Intelligent analysis method, device and equipment for logistics anomaly and storage medium

Abstract

The present application relates to the technical field of data analysis, and more particularly to a logistics anomaly intelligent analysis method, device and equipment and storage medium, the method extracts the multi-modal features of each waybill in the constructed logistics space-time graph, based on the logistics space-time graph and the multi-modal features, the abnormal probability is calculated through the space-time graph convolution network and the abnormal waybill is screened, the complete logistics track sequence of the abnormal waybill is queried, the logistics stage is determined by backtracking the track, the problem piece data is generated by matching the site; based on the problem piece data and the problem receiving site, the processing result corresponding to the abnormal waybill is obtained to generate an abnormal processing information set, the root cause is identified through a causal inference model to obtain an overall cause list. The method solves the problems of missing anomaly monitoring and subjective and fuzzy determination results in the existing mode, significantly reduces the cost of logistics anomaly analysis and processing, improves the determination and disposal accuracy, and is suitable for multi-site logistics scenarios.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to an intelligent analysis method, apparatus, device, and storage medium for logistics anomalies. Background Technology

[0002] Currently, the intelligent analysis of logistics anomalies in the logistics industry, including waybill circulation, site operations, and personnel operations, faces multiple challenges. Existing anomaly control models rely excessively on manual post-event verification and experience-based judgment. Anomaly judgment standards largely depend on human experience and simple rules, lacking objective and quantifiable multi-dimensional data support. This leads to doubts about the accuracy of anomaly identification and fails to provide clear and feasible improvement directions for optimizing logistics processes.

[0003] Furthermore, the anomaly management process only extends to anomaly detection and final work order closure, lacking effective monitoring and real-time correction of the entire process trajectory of the waybill from pickup, transit, delivery to receipt. In the actual flow of waybills, problems such as missing trajectories, node delays, and non-standard operations are prone to occur. Moreover, the logistics anomaly analysis system is isolated from the core logistics business system, and the synchronization of trajectory data and waybill data is lagging, resulting in insufficient authenticity and real-time information of anomalies.

[0004] Furthermore, the existing logistics anomaly handling platforms have cumbersome process logic and lengthy feedback links, making them difficult for frontline stations and operators to use. This makes it difficult to meet the needs of scenarios that require rapid anomaly location and timely handling and rectification, further reducing anomaly handling efficiency and closure rate.

[0005] Frequently occurring logistics problems such as waybill delays, misclassification and misdelivery, abnormal tracking, and untimely handling at stations have long been unable to be systematically resolved due to the lack of efficient tracking extraction, intelligent identification, and closed-loop feedback mechanisms, which seriously affects the quality of logistics services and overall operational efficiency.

[0006] It is evident that existing technologies still need improvement and enhancement. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, the present invention aims to provide an intelligent analysis method for logistics anomalies. By combining the processing results of each abnormal waybill to perform global root cause analysis, it realizes integrated intelligent management and control of accurate identification and closed-loop rectification of logistics anomalies.

[0008] The first aspect of this invention provides an intelligent analysis method for logistics anomalies, comprising: constructing a logistics spatiotemporal graph based on logistics information; extracting multimodal features of each waybill in the logistics spatiotemporal graph; calculating anomaly probability using a pre-trained spatiotemporal graph convolutional network based on the logistics spatiotemporal graph and the multimodal features to obtain an anomaly probability score for each waybill; filtering out abnormal waybills based on the anomaly probability score; querying the complete logistics trajectory sequence of each abnormal waybill; tracing the complete logistics trajectory sequence back in reverse chronological order and matching it with a preset stage identifier code; determining the current logistics stage of each abnormal waybill based on the matching result; determining the problem receiving station corresponding to each abnormal waybill based on the logistics stage using a preset station matching rule; generating problem shipment data based on the complete logistics trajectory sequence and the logistics stage; obtaining the processing result corresponding to each abnormal waybill based on the problem shipment data and the problem receiving station; generating an anomaly processing information set based on the processing result; and performing root cause identification on the anomaly processing information set using a pre-trained causal inference model to obtain an overall causal list.

[0009] Optionally, in a first implementation of the first aspect of the present invention, the step of constructing a logistics spatiotemporal graph based on logistics information includes: generating a node set including station nodes, personnel nodes, and waybill nodes based on station information, personnel information, and waybill information in the logistics information; establishing a transportation edge connecting two station nodes, a membership edge connecting a personnel node and a station node, and a flow edge connecting a waybill node and a station node or personnel node based on the logistics flow relationship in the logistics information, thereby generating an edge set; configuring a unique identifier and basic attributes for each node in the node set, and configuring a timestamp and association attributes for each edge in the edge set; and constructing the logistics spatiotemporal graph based on the configured node set and edge set.

[0010] Optionally, in a second implementation of the first aspect of the present invention, the step of calculating the anomaly probability score for each waybill based on the logistics spatiotemporal graph and the multimodal features using a pre-trained spatiotemporal graph convolutional network includes: performing standardized preprocessing on the multimodal features of each waybill, wherein the multimodal features include text features, trajectory features, image features, and structured features; encoding the standardized preprocessed multimodal features to obtain a feature vector corresponding to each feature; weighting and fusing multiple feature vectors through an attention mechanism to generate a unified representation vector; and inputting the unified representation vector and the logistics spatiotemporal graph together into the spatiotemporal graph convolutional network to obtain the anomaly probability score for each waybill.

[0011] Optionally, in a third implementation of the first aspect of the present invention, the step of querying the complete logistics trajectory sequence of each abnormal waybill, backtracking the complete logistics trajectory sequence in reverse chronological order and matching it with a preset stage identifier code, and determining the current logistics stage of each abnormal waybill based on the matching result includes: obtaining the complete logistics trajectory sequence of each abnormal waybill in batches through a preset logistics trajectory interface, wherein the complete logistics trajectory sequence includes multiple trajectory points arranged in chronological order, and each trajectory point includes a scan time and a stage identifier code; starting from the trajectory point with the latest scan time in the complete logistics trajectory sequence, traversing each trajectory point in reverse chronological order; matching the stage identifier code in the traversed trajectory points with a preset set of stage identifier codes, stopping the traversal when any preset stage identifier code is matched, and determining the current logistics stage of the abnormal waybill based on the matched stage identifier code.

[0012] Optionally, in a fourth implementation of the first aspect of the present invention, determining the problem receiving station corresponding to each abnormal waybill according to the logistics stage and by using preset station matching rules includes: Obtain the latest scanning station in the complete logistics trajectory sequence corresponding to the abnormal waybill; when the logistics stage is the sorting stage, transit stage, or delivery stage, if the latest scanning station has the authority to handle problematic items, then the latest scanning station is determined as the problematic receiving station; otherwise, the parent station of the latest scanning station is determined as the problematic receiving station; when the logistics stage is the non-pickup stage, the station where the sender of the abnormal waybill is located is determined as the problematic receiving station; when the logistics stage is the return stage, the return destination station of the abnormal waybill is determined as the problematic receiving station.

[0013] Optionally, in a fifth implementation of the first aspect of the present invention, the step of generating problematic item data based on the complete logistics trajectory sequence and the logistics stage includes: finding the problematic item type code corresponding to the logistics stage according to a preset stage-code mapping relationship; extracting the latest scanning time, the latest scanning station, and key trajectory notes associated with the logistics stage from the complete logistics trajectory sequence; filling the waybill number of the abnormal waybill, the problematic item type code, the latest scanning time, the latest scanning station, and the key trajectory notes into a preset problem description template to generate a standardized problem description; and assembling the standardized problem description, the problem receiving station, and the current timestamp into problematic item data.

[0014] Optionally, in a sixth implementation of the first aspect of the present invention, the processing result includes a processing status and processing details. The step of generating an abnormal processing information set based on the processing result, and performing root cause identification on the abnormal processing information set using a pre-trained causal inference model to obtain an overall cause list, includes: when the processing status is processing failure, obtaining processing details and extracting failure cause categories from the processing details; fusing the abnormal probability score, logistics stage, failure cause category, and trajectory features in the complete logistics trajectory sequence corresponding to each abnormal waybill to obtain the abnormal processing information set; inputting the abnormal processing information set into the pre-trained causal inference model for causal analysis to generate a root cause list and intervention effect values ​​corresponding to each root cause; sorting each root cause in the root cause list in descending order according to the intervention effect values, determining the root cause ranked first as the overall core cause, determining the root causes ranked second and third as overall secondary causes, and outputting the overall core cause and the overall secondary causes as the overall cause list.

[0015] A second aspect of the present invention provides an intelligent analysis device for logistics anomalies, comprising: a calculation module, configured to construct a logistics spatiotemporal graph based on logistics information, extract multimodal features of each waybill in the logistics spatiotemporal graph, and calculate anomaly probability using a preset spatiotemporal graph convolutional network based on the logistics spatiotemporal graph and the multimodal features to obtain an anomaly probability score for each waybill; a filtering module, configured to filter out abnormal waybills based on the anomaly probability score, query the complete logistics trajectory sequence of each abnormal waybill, backtrack the complete logistics trajectory sequence in reverse chronological order and match it with a preset stage identifier code, and determine the current logistics stage of each abnormal waybill based on the matching result; a sending module, configured to determine the problem receiving station corresponding to each abnormal waybill based on the logistics stage and a preset station matching rule, generate problem item data based on the complete logistics trajectory sequence and the logistics stage, and send the problem item data to the corresponding problem receiving station by calling a preset problem interface; and an identification module, configured to receive the processing result of each abnormal waybill returned by the problem receiving station, generate an anomaly processing information set based on the processing result, and perform root cause identification on the anomaly processing information set using a pre-trained causal inference model to obtain an overall causal list.

[0016] A third aspect of the present invention provides an intelligent analysis device for logistics anomalies, the intelligent analysis device for logistics anomalies comprising: a memory and at least one processor, the memory storing instructions; at least one processor calling the instructions in the memory to cause the intelligent analysis device for logistics anomalies to execute the various steps of the intelligent analysis method for logistics anomalies described above.

[0017] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the intelligent analysis method for logistics anomalies described in any of the preceding claims.

[0018] The technical solution of this invention overcomes the subjective limitations of traditional manual anomaly verification. By combining multimodal trajectory feature preprocessing with spatiotemporal graph convolutional networks, it achieves objective quantification and accurate identification of logistics anomaly judgment. Through a hierarchical site matching mechanism and full-process trajectory data management, it ensures the continuity, traceability, and reproducibility of logistics anomaly analysis and handling based on logistics trajectories. This method not only specifically addresses core technical problems in existing models such as lack of anomaly monitoring, isolated data systems, inefficient handling execution, and subjective ambiguity in judgment results, but also significantly reduces the manpower and time costs of logistics anomaly analysis and handling through process closure and algorithm empowerment. It effectively shortens the anomaly closure time, improves the accuracy of anomaly judgment and handling, and enhances the fairness and standardization of anomaly governance. It provides systematic technical support for optimizing logistics operation efficiency and improving service quality, and is applicable to the standardized management needs of various logistics scenarios with multi-site flow and full-process trajectory tracking. Attached Figure Description

[0019] Figure 1 A first flowchart of the intelligent analysis method for logistics anomalies provided in an embodiment of the present invention; Figure 2 A second flowchart of the intelligent analysis method for logistics anomalies provided in an embodiment of the present invention; Figure 3 A third flowchart of the intelligent analysis method for logistics anomalies provided in this embodiment of the invention; Figure 4 The fourth flowchart of the intelligent analysis method for logistics anomalies provided in this embodiment of the invention; Figure 5 The fifth flowchart of the intelligent analysis method for logistics anomalies provided in this embodiment of the invention; Figure 6 The sixth flowchart of the intelligent analysis method for logistics anomalies provided in the embodiments of the present invention; Figure 7 The seventh flowchart of the intelligent analysis method for logistics anomalies provided in the embodiments of the present invention; Figure 8 A schematic diagram of the intelligent analysis device for logistics anomalies provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the intelligent analysis device for logistics anomalies provided in an embodiment of the present invention. Detailed Implementation

[0020] This invention provides an intelligent analysis method, apparatus, device, and storage medium for logistics anomalies. In this invention, the terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0021] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent analysis method for logistics anomalies in this invention includes: 101. Construct a logistics spatiotemporal graph based on logistics information, extract multimodal features of each waybill in the logistics spatiotemporal graph, and calculate the anomaly probability based on the logistics spatiotemporal graph and the multimodal features through a pre-trained spatiotemporal graph convolutional network to obtain the anomaly probability score of each waybill. In this embodiment, the logistics spatiotemporal graph refers to a topological structure graph that abstracts stations, personnel, and waybills in the logistics system as nodes, and abstracts relationships such as transportation between stations, personnel affiliation, and waybill flow as edges. Based on station information, personnel information, and waybill information in the logistics information, corresponding station nodes, personnel nodes, and waybill nodes are generated respectively. Then, transportation edges, affiliation edges, and flow edges are established according to the logistics flow relationships, thereby forming a set of nodes and a set of edges. By configuring unique identifiers and basic attributes for each node, and timestamps and association attributes for each edge, the construction of the logistics spatiotemporal graph is completed.

[0022] The multimodal features refer to the collective term for textual features (features extracted from the waybill text, including sender and recipient addresses, shipment remarks, etc.), trajectory features (including historical shipment flow nodes, dwell time at each node, etc.), image features (including shipment label photos, photos of damaged / abnormal scenes, etc.), and structured features (including shipment weight, volume, estimated delivery time, station type, etc.) extracted from the waybill-related data. From the constructed logistics spatiotemporal map, multidimensional data related to each waybill is extracted to obtain the corresponding waybill's multimodal features. Because the multimodal features originate from the association data between waybills, stations, personnel, and trajectories in the logistics spatiotemporal map, they can more comprehensively reflect the status information of waybills in the actual logistics scenario.

[0023] The extracted multimodal features are standardized and preprocessed and encoded. Multiple feature vectors are weighted and fused through an attention mechanism to generate a unified representation vector. This unified representation vector and the logistics spatiotemporal graph are input into a pre-trained spatiotemporal graph convolutional network. The network jointly learns the spatiotemporal correlation features and multimodal features of the waybill, and finally outputs the anomaly probability score of each waybill.

[0024] By comprehensively utilizing the topological information of the logistics network and the multimodal feature information of waybills, the system can automatically identify abnormal waybills, avoiding the inefficiency and subjectivity of manual inspection and providing accurate screening criteria for subsequent batch processing of problematic shipments.

[0025] 102. Based on the abnormal probability score, abnormal waybills are selected, the complete logistics trajectory sequence of each abnormal waybill is queried, the complete logistics trajectory sequence is traced back in reverse time and matched with the preset stage identifier code, and the current logistics stage of each abnormal waybill is determined based on the matching result. In this embodiment, the abnormal waybill refers to a waybill that may have logistics anomalies, selected from all waybills based on its abnormality probability score. Specifically, a preset threshold (e.g., 0.7) is set based on the abnormality probability score (output range [0,1]). When the abnormality probability score of a waybill is greater than or equal to the threshold, the waybill is marked as an abnormal waybill; when the abnormality probability score is lower than the threshold, it is considered a normal waybill and does not proceed to the subsequent processing flow.

[0026] The complete logistics trajectory sequence refers to the collection of all logistics operation records experienced by a waybill from its creation to the current moment. Each trajectory point corresponds to one logistics scan operation, including fields such as scan time, scan site, scan type code, scanner, scan equipment, and trajectory remarks. For each filtered abnormal waybill, the internal logistics trajectory query interface is called (e.g., the interface address is / internal / logistics / track / query, and the request method is POST), passing the waybill number of the abnormal waybill as the query parameter. To improve processing efficiency, this interface supports batch queries, with a maximum of 1000 waybill numbers allowed per query. The data returned by the interface is in JSON format, containing a trajectory list for each waybill number, where each trajectory record includes information such as scan time, scan site, and scan type code.

[0027] The reverse chronological backtracking refers to starting from the latest track point in the logistics trajectory sequence (i.e., the most recently occurred operation record) and traversing each track point sequentially in the direction of earlier times. Starting from the track point with the latest scan time, the traversal proceeds backward in reverse chronological order. Because traversing backward from the latest track point prioritizes matching the current most recent stage of the waybill's flow, it improves the timeliness of stage identification. The logistics stage refers to a classification identifier used to characterize the current flow stage of an abnormal waybill, including sorting stage, transit stage, delivery stage, uncollected stage, and return stage. The stage identifier codes in the traversed track points are matched with a preset set of stage identifier codes. Traversal stops when any preset stage identifier code is matched, and the current logistics stage of the abnormal waybill is determined based on that stage identifier code.

[0028] By screening abnormal waybills by anomaly probability scores and matching stage identifier codes in reverse chronological order based on the complete logistics trajectory sequence, the current logistics stage of each abnormal waybill can be accurately determined, providing a unified and objective basis for subsequent abnormal handling station allocation.

[0029] 103. Based on the logistics stage, determine the problem receiving station corresponding to each abnormal waybill through preset station matching rules, and generate problem shipment data based on the complete logistics trajectory sequence and the logistics stage; In this embodiment, the problem receiving station refers to a logistics station matched according to preset rules based on the logistics stage of the abnormal waybill. The problem receiving station is responsible for verifying and processing abnormal waybills. The problem shipment data refers to structured data used to push abnormal information to the problem receiving station.

[0030] First, the latest scanning station in the complete logistics trajectory sequence corresponding to the abnormal waybill is obtained. For example, when the logistics stage is sorting, transit, or delivery, the latest scanning station is identified as the problem receiving station; when the logistics stage is not yet picked up, the sender's station is identified as the problem receiving station. Then, the problem item type code corresponding to the logistics stage is determined according to a preset mapping relationship. The latest scanning time, latest scanning station, and key trajectory notes are extracted from the complete logistics trajectory sequence. The waybill number, problem item type code, and relevant trajectory information are filled into a preset template to form a standardized problem description. Then, the standardized problem description, problem receiving station, and current timestamp are assembled into problem item data. Finally, the problem item data is sent to the problem receiving station.

[0031] By intelligently matching problem receiving stations according to logistics stages and automatically generating standardized problem item data, the system can accurately push abnormal information and efficiently collect handling results, providing reliable handling data support for subsequent root cause analysis of abnormalities.

[0032] 104. Based on the problematic shipment data and the problematic receiving station, obtain the processing result corresponding to each abnormal waybill, generate an abnormal processing information set according to the processing result, and use a pre-trained causal inference model to identify the root cause of the abnormal processing information set to obtain an overall cause list. In this embodiment, the processing result refers to the structured information formed by the problem receiving station after verifying, rectifying, and providing feedback on the problem data of a single abnormal waybill. The processing result includes the processing status and processing details. Based on the pushed problem data and the corresponding problem receiving station, the processing result corresponding to each abnormal waybill is obtained, that is, the corresponding processing result returned by the problem receiving station is received according to the problem data. The abnormal processing information set refers to a dataset formed by fusing the multi-dimensional features and handling information of each abnormal waybill. When the problem receiving station indicates that the processing status of an abnormal order is "processing failure," it extracts the failure reason category from the processing details and fuses the abnormal probability score, logistics stage, failure reason category, and trajectory features in the corresponding complete logistics trajectory sequence of each abnormal waybill to obtain the abnormal processing information set.

[0033] The causal inference model employs a Bayesian network model to identify the strength of the causal relationship between each candidate factor and the occurrence of the anomaly. A Bayesian network is a probabilistic graphical model based on a directed acyclic graph (DAG), which describes the dependencies between variables through conditional probability distributions, supporting causal inference and intervention effect calculation. When the anomaly handling information set is input into the causal inference model, the model automatically identifies the root causes and outputs a comprehensive list of causes.

[0034] By collecting the processing results of each abnormal waybill and constructing an abnormal processing information set, and using a causal inference model to identify global root causes, a closed loop is achieved from individual handling of logistics abnormalities to overall root cause analysis, providing a quantitative basis for systematically optimizing logistics operation processes.

[0035] This invention discloses an intelligent analysis method for logistics anomalies, which overcomes the subjective limitations of traditional manual anomaly checks. By combining multimodal trajectory feature preprocessing with spatiotemporal graph convolutional networks, it achieves objective quantification and accurate identification of logistics anomalies. Through a hierarchical site matching mechanism and full-process trajectory data management, it ensures the continuity, traceability, and reproducibility of logistics anomaly analysis and handling based on logistics trajectories. This method not only specifically addresses core technical problems in existing models such as lack of anomaly monitoring, isolated data systems, inefficient handling execution, and subjective ambiguity in judgment results, but also significantly reduces the manpower and time costs of logistics anomaly analysis and handling through process closure and algorithm empowerment. It effectively shortens the anomaly closure time, improves the accuracy of anomaly judgment and handling, and enhances the fairness and standardization of anomaly governance. It provides systematic technical support for optimizing logistics operation efficiency and improving service quality, and is applicable to the standardized management needs of various logistics scenarios with multi-site flow and full-process trajectory tracking.

[0036] Please see Figure 2 In this embodiment of the invention, constructing a logistics spatiotemporal map based on logistics information includes: 201. Based on the site information, personnel information, and waybill information in the logistics information, generate a node set including site nodes, personnel nodes, and waybill nodes; In this embodiment, the node refers to the graph vertex in the logistics spatiotemporal diagram used to represent physical entities in the logistics network. Specifically, based on the site information, personnel information, and waybill information in the logistics information, a node set including site nodes, personnel nodes, and waybill nodes is generated. Site nodes correspond to physical sites in the logistics network such as sorting centers, transit points, and sales outlets, and each site node is associated with attributes such as site code, site type, geographical location, operating hours, and carrying capacity. Personnel nodes correspond to logistics workers such as sorters, delivery personnel, and transit specialists, and each personnel node is associated with attributes such as employee ID, job type, responsible area, and on-duty status. Waybill nodes correspond to each express waybill, and each waybill node is associated with attributes such as waybill number, express type, weight, volume, sender and recipient address, and estimated delivery time. By constructing the node set, a data foundation is provided for the construction of the logistics spatiotemporal diagram.

[0037] 202. Based on the logistics flow relationship in the logistics information, establish a transportation edge connecting two station nodes, a membership edge connecting a personnel node and a station node, and a flow edge connecting a waybill node and a station node or personnel node, and generate an edge set. In this embodiment, the edge refers to the graph connection used in the logistics spatiotemporal graph to represent the relationship between nodes. The logistics flow relationship refers to the association rules extracted from the logistics information system that characterize the dynamic interaction and static affiliation between various entities in the logistics network.

[0038] Specifically, three types of edges are established based on the logistics flow relationship. First, there are transportation edges, used to connect two station nodes, representing the physical transportation path of express parcels between sorting centers, transit hubs, and sales outlets. Each transportation edge is associated with attributes such as the means of transport (e.g., trucks, airplanes), transportation route, departure time, estimated arrival time, and actual arrival time. For example, the transportation edge from the "East China Sorting Center" to the "North China Transit Hub" records trunk line transportation information departing at 3:00 AM daily. Second, there are membership edges, used to connect personnel nodes and station nodes, representing the affiliation relationship between logistics practitioners such as sorters, delivery personnel, and transit specialists and their respective stations. Each membership edge associates affiliation relationships when they take effect. The attributes include the scope of business and the responsibility of the delivery personnel. For example, the deliveryman "Zhang San" is associated with "Chaoyang Business Point" through the subordinate edge, indicating that he is responsible for the delivery tasks in the area covered by that business point. The third is the flow edge, which is used to connect the waybill node with the station node or personnel node. It represents the interaction records of each express with the station or personnel in each logistics link such as pickup, sorting, transfer, delivery and signing. Each flow edge is associated with attributes such as flow type, operation time, operator identification and scanning device identification. For example, a waybill is associated with "Transfer Center A" and operator "Li Si" through the flow edge, which records that the express completed the transfer and arrival operation at a specific time.

[0039] By establishing three types of edges, the originally discrete logistics events are organized into a set of graph edges with clear semantics and temporal relationships, providing a complete graph structure foundation for subsequent anomaly detection.

[0040] 203. Configure a unique identifier and basic attributes for each node in the node set, and configure a timestamp and association attributes for each edge in the edge set; In this embodiment, the unique identifier and basic attributes refer to the set of fields assigned to each node to uniquely identify the node and describe its inherent characteristics.

[0041] Specifically, for station nodes, a station code (e.g., WH001) is configured as a unique identifier, along with basic attributes such as station type (sorting center, transit point, or sales point), geographical location (latitude and longitude coordinates), operating hours (e.g., 00:00-24:00), and capacity (e.g., daily processing volume of 50,000 pieces). These attributes are used to determine whether the station has the authority and capability to handle specific problematic items. For personnel nodes, employee ID is configured as a unique identifier, along with basic attributes such as job type (sorter, delivery person, or transit specialist), responsible area (e.g., Chaoyang District), and on-duty status (on-duty, off-duty, or on leave). These attributes are used to assess the impact of personnel workload and on-duty rate on the occurrence of anomalies. For waybill nodes, waybill number is configured as a unique identifier, along with basic attributes such as parcel type (ordinary parcel, fragile parcel, or perishable parcel), weight, volume, sender and recipient address, and estimated delivery time (e.g., next-day delivery). These attributes are used to identify the differences in anomaly patterns among different types of parcels (e.g., perishable parcels are time-sensitive, and delays have a greater impact).

[0042] For each edge in the edge set, configure timestamps and associated attributes: Transportation edges are configured with timestamp fields such as departure time, estimated arrival time, and actual arrival time to calculate transportation timeliness deviations; Membership edges are configured with the effective time of the membership relationship to determine whether the affiliation relationship between personnel and stations is valid at a specific point in time; Flow edges are configured with operation time as a timestamp and associated attributes such as flow type (pickup, sorting, transfer, delivery, or receipt), operator ID, and scanning device ID to trace the precise time point and operation subject of anomalies.

[0043] Through the above configuration, the originally abstract graph nodes and edges are given specific business semantics and computable numerical attributes, which transforms the logistics spatiotemporal graph from a static structural skeleton into a dynamic data model carrying rich business information. This provides multi-dimensional input data containing node state features and edge temporal features for subsequent anomaly detection based on spatiotemporal graph convolutional networks.

[0044] 204. Construct the logistics spatiotemporal graph based on the configured node set and edge set; In this embodiment, the logistics spatiotemporal graph refers to a graph structure data model that is based on the aforementioned configured set of nodes and edge sets, and incorporates a dynamic update mechanism of the time dimension to form a graph structure data model that can map the changes in the actual logistics status in real time.

[0045] Specifically, after generating and configuring unique identifiers and basic attributes for station nodes, personnel nodes, and waybill nodes, as well as establishing transportation edges, membership edges, and circulation edges and configuring timestamps and associated attributes, these nodes and edges are assembled according to their topological relationships: transportation edges connect two station nodes to form a transportation network topology, membership edges connect personnel nodes and station nodes to form an organizational affiliation topology, and circulation edges connect waybill nodes and station nodes or personnel nodes to form an operation trajectory topology. Together, these three constitute a heterogeneous graph structure containing multiple types of nodes and multiple types of edges, thus completing the construction of the logistics spatiotemporal graph.

[0046] Because the logistics system is highly dynamic, incremental data can be synchronized from the logistics system according to a preset period (such as every 5 minutes) to dynamically update the attributes of nodes and edges in the graph.

[0047] By unifying and integrating discrete logistics entities and their flow relationships into a structured graph structure, the spatiotemporal correlation information in the logistics process can be fully preserved, providing a unified data carrier for subsequent waybill feature extraction and anomaly analysis.

[0048] Please see Figure 3 In this embodiment of the invention, the step of calculating the anomaly probability score for each waybill by using a pre-trained spatiotemporal graph convolutional network based on the logistics spatiotemporal graph and the multimodal features includes: 301. Perform standardized preprocessing on the multimodal features of each waybill, wherein the multimodal features include text features, trajectory features, image features, and structured features; In this embodiment, the multimodal features refer to the multidimensional raw data representing the status of waybills obtained from different data sources, including four categories: text features, trajectory features, image features, and structured features. Since these four types of data have significant differences in format, dimension, and distribution characteristics, they cannot be directly fused. Therefore, they need to be standardized and preprocessed separately.

[0049] Specifically, the text features are derived from unstructured text data such as sender and recipient addresses, package remarks, and error feedback text. The original text contains a large number of stop words such as "express delivery" and "package," as well as invalid characters such as special symbols and line breaks. During the standardization preprocessing, the text is first segmented into words, and then stop words and invalid characters are removed. For example, "[Error] The package has been stuck in the transit center for more than 24 hours, please process it as soon as possible" is preprocessed to "Error transit center stuck for 24 hours processing" to eliminate the interference of noise information on semantic encoding.

[0050] The trajectory features originate from the historical flow nodes of the express shipment, the dwell time at each node, and the transportation time between adjacent nodes. The dwell time and transportation time vary considerably; for example, some nodes may only have a dwell time of a few minutes, while others may have a dwell time of several days. During standardization preprocessing, the Min-Max normalization method is used to map the dwell time and transportation time to the [0,1] interval. The formula is: Normalized value = (Original value - Minimum value) / (Maximum value - Minimum value). For example, if a waybill has a dwell time of 8 hours at transit point A, and the minimum dwell time among all samples is 0.5 hours and the maximum is 24 hours, then the normalized value is (8-0.5) / (24-0.5) = 7.5 / 23.5 ≈ 0.32. This normalization process eliminates the difference in dimensions, making features at different time scales comparable in subsequent encoding.

[0051] Image features are derived from photos of express delivery labels, damaged or abnormal scenes. Different photos vary in resolution, lighting conditions, and shooting angles, and some photos may be blurry or lack meaningful content. During standardization preprocessing, images are first uniformly scaled to a fixed size (e.g., 224×224 pixels), then pixel values ​​are normalized (divided by 255 to map to the [0,1] interval), and blurry or meaningless images are marked with zero vector padding to ensure correct processing by the subsequent encoding model.

[0052] Structured features originate from discrete or continuous structured data such as package weight, volume, estimated delivery time, station type, and staff position. Different types of features require different encoding methods: For discrete features such as station type (values ​​include sorting center, transit point, and sales point), one-hot encoding is required, for example, encoding "transit point" as [0,1,0] (assuming there are three categories); for continuous features such as weight (2.5kg), Z-Score standardization is required, i.e., subtracting the mean and dividing by the standard deviation, the formula is: Standardized value = (Original value - Mean) / Standard deviation.

[0053] By unifying the original multimodal data with different formats and dimensions into a standardized format suitable for the input of their respective coding models, a data foundation is laid for subsequent feature encoding and fusion.

[0054] 302. Encode the standardized preprocessed multimodal features separately to obtain the feature vector corresponding to each feature; In this embodiment, the encoding process refers to converting the standardized preprocessed features of different types into fixed-dimensional feature vectors using their respective suitable encoding models, thereby mapping the original data to a unified semantic space. Since the data structures of text, trajectory, image, and structured features differ significantly, a single encoding model cannot effectively handle all types; therefore, specialized encoding models are required for feature extraction.

[0055] Specifically, for text features, the standardized preprocessed text sequence is input into the BERT pre-trained language model. The BERT model, based on the Transformer architecture, captures the contextual semantic information of words in the text through a bidirectional encoding mechanism. For example, for the preprocessed text sequence "processing of 24-hour delays during abnormal transit," the BERT model calculates the association weights between each word and other words through a multi-head attention mechanism, generating an embedding representation that incorporates contextual information, resulting in a 64-dimensional text feature vector.

[0056] For trajectory features, the normalized time-series data (including the dwell time sequence of each node and the transportation time sequence of adjacent nodes) is input into the LSTM (Long Short-Term Memory) recurrent neural network model. The LSTM model can effectively capture the long-short-term dependencies in the time-series data through its unique gating mechanism (forget gate, input gate, output gate) and output the corresponding trajectory feature vector.

[0057] For image features, the standardized preprocessed image is input into a ResNet50 convolutional neural network model to extract image feature vectors. For structured features, discrete features after one-hot encoding are concatenated with standardized continuous features. For example, site type is generated as a 3D vector after one-hot encoding, personnel position is generated as a 4D vector after one-hot encoding, and continuous features such as weight, volume, and expected timeliness are each 1D after Z-score standardization. These vectors are concatenated in sequence to form a multi-dimensional vector, which is then mapped to a 64-dimensional structured feature vector through a fully connected layer. The weight parameters of this fully connected layer are learned during model training, enabling the final structured feature vector to effectively represent the influence patterns of different combinations of structured features on anomaly occurrence.

[0058] By uniformly encoding multimodal features of different forms into numerical vectors of the same type, the model adaptation problem caused by differences in feature forms can be eliminated, providing a unified feature foundation for subsequent multi-feature fusion and anomaly calculation.

[0059] 303. A unified representation vector is generated by weighted fusion of multiple feature vectors through an attention mechanism; In this embodiment, multiple feature vectors are weighted and fused using an attention mechanism to generate a unified representation vector.

[0060] Specifically, the four feature vectors obtained from the aforementioned encoding process are used as input to the attention mechanism. The attention mechanism first calculates an attention score for each feature vector, which reflects the importance of that modality feature in the current waybill anomaly assessment. Specifically, each feature vector is linearly transformed using a learnable weight matrix, and then the transformed score is normalized using the Softmax function so that the sum of the four scores is 1. For example, for a waybill with a complete trajectory and a note explicitly mentioning "delay," the attention mechanism might calculate a trajectory feature weight of 0.45, a text feature weight of 0.30, a structured feature weight of 0.15, and an image feature weight of 0.10; while for another waybill with a blurry photo but a normal trajectory, the image feature weight might approach 0, while the text and trajectory features receive higher weights.

[0061] After obtaining the attention weights for each feature vector, the attention mechanism performs a weighted summation of the four feature vectors. The resulting fused feature vector integrates information from all four modalities, but the contribution of each dimension is modulated by the attention weights. For example, if the image feature weights are low, the contribution of image features to the fusion result will be suppressed. If the dimension of the fused feature vector does not match the input dimension (256 dimensions) of the subsequent spatiotemporal graph convolutional network, a further dimensionality transformation is performed. The fused feature vector is then input to a fully connected layer containing 256 neurons. Each neuron is fully connected to the 64-dimensional input vector through learnable weight parameters. Through this linear transformation, the 64-dimensional fused feature vector is mapped to a 256-dimensional unified representation vector.

[0062] By using an attention mechanism for weighted fusion, irrelevant features are avoided from interfering with model calculations, enabling the generated unified representation vector to more accurately capture the abnormal features of waybills and provide more targeted feature support for subsequent anomaly probability calculations.

[0063] 304. Input the unified representation vector and the logistics spatiotemporal graph into the spatiotemporal graph convolutional network to obtain the anomaly probability score of each waybill; In this embodiment, the Spatio-Temporal Graph Convolutional Network (ST-GCN) is a deep learning model capable of simultaneously processing graph-structured data (spatial dimension) and temporal data (temporal dimension). Specifically, a historical training set is constructed based on the logistics spatio-temporal graph of the entire flow of waybills, multimodal features, and corresponding actual result labels (divided into abnormal and normal samples). The initial spatio-temporal graph convolutional network is trained using this historical training set. During the training process, the network is further optimized using a cross-entropy loss function and the Adam optimizer to obtain the trained spatio-temporal graph convolutional network.

[0064] The unified representation vector and the logistics spatiotemporal graph are input into the spatiotemporal graph convolutional network. First, the logistics spatiotemporal graph is used as the topological input of the network to clarify the flow and relationship between express packages, stations, and personnel. Then, the unified representation vector is used as the feature input of each node to assign complete express package attribute information to each node. The network performs spatiotemporal feature extraction and convolution calculation according to three preset hidden layers (the number of neurons is 128, 64, and 32 respectively). The output value of the Sigmoid activation function is a value in the range [0,1], which is the abnormal probability score corresponding to each waybill. The score is positively correlated with the probability of express package abnormality.

[0065] By calculating the anomaly probability score using a unified representation vector and the aforementioned logistics spatiotemporal diagram, anomaly-related factors can be captured more comprehensively, thereby improving the comprehensiveness and accuracy of anomaly identification and providing a reliable quantitative basis for subsequent screening and batch processing of abnormal parcels.

[0066] Please see Figure 4 In this embodiment of the invention, querying the complete logistics trajectory sequence of each abnormal waybill, tracing the complete logistics trajectory sequence back in reverse chronological order and matching it with a preset stage identifier code, and determining the current logistics stage of each abnormal waybill based on the matching result includes: 401. Obtain the complete logistics trajectory sequence of each abnormal waybill in batches through the preset logistics trajectory interface. The complete logistics trajectory sequence includes multiple trajectory points arranged in chronological order, and each trajectory point includes the scan time and stage identifier code. In this embodiment, the preset logistics trajectory interface is a standardized interface provided by the logistics intranet for batch querying waybill trajectory data, used to ensure the standardization, batch processing, and timeliness of trajectory acquisition; the complete logistics trajectory sequence is a collection of all trajectory records of a single abnormal waybill from its creation to the current moment, consisting of several trajectory points sorted by scanning time, used to completely reconstruct the express delivery flow process; the trajectory point is the smallest unit of the logistics trajectory sequence, used to record the specific operational status of the express delivery at a certain moment and at a certain station; the stage identifier code is the preset scan type code (scanType), used to identify the current operation type and flow stage of the express delivery.

[0067] The complete logistics trajectory sequence of each abnormal waybill is obtained in batches through a preset logistics trajectory interface. Specifically, a POST request is used to call the internal logistics trajectory query interface, with the list of abnormal waybill numbers, request timestamp, and interface key as input parameters. A maximum of 1,000 waybill numbers can be retrieved in a single batch. The interface returns JSON format data, from which the trajectory list of each waybill is extracted and sorted from earliest to latest by scan time to form a continuous trajectory sequence. Each trajectory point includes scan time, scan site, scan type code (stage identifier code), scanner, scan device, and trajectory remarks. The returned data is verified for completeness, and waybills with empty trajectories or missing fields are removed and marked separately. At the same time, valid trajectory data is cached in Redis for 30 minutes to avoid repeated interface calls.

[0068] By using batch acquisition and structured parsing, the trajectory of multiple abnormal waybills can be collected in a single request, reducing the frequency of interface calls and improving data acquisition efficiency. At the same time, core fields such as scanning time and stage identifier code are fully preserved, providing a complete and standardized data foundation for subsequent trajectory backtracking and stage determination, and reducing judgment errors caused by missing trajectories or inconsistent formats.

[0069] 402. Starting from the trajectory point with the latest scanning time in the complete logistics trajectory sequence, traverse each trajectory point in reverse chronological order. In this embodiment, the trajectory point with the latest scanning time is the latest operation record in the complete logistics trajectory sequence, representing the latest status of the express delivery.

[0070] Starting from the track point with the latest scan time in the complete logistics trajectory sequence, the system traverses each track point backward in reverse chronological order. Specifically, the complete logistics trajectory sequence is first sorted and confirmed according to the scan time, and the track record with the largest time value is taken as the starting point of the traversal. Then, based on the starting point, the system reads the track points with earlier scan times in sequence and matches the preset stage identifier codes one by one. During the traversal, the system stops backtracking once a core scan type code is matched. If no valid code is matched when the system reaches the starting point of the sequence, the system is marked as having a trajectory abnormality.

[0071] By using reverse traversal, the latest operational status of the package can be prioritized, shortening the trajectory retrieval path and reducing the number of invalid traversals, thereby improving node positioning efficiency. At the same time, it can more accurately pinpoint the current stage of the package's flow, providing a reliable basis for subsequent problem package judgment.

[0072] 403. Match the stage identifier codes in the traversed trajectory points with a preset set of stage identifier codes. Stop traversing when any preset stage identifier code is matched. Determine the current logistics stage of the abnormal waybill based on the matched stage identifier code. In this embodiment, the stage identifier code is the scan type code, which is used to identify the business type of a single operation of the express shipment; the preset stage identifier code set is a pre-configured set of core scan codes used to determine the stage of the express shipment flow, and is set according to the categories of pickup, sorting, transit, delivery, signing, and return; the logistics stage is the flow link divided according to the latest valid operation of the express shipment, which is used to clarify the object and direction of handling problem shipments.

[0073] During the reverse backtracking process, the stage identifier code carried by each traversed trajectory point is compared with a preset set. As long as the stage identifier code belongs to a certain preset category (e.g., 203 belongs to the transit category), it is considered a successful match, and the traversal immediately stops, no longer continuing to backtrack. If multiple scan records exist at the same time, the one with the higher priority is used (priority order: signed for > delivered > transit > sorted > picked up > returned). For example, if both transit and delivery records exist, the delivery category is matched first. If no preset stage identifier code is matched after traversing the entire trajectory sequence, the waybill is marked as a "trajectory abnormal" stage.

[0074] Taking the aforementioned example waybill as an example, its reverse backtracking process is as follows: First, iterate through the latest trajectory point D (14:00:00, 301 in delivery). 301 belongs to the preset set of delivery categories. If a match is found, the backtracking stops immediately. Based on the matched stage identifier 301 (in delivery), combined with the condition that no signed receipt record was matched, it is determined that the waybill is currently in the "delivery stage". If the latest trajectory point of the waybill is 203 (transit delay), then a transit category is matched, and it is determined to be in the "transit stage"; if there is no match and the waybill was created more than 24 hours ago, it is determined to be in the "uncollected stage".

[0075] By using priority matching and a hit-and-stop traversal rule, the latest valid flow node of the package can be quickly located, avoiding the time-consuming process of full traversal. At the same time, by completing the stage judgment based on standardized code, the consistency and accuracy of stage division can be improved, providing stable support for subsequent receiving site matching and problem package type mapping.

[0076] Please see Figure 5 In this embodiment of the invention, determining the problem receiving station corresponding to each abnormal waybill based on the logistics stage and a preset station matching rule includes: 501. Obtain the latest scanning station in the complete logistics trajectory sequence corresponding to the abnormal waybill; In this embodiment, the latest scanning station refers to the station code or station name corresponding to the latest scanned trajectory point in the complete logistics trajectory sequence of the abnormal waybill. This station represents the current physical location of the waybill or the location of the last scanned operation, and is the core basis for determining the receiving station of the problematic shipment.

[0077] Specifically, the system first locates the track point with the latest scanning time (i.e., the starting point for reverse tracing), and then reads the scanning station information from that track point. For example, the complete logistics track sequence of an abnormal waybill, in chronological order, is as follows: Track point A (08:00:00, 001 pickup completed, pickup station "Chaoyang Business Point"), Track point B (10:00:00, 101 sorting completed, sorting center "East District Sorting Center"), and Track point C (12:00:00, 203 transit delay, transit center "South China Transit Center"). The track point with the latest scanning time is track point C, and its corresponding scanning station is "South China Transit Center". Therefore, the latest scanning station is "South China Transit Center".

[0078] By directly extracting the station corresponding to the latest trajectory point, the responsible station currently associated with the package can be quickly located, avoiding the time-consuming process of traversing the entire trajectory. At the same time, with the addition of validity verification, the accuracy and usability of the station information can be guaranteed, providing a true and reliable basic data for matching the receiving station of subsequent problematic packages.

[0079] 502. When the logistics stage is the sorting stage, the transit stage, or the delivery stage, if the latest scanning station has the authority to handle problem items, then the latest scanning station is determined as the problem receiving station; otherwise, the parent station of the latest scanning station is determined as the problem receiving station. In this embodiment, the logistics stage is the current flow link of the express shipment determined by the matched stage identifier code, including the sorting stage, transit stage, delivery stage, etc.; the problem receiving station is the responsible station for handling the abnormal waybill, and is used to receive and handle problem shipment tasks.

[0080] When the logistics stage is the sorting stage, transit stage, or delivery stage, the internal site management interface is called to query the latest scanned site's operational status and problem item handling permissions. If the query result indicates that the site has the permissions, the site is directly set as a problem receiving site. If the site does not have the permissions or is invalid, the upper-level sorting center or transit center configured for the site is automatically obtained, the upper-level site is identified as a problem receiving site, and the site adjustment log is recorded.

[0081] By verifying permissions first and then matching upwards step by step, it is possible to ensure that problematic items are accurately pushed to valid sites with processing capabilities, reducing processing delays caused by sites lacking permissions, while also improving the automation of problematic item allocation and reducing the cost of manual intervention.

[0082] 503. When the logistics stage is the non-collection stage, the sender's station of the abnormal waybill is identified as the problem receiving station; In this embodiment, when the logistics stage is the non-collection stage, the sender's address and sender's area code are extracted from the structured information of the abnormal waybill. Based on a preset address-site affiliation mapping relationship, the corresponding collection point is matched. After the site is verified as valid by the internal site management interface, it is designated as the problem receiving site. By directing non-collection problem shipments to the sender's site, the responsibility for problem handling is aligned with the responsibility for collection, facilitating direct contact with collection personnel to verify the cause and follow up on the handling, thus improving the efficiency and accuracy of handling non-collection issues.

[0083] 504. When the logistics stage is the return stage, the return destination station of the abnormal waybill shall be determined as the problem receiving station; In this embodiment, when the logistics stage is the return stage, the return address and return target code are extracted from the structured information and trajectory data of the abnormal waybill. The return destination station is obtained by matching the address with the station affiliation. The validity and operational status of the station are verified through the internal station management interface. If the verification is successful, the station is identified as the problem receiving station. By assigning return-related problem items to the return destination station, the processing responsibility is kept consistent with the target node of the return link. This facilitates the station to directly verify issues such as return delays and return anomalies, and helps improve the targeting and timeliness of return item processing.

[0084] Please see Figure 6 In this embodiment of the invention, generating problem item data based on the complete logistics trajectory sequence and the logistics stage includes: 601. Based on the preset stage and code mapping relationship, find the problem item type code corresponding to the logistics stage; In this embodiment, the preset stage-code mapping relationship refers to the corresponding rules for converting standardized logistics stages into standardized problem item type codes. Different logistics stages correspond to different handling requirements for abnormal issues. For example, transit delays require urging the transit process to speed up, delivery delays require urging the delivery person, and abnormal signatures require verifying the authenticity of the signature.

[0085] The system reads preset mapping rules: the transit / delay stage corresponds to code 01 (urging transit), the delivery / delivery delay stage corresponds to code 02 (urging delivery), the return / delay stage corresponds to code 03 (urging return), and the receipt stage with recipient feedback of non-receipt corresponds to code 04 (signed but not received). It compares each determined logistics stage with its mapping relationship, and outputs the corresponding problem item type code upon successful match. For example, if an abnormal waybill is determined to be in the "transit stage" through trajectory backtracking, the problem item type code is "01" (urging transit) based on the above mapping relationship. If another abnormal waybill is determined to be in the "delivery stage," the mapping code is "02" (urging delivery).

[0086] Automatic matching based on fixed mapping relationships can standardize and unify the correspondence between stages and problem types, reduce the bias caused by manual judgment, and improve the generation speed of type codes, providing standardized and directly usable field data for subsequent batch submission interface calls.

[0087] 602. Extract the latest scanning time, the latest scanning station, and key trajectory notes associated with the logistics stage from the complete logistics trajectory sequence; In this embodiment, the latest scan time is the operation time corresponding to the latest scan record in the trajectory sequence; the latest scan station is the operation station corresponding to the latest scan record in the trajectory sequence; and the key trajectory notes are trajectory description information directly related to the current logistics stage and used to explain abnormal states.

[0088] Specifically, first locate the trajectory point with the latest scanning time in the complete logistics trajectory sequence, and extract the latest scanning time and latest scanning site fields from it; then, based on the determined logistics stage, extract the notes related to delays, anomalies, and failures from the trajectory point or adjacent trajectory points as key trajectory notes; perform null value and validity checks on the extracted information, remove invalid characters and meaningless content, and ensure that the fields are usable.

[0089] By accurately extracting the latest time, site, and stage-related notes, the generated problem item description information can be complete and clearly pointed, making it easier for the processing site to quickly understand the abnormal scenario, reducing information verification costs, and improving the efficiency of problem item processing.

[0090] 603. Fill the waybill number of the abnormal waybill, the problem item type code, the latest scan time, the latest scan station, and the key trajectory remarks into the preset problem description template to generate a standardized problem description; In this embodiment, the preset problem description template is a pre-configured problem item description framework containing fixed field placeholders, used to ensure a uniform description format and complete information. Specifically, the corresponding type template preset by the system is called, and the waybill number, problem item type code, latest scan time, latest scan site, and key trajectory remarks are filled in sequentially according to the placeholder positions. Then, the filled content is format-validated and redundant characters are removed to ensure that the sentences are fluent and the information is unambiguous. The corresponding standardized description text is output according to four types: urging transit, urging delivery, urging return, and signed for but not received.

[0091] By automatically filling in standardized descriptions using templates, each issue item's description format becomes uniform, complete in elements, and clear in its direction. This reduces errors and time consumption associated with manual writing, facilitates rapid identification of anomalies at processing sites, and improves the efficiency of issue item circulation and processing.

[0092] 604. Assemble the standardized problem description, the problem receiving station, and the current timestamp into problem item data; In this embodiment, the current timestamp is the system time at the moment the problem shipment data is assembled, used to mark the submission time. Specifically, using a structured data format as the carrier, the standardized problem description, problem receiving station code, and current system timestamp are combined as required fields; simultaneously, the waybill number and problem shipment type code of the corresponding abnormal waybill are associated and bound to form a complete problem shipment data structure that conforms to the interface specification; the assembled data undergoes field integrity verification to ensure no missing or format errors. For example, the generated standardized problem description is extracted, and waybill number [YT20260410001] is currently in the [transit delay] stage (latest scan time: 2026-04-10 14:30:00). Please [WH003] process the transit process as soon as possible, speed up the shipment flow, and update the transit trajectory in a timely manner. The site code of the issue receiving site is WH003 (South China Foshan Chancheng Transit Center); obtain the current timestamp, 2026-04-10 14:35:22; assemble according to the JSON structure required by the interface, and integrate the above information with fields such as waybill number and issue item type code to form issue item data that can be directly submitted.

[0093] By automatically assembling key information according to a fixed structure, standardized problem data that can be directly identified by the interface can be generated, reducing manual assembly errors and improving the success rate and data accuracy of subsequent batch submissions.

[0094] Please see Figure 7 In this embodiment of the invention, the processing result includes processing status and processing details. The step of generating an anomaly processing information set based on the processing result, and then using a pre-trained causal inference model to identify the root causes of the anomaly processing information set to obtain an overall causal list, includes: 701. When the processing status is processing failure, obtain the processing details and extract the failure reason category from the processing details; In this embodiment, the processing status is based on the problematic shipment data and the abnormal shipment processing information obtained by the problematic receiving station. Specifically, the problematic shipment data is sent to the problematic receiving station, the station processes the shipment according to the problem description, and returns the processing information after processing. By matching the corresponding processing information according to the problematic shipment data, the processing result corresponding to each abnormal waybill can be obtained.

[0095] Specifically, the system first obtains the parcel processing status corresponding to the problematic parcel data through the problematic parcel processing result query interface. After determining that the processing has failed, it retrieves the complete processing details of the corresponding waybill. The processing details are structured text information including the failure reason category. According to the preset classification rules, it matches failure reason keywords such as signed for, not picked up, no track, invalid station, and processing timeout from the processing details, extracts and classifies them into the corresponding failure reason category, and performs standardized conversion on the extraction results to ensure that the category names are consistent and can be used for statistical analysis.

[0096] By automatically extracting and classifying the causes of failure, the core factors that caused the failure to process problematic items can be quickly identified, reducing the time spent on manual analysis of details, and providing a standardized data source for subsequent causal inference and process optimization.

[0097] 702. The abnormal probability score, logistics stage, failure reason category, and trajectory features in the complete logistics trajectory sequence corresponding to each abnormal waybill are fused to obtain the abnormal handling information set. In this embodiment, the waybill number is used as the unique primary key to match and aggregate the anomaly probability score (e.g., 0.87), logistics stage (e.g., transit stage), and failure reason category (e.g., processing timeout) corresponding to the same waybill. Then, trajectory features are extracted from the trajectory sequence of the waybill, including transit time, inter-station transportation time, and flow node order. The above continuous numerical features are normalized to map the original feature values ​​to a unified range of [0,1]. For example, the original transit time "6 hours" is mapped to 0.62 and the original transportation time "2.5 hours" is mapped to 0.31 to eliminate the calculation impact caused by the difference in the units of different features. Finally, the waybill identifier, anomaly probability score, logistics stage, failure reason category, and normalized trajectory features are spliced ​​and integrated according to a fixed structure to form an anomaly handling information set for root cause analysis.

[0098] By fusing multi-dimensional information by waybill primary key, the data dimensions can be complete and the correlation can be clear. This can comprehensively cover the key information of the entire chain from detection to processing of problematic items, providing comprehensive and standardized input data for subsequent causal inference models to identify root causes and calculate intervention effects.

[0099] 703. Input the set of anomaly handling information into a pre-trained causal inference model for causal analysis, and generate a list of root causes and the corresponding intervention effect values ​​for each root cause; In this embodiment, the intervention effect value is a quantitative indicator used to characterize the strength of the root cause's influence. An anomaly handling information set, including waybill status, trajectory features, and failure reasons (e.g., anomaly probability score of 0.87, logistics stage as transit stage, transit dwell time feature of 0.62, failure reason as processing timeout), is input into a trained Bayesian network causal inference model. The model calculates the degree of influence of each candidate root cause on the occurrence of the anomaly based on the conditional probability distribution learned from the training data. For example, model analysis reveals that under the condition of "transit yard sorting efficiency below the threshold," the probability of "transit dwell anomaly" increases from the baseline of 5% to 35%, a probability difference of 30 percentage points, indicating a strong causal relationship. The model calculates a causal relationship strength value for each candidate root cause, reflecting the closeness of the association between the root cause and the occurrence of the anomaly.

[0100] Secondly, the model outputs a list of root causes and the causal strength of each root cause. For example, the analysis results for a batch of abnormal transit delays show that the causal strength of root cause A (low sorting efficiency at transit point A) is 0.65, the causal strength of root cause B (overloading due to a surge in cargo volume) is 0.48, the causal strength of root cause C (sorting equipment malfunction) is 0.32, and the causal strength of root cause D (insufficient night shift staff) is 0.25.

[0101] Then, the model calculates the intervention effect for each root cause. The intervention effect refers to the magnitude of change in the anomaly incidence rate after altering the value of a root cause variable. The formula is: Intervention effect = P(anomaly | do(root cause = normal state)) - P(anomaly | root cause = current state), where the do operator represents active intervention rather than passive observation, P is the probability, and | is the probability symbol. The larger the absolute value of the intervention effect, the more significant the effect of intervening in that root cause on reducing the anomaly incidence rate, and the more critical that root cause.

[0102] By inputting complete anomaly information into a dedicated causal model for analysis, the core causes of anomalies can be objectively identified, rather than just performing correlation statistics. This facilitates the identification of key optimizable links and provides quantifiable and actionable data support for subsequent logistics process improvements and resource scheduling.

[0103] 704. Sort the root causes in the root cause list in descending order according to the intervention effect value, determine the root cause ranked first as the overall core cause, determine the root causes ranked second and third as the overall secondary causes, and output the overall core cause and the overall secondary cause as the overall cause list. In this embodiment, the root cause list is a set of all causes that lead to the anomaly of the problematic item, obtained through causal analysis; the overall core cause is the root cause with the highest impact and plays a dominant role in the occurrence of the anomaly; the overall secondary cause is the root cause with the next highest impact and can also significantly cause the anomaly; the overall cause list is a set of key causes for process optimization, composed of the core cause and the secondary cause.

[0104] Specifically, the intervention effect values ​​corresponding to each root cause in the root cause list are arranged from largest to smallest. For example, the root cause "low sorting efficiency in the transfer station" corresponds to an intervention effect value of 0.78, "saturated capacity of the transfer station" corresponds to 0.53, "insufficient on-duty rate of delivery personnel" corresponds to 0.41, and "congestion of transportation routes" corresponds to 0.29. After being arranged in descending order, the order is as follows: low sorting efficiency in the transfer station, saturated capacity of the transfer station, insufficient on-duty rate of delivery personnel, and congestion of transportation routes. The root cause ranked first is identified as the overall core cause, and the root causes ranked second and third are identified as the overall secondary causes. Root causes ranked third and below are eliminated. The overall core causes and overall secondary causes are combined in a fixed structure to form and output the overall cause list.

[0105] By sorting the intervention effect values ​​in descending order and classifying and screening the core and secondary causes, the most influential key factors can be highlighted, and interference from irrelevant factors can be avoided. This makes it easier for logistics operators to focus on high-value optimization points and accurately invest resources to reduce the occurrence rate of defective items.

[0106] The above describes the intelligent analysis method for logistics anomalies in the embodiments of the present invention. The following describes the intelligent analysis device for logistics anomalies in the embodiments of the present invention. Please refer to [link / reference]. Figure 8 One embodiment of the intelligent analysis device for logistics anomalies in this invention includes: The calculation module 801 is used to construct a logistics spatiotemporal graph based on logistics information, extract multimodal features of each waybill in the logistics spatiotemporal graph, and calculate the anomaly probability based on the logistics spatiotemporal graph and the multimodal features through a preset spatiotemporal graph convolutional network to obtain the anomaly probability score of each waybill. The filtering module 802 is used to filter out abnormal waybills based on the abnormal probability score, query the complete logistics trajectory sequence of each abnormal waybill, trace back the complete logistics trajectory sequence in reverse time and match it with a preset stage identifier code, and determine the current logistics stage of each abnormal waybill based on the matching result. The sending module 803 is used to determine the problem receiving station corresponding to each abnormal waybill according to the logistics stage and a preset station matching rule, generate problem item data based on the complete logistics trajectory sequence and the logistics stage, and call a preset problem interface to send the problem item data to the corresponding problem receiving station. The identification module 804 is used to receive the processing result of each abnormal waybill returned by the problem receiving station, generate an abnormal processing information set based on the processing result, and perform root cause identification on the abnormal processing information set through a pre-trained causal inference model to obtain an overall cause list.

[0107] Based on the same ideas as the methods in the above embodiments, the apparatus provided by the present invention can implement the methods in the above embodiments.

[0108] The above combination Figure 8 The intelligent analysis device for logistics anomalies in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The intelligent analysis device for logistics anomalies in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0109] Figure 9 This is a schematic diagram of the structure of an intelligent analysis device for logistics anomalies provided in an embodiment of the present invention. The intelligent analysis device 900 for logistics anomalies can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more storage media 930 (e.g., one or more mass storage devices) for storing application programs 933 or data 932. The memory 920 and storage media 930 can be temporary or persistent storage. The program stored in the storage media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the intelligent analysis device 900 for logistics anomalies. Furthermore, the processor 910 may be configured to communicate with the storage media 930 and execute the series of instruction operations in the storage media 930 on the intelligent analysis device 900 for logistics anomalies to implement the steps of the intelligent analysis method for logistics anomalies provided in the above-described method embodiments.

[0110] The intelligent analysis device 900 for logistics anomalies may also include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 9 The illustrated intelligent analysis device structure for logistics anomalies does not constitute a limitation on the intelligent analysis device for logistics anomalies. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0111] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of an intelligent analysis method for logistics anomalies.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0113] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0114] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent analysis method for logistics anomalies, characterized in that, include: A logistics spatiotemporal graph is constructed based on logistics information. Multimodal features of each waybill in the logistics spatiotemporal graph are extracted. Based on the logistics spatiotemporal graph and the multimodal features, anomaly probability is calculated through a pre-trained spatiotemporal graph convolutional network to obtain anomaly probability score for each waybill. Abnormal waybills are selected based on the abnormal probability score. The complete logistics trajectory sequence of each abnormal waybill is queried. The complete logistics trajectory sequence is traced back in reverse time and matched with the preset stage identifier code. The current logistics stage of each abnormal waybill is determined based on the matching result. Based on the logistics stage, the problem receiving station corresponding to each abnormal waybill is determined by a preset station matching rule, and problem shipment data is generated based on the complete logistics trajectory sequence and the logistics stage. Based on the problematic shipment data and the problematic receiving station, the processing result corresponding to each abnormal waybill is obtained. An abnormal processing information set is generated according to the processing result. The root cause is identified in the abnormal processing information set through a pre-trained causal inference model to obtain an overall cause list.

2. The intelligent analysis method for logistics anomalies according to claim 1, characterized in that, The construction of a logistics spatiotemporal map based on logistics information includes: Based on the site information, personnel information, and waybill information in the logistics information, generate a node set including site nodes, personnel nodes, and waybill nodes; Based on the logistics flow relationships in the logistics information, establish transportation edges connecting two station nodes, membership edges connecting personnel nodes and station nodes, and flow edges connecting waybill nodes and station nodes or personnel nodes, and generate an edge set. Configure a unique identifier and basic attributes for each node in the node set, and configure a timestamp and association attributes for each edge in the edge set; Based on the configured set of nodes and set of edges, the logistics spatiotemporal graph is constructed.

3. The intelligent analysis method for logistics anomalies according to claim 1, characterized in that, Based on the logistics spatiotemporal graph and the multimodal features, anomaly probability calculation is performed using a pre-trained spatiotemporal graph convolutional network to obtain an anomaly probability score for each waybill, including: The multimodal features of each waybill are standardized and preprocessed, including text features, trajectory features, image features, and structured features; The standardized preprocessed multimodal features are encoded separately to obtain the feature vector corresponding to each feature; Multiple feature vectors are weighted and fused using an attention mechanism to generate a unified representation vector; The unified representation vector and the logistics spatiotemporal graph are input together into the spatiotemporal graph convolutional network to obtain the anomaly probability score for each waybill.

4. The intelligent analysis method for logistics anomalies according to claim 1, characterized in that, The process involves querying the complete logistics trajectory sequence for each abnormal waybill, tracing the complete logistics trajectory sequence backward in reverse chronological order, and matching it with a preset stage identifier code. Based on the matching result, the current logistics stage of each abnormal waybill is determined, including: The complete logistics trajectory sequence of each abnormal waybill is obtained in batches through a preset logistics trajectory interface. The complete logistics trajectory sequence includes multiple trajectory points arranged in chronological order, and each trajectory point includes the scan time and stage identifier code. Starting from the trajectory point with the latest scanning time in the complete logistics trajectory sequence, traverse each trajectory point in reverse chronological order. The stage identifier codes in the traversed trajectory points are matched with a preset set of stage identifier codes. The traversal stops when any preset stage identifier code is matched, and the current logistics stage of the abnormal waybill is determined based on the matched stage identifier code.

5. The intelligent analysis method for logistics anomalies according to claim 1, characterized in that, The step of determining the problem receiving station corresponding to each abnormal waybill based on the logistics stage and through preset station matching rules includes: Obtain the latest scanning station in the complete logistics trajectory sequence corresponding to the abnormal waybill; When the logistics stage is the sorting stage, the transit stage, or the delivery stage, if the latest scanning station has the authority to handle problem items, then the latest scanning station is determined as the problem receiving station; otherwise, the parent station of the latest scanning station is determined as the problem receiving station. When the logistics stage is the non-pickup stage, the sender's station of the abnormal waybill is identified as the problem receiving station; When the logistics stage is the return stage, the return destination station of the abnormal waybill is determined as the problem receiving station.

6. The intelligent analysis method for logistics anomalies according to claim 1, characterized in that, The generation of problem item data based on the complete logistics trajectory sequence and the logistics stage includes: Based on the preset stage and code mapping relationship, find the problem item type code corresponding to the logistics stage; Extract the latest scanning time, the latest scanning station, and key trajectory notes associated with the logistics stage from the complete logistics trajectory sequence; Fill the waybill number of the abnormal waybill, the problem item type code, the latest scan time, the latest scan station, and the key trajectory remarks into the preset problem description template to generate a standardized problem description; The standardized problem description, the problem receiving station, and the current timestamp are assembled into problem item data.

7. The intelligent analysis method for logistics anomalies according to claim 1, characterized in that, The processing result includes processing status and processing details. An anomaly processing information set is generated based on the processing result. A pre-trained causal inference model is used to identify the root causes of the anomaly processing information set to obtain an overall list of causes, including: When the processing status is processing failure, obtain the processing details and extract the failure reason category from the processing details; The abnormality probability score, logistics stage, failure reason category, and trajectory features in the complete logistics trajectory sequence corresponding to each abnormal waybill are fused to obtain the abnormality handling information set. The anomaly handling information set is input into a pre-trained causal inference model for causal analysis, generating a root cause list and the corresponding intervention effect value for each root cause. Based on the intervention effect value, the root causes in the root cause list are sorted in descending order. The root cause ranked first is determined as the overall core cause, and the root causes ranked second and third are determined as the overall secondary causes. The overall core cause and the overall secondary cause are output as the overall cause list.

8. An intelligent analysis device for logistics anomalies, characterized in that, include: The calculation module is used to construct a logistics spatiotemporal graph based on logistics information, extract multimodal features of each waybill in the logistics spatiotemporal graph, and calculate the anomaly probability based on the logistics spatiotemporal graph and the multimodal features through a preset spatiotemporal graph convolutional network to obtain the anomaly probability score of each waybill. The filtering module is used to filter out abnormal waybills based on the abnormal probability score, query the complete logistics trajectory sequence of each abnormal waybill, trace the complete logistics trajectory sequence back in reverse time and match it with a preset stage identifier code, and determine the current logistics stage of each abnormal waybill based on the matching result. The sending module is used to determine the problem receiving station corresponding to each abnormal waybill according to the logistics stage and a preset station matching rule, generate problem item data based on the complete logistics trajectory sequence and the logistics stage, and call a preset problem interface to send the problem item data to the corresponding problem receiving station. The identification module is used to receive the processing result of each abnormal waybill returned by the problem receiving station, generate an abnormal processing information set based on the processing result, and perform root cause identification on the abnormal processing information set through a pre-trained causal inference model to obtain an overall cause list.

9. An intelligent analysis device for logistics anomalies, characterized in that, The intelligent analysis device for logistics anomalies includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the intelligent analysis device for logistics anomalies to perform the steps of the intelligent analysis method for logistics anomalies as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the steps of the intelligent analysis method for logistics anomalies as described in any one of claims 1-7.

Images