechanism. The text block with the highest probability is selected as the value of the corresponding field, and the probability value is output as the field confidence. Values ​​of the same field from different sources are compared. If inconsistencies are found, a weighted voting decision is made for discrete fields, and a Bayesian inference fusion is performed for continuous fields. The basic credibility weights of each data source are dynamically adjusted based on the accuracy of historical data.

3. The intelligent identification and sorting method for express delivery terminals based on agent collaboration according to claim 1, characterized in that, The pre-defined intelligent agent compares standardized package data with pre-defined anomaly detection rules, adds anomaly tags to packages that trigger the rules, and assigns them to an anomaly processing queue, specifically including: Extract package feature vectors from standardized package data and import them into the preset intelligent agent. The package feature vectors include data completeness, data source type, physical attribute availability, and courier company identification. The preset intelligent agent filters rules based on the package feature vector and the preconditions of each rule in the rule base, and calculates the weighted sum of the activation weight and computation cost of each rule. The agent generates a rule execution sequence from high to low according to the weighted sum result, where the activation weight is a positive correlation factor and the computation cost is a negative correlation factor. The computation cost is mapped to a dimensionless cost score through a conversion coefficient. If the number of triggered anomalies reaches the preset limit or the package is determined to be of the highest anomaly level during the rule execution process, the execution of subsequent rules will be terminated. The preset intelligent agent calls the pre-trained variational autoencoder to calculate the reconstruction error of the package feature vector, searches the boundary region of the normal package feature distribution through the adversarial training mechanism of the generative adversarial network, and performs density peak clustering on the boundary samples to generate a set of abnormal sensitive points. Calculate the Mahalanobis distance from the package feature vector to the nearest anomaly sensitive point, and combine the reconstruction error with the reciprocal of the Mahalanobis distance to synthesize an anomaly score. Dynamically calculate the anomaly score threshold based on the distribution statistics of all package anomaly scores within a preset time period, and determine packages with anomaly scores greater than the anomaly score threshold as anomalies. The judgment results of each rule in the rule execution sequence are fused with the anomaly score to generate a comprehensive anomaly judgment result, and the anomaly inference chain is recorded. Based on the comprehensive anomaly judgment result, the corresponding anomaly tag is added to the package, the abnormal package is assigned to the corresponding processing queue, and the anomaly inference chain is attached to the package data.

4. The intelligent identification and sorting method for express delivery terminals based on agent collaboration according to claim 1, characterized in that, For packages that do not trigger the exception judgment rules, the preset intelligent agent calls an external interface to obtain time-series idle data based on the recipient identifier, generates appointment time data, and writes it into the package data, specifically including: Extract the recipient's historical pickup records, package query logs, and static attribute data, import them into a time-series graph attention network, and generate user behavior embedding vectors. The user behavior embedding vector, the time-series idle data obtained from the external interface, the current delivery resource status, and the time distribution of reserved packages are used as the state space and imported into a multi-objective deep reinforcement learning scheduler. In the multi-objective deep reinforcement learning scheduler, a multi-objective proximal policy optimization algorithm is adopted to output the selection probability distribution of each candidate time window, and the time window with the highest probability is selected as the initial reservation time. Conflict detection is performed on the initial reservation time, and a conflict graph is constructed with allocated reservations as nodes and resource capacity conflicts or user time conflicts as edges. If a conflict is detected, a constraint satisfaction search is performed until the conflict is eliminated. The reservation confidence score is calculated based on historical punctuality rate, real-time traffic congestion index, weather impact coefficient, vehicle status, and current vehicle parcel density. The corresponding reservation time format is output according to the threshold range of the confidence score, and the final reservation time data and confidence score are written into the parcel data.

5. The intelligent identification and sorting method for express delivery terminals based on agent collaboration according to claim 1, characterized in that, The preset intelligent agent plans the flow path based on the address information and appointment time data in the package data, combined with the sorting line load status data, specifically including: A digital twin model of the sorting line is constructed, and the dynamic attributes of the sorting line are extracted. The dynamic attributes include the current load rate, queue length, overflow flag and health status. The state of each dynamic attribute is estimated and smoothed real-time state data is output. Each package is modeled as an independent package agent, and each package agent carries priority weight, time constraints, and physical constraints. The pre-set agents periodically publish a list of available resource units. Each package agent calculates the bid amount for the resource unit based on its own attributes and submits a bid. The package agent with the highest bid wins the right to use the resource unit. Before executing the bid, the package agent performs a finite-time Pareto optimal path search, outputs a set of candidate paths, and selects the execution path from the candidate paths based on the remaining virtual budget.

6. The intelligent identification and sorting method for express delivery terminals based on agent collaboration according to claim 5, characterized in that, Generate sorting instructions, specifically including: Establish queuing theory models for each key node to predict the queue length and congestion probability for multiple future time windows. When the predicted waiting time of any node exceeds the time constraint margin of the package or the congestion probability exceeds the set threshold, a forward-looking rerouting is triggered to generate an alternative path and perform group rerouting coordination for multiple packages that are rerouting simultaneously. Real-time monitoring of instruction queue length, average instruction processing latency, and PLC occupancy rate; load levels are classified according to comprehensive load score, including low load level, medium load level, and high load level. Generate fine-grained instructions at low load levels, with each package containing independent execution parameters; Under medium load levels, a package clustering algorithm is executed to merge packages whose path and time similarity meet preset requirements into batch instructions. Under high load levels, it matches pre-generated path templates to generate coarse-grained instructions that include path template identifiers and a package list.

7. The intelligent identification and sorting method for express delivery terminals based on agent collaboration according to claim 1, characterized in that, Upon receiving the sorting completion event signal, the preset intelligent agent updates the package data status to "sorted" and associates the package data with the delivery vehicle identifier before sending it to the delivery vehicle system, specifically including: After the package enters the sorting process, the preset intelligent agent queries the delivery vehicle responsible for the corresponding area based on the package's scheduled time data and target area identifier, reserves a matching compartment on the vehicle, and writes the reserved information into the package data. Obtain the planned delivery routes of delivery vehicles, assign a sequence index to each delivery point, sort the packages to be loaded in ascending order according to the delivery sequence index, generate a loading sequence, and recalculate the sequence index and update the loading sequence when the delivery route changes. After the package falls into the compartment, the sorting end generates a falling signal and sends it to the vehicle end. After receiving the signal, the vehicle end locks the compartment and generates a locking confirmation and sends it to the system end. After receiving the confirmation, the system end updates the package status to loaded and generates a record receipt and sends it to both the sorting end and the vehicle end. The confirmed loading plan and loading sequence will be sent to the delivery vehicle's onboard system.

8. The intelligent identification and sorting method for express delivery terminals based on agent collaboration according to claim 1, characterized in that, The preset intelligent agent generates a state change event when the package's state changes, pushes it to the digital twin system interface, and writes it into the time-series database. Specifically, this includes: Each state change event is written to both hot storage and cold storage. The hot storage is used to store events within the most recent preset time period, and the cold storage is used to permanently store all historical events. Perform summary aggregation on consecutive status change events of the same package within a preset time window, merging multiple fine-grained events into a summary event; Based on the event type, query the preset push conditions, combine them with the user preference configuration to determine whether to push the status change event, send the event to the digital twin system interface according to the push strategy, and write it into the time series database.

9. A smart identification and sorting system for last-mile delivery based on agent collaboration, characterized in that, For implementing the intelligent identification and sorting method for express delivery terminals based on agent collaboration as described in any one of claims 1-8, the system comprises: The intelligent recognition module is used to collect multimodal data of the package; A pre-defined intelligent agent is used to receive and standardize the multimodal data, compare the standardized data with pre-defined anomaly judgment rules, add anomaly tags to packages that trigger the rules and allocate them to the anomaly processing queue; for packages that do not trigger the rules, it calls an external interface to obtain time-series idle data and generate appointment time data based on the recipient identifier; based on the package's address information and appointment time data, it plans the flow path and generates sorting instructions in conjunction with the sorting line load status data; after receiving a sorting completion event signal, it updates the package data status to "sorted" and associates the package data with the delivery vehicle identifier before sending it to the delivery vehicle system; when the package status changes, it generates a status change event, pushes it to the digital twin system interface and writes it into the time-series database; The sorting control system is used to receive the sorting instructions, convert the sorting instructions into electromechanical control signals, and provide real-time feedback on the operating status of the physical equipment. The multi-layer automated sorting and conveying system uses lifting or lowering mechanisms to transfer packages between layers. Each layer is equipped with multiple sorting slots, at least one of which is directly connected to a smart express cabinet. The delivery collaboration system is used to realize two-way data exchange between the preset intelligent agent and the on-board system of the delivery vehicle, including the distribution of loading list and route planning data, as well as the receipt of status information transmitted back by the vehicle. The fulfillment platform is used to encapsulate the connection capabilities with external data sources, obtain time-series idle data based on the recipient identifier, calculate and return the scheduled time window; A digital twin and visualization system is used to receive status change events and visualize the operational status in real time, including data dashboards and logistics maps. The data storage system is used to persistently store state change events and package status data, including time-series databases, event log storage, hot storage, and cold storage.

Images