Logistics supply chain dynamic risk identification method based on large language model

Abstract

The invention belongs to the technical field of logistics supply chain management, and discloses a logistics supply chain dynamic risk identification method based on a large language model. A real-time heterogeneous data stream is subjected to space-time normalization processing through a dynamic sliding window mechanism, a three-dimensional space-time data tensor set with entity types, timestamps and space grid codes as dimensions is constructed, dynamic entities in a logistics supply chain and the incidence relation of the dynamic entities are recognized through an entity-relation-time triple extraction module, and the real-time heterogeneous data stream is obtained. Constructing a dynamic knowledge graph with space-time attributes; generating an incremental graph version according to the change event of the entity state, performing multi-dimensional anomaly detection in combination with a corresponding graph change log, and generating a structured risk tag; when a risk event occurs, the time-space coordinates of the root cause of the risk event are accurately positioned through a version backtracking function. The real-time performance, the accuracy and the interpretability of supply chain risk identification are remarkably improved, and a systematic solution is provided for dynamic risk management of a complex logistics network.

Description

Technical Field

[0001] This invention relates to the field of logistics supply chain management technology, and in particular to a method for dynamic risk identification in logistics supply chains based on a large language model. Background Technology

[0002] Logistics supply chain risk management is a core component in ensuring the efficient and secure operation of the supply chain. Traditional methods rely primarily on human experience, static rules, or single data sources for analysis, making it difficult to handle dynamically changing spatiotemporal data and complex entity relationships. In the face of increasingly complex global supply chain networks, how to dynamically and efficiently identify risks in the logistics supply chain has become a pressing issue.

[0003] Modern logistics supply chains encompass heterogeneous data from multiple sources, including IoT sensors, transportation routes, and market sentiment. These data differ significantly in format, frequency, and semantics, making effective integration difficult with traditional methods. Furthermore, these methods are prone to failure due to changes in data sources, leading to information silos. Simultaneously, traditional risk detection models rely on fixed time windows, making them unable to adapt to sudden fluctuations in data flow. Moreover, timestamp discrepancies and spatial mapping errors in multi-source data are not effectively corrected, resulting in distorted analysis results.

[0004] Existing knowledge graph construction largely relies on experts manually labeling entity relationships, resulting in long update cycles and difficulty in reflecting dynamic changes in the supply chain in a timely manner. Static graphs also cannot automatically link historical events with real-time data, making it difficult to predict systemic risks. Furthermore, traditional machine learning models are prone to catastrophic forgetting during iterative updates, thereby reducing the accuracy of long-term risk predictions.

[0005] In logistics and supply chain management, dynamic risk identification is a key link in ensuring the efficient and safe operation of the entire chain. With the increasing complexity of global supply chain networks and the acceleration of digital transformation, overcoming the bottlenecks of traditional technologies in data fusion, real-time response, and knowledge evolution can provide accurate and agile risk management support for complex supply chain networks. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a dynamic risk identification method for the logistics supply chain based on a large language model. The method dynamically integrates multi-source data to construct a spatiotemporally aligned three-dimensional tensor, extracts entity relationships using a large language model to generate an incremental spatiotemporal knowledge graph, generates risk labels based on graph anomaly detection, and finally optimizes the system through a feedback mechanism.

[0007] The technical solution of this invention is as follows: A method for dynamic risk identification in logistics supply chains based on a large language model, comprising the following steps:

[0008] Access real-time heterogeneous data streams covering the entire logistics supply chain, and perform spatiotemporal normalization processing on the real-time heterogeneous data streams through a dynamic sliding window mechanism to construct a three-dimensional spatiotemporal data tensor set with entity type, timestamp, and spatial grid encoding as dimensions;

[0009] The three-dimensional spatiotemporal data tensor set is input into a pre-trained large language model. Through the entity-relationship-time triple extraction module, dynamic entities and their relationships in the logistics supply chain are identified, a dynamic knowledge graph with spatiotemporal attributes is constructed, and an incremental graph version is generated based on the entity state change events. The corresponding graph change log is stored in the temporal graph database.

[0010] Based on incremental graph versions, multi-dimensional anomaly detection is performed in conjunction with the corresponding graph change logs to generate structured risk labels; when a risk event occurs, the spatiotemporal coordinates of the root cause of the risk event are accurately located through the version backtracking function of the temporal graph database.

[0011] The dynamic sliding window mechanism specifically includes:

[0012] The time window is dynamically adjusted according to the update frequency of the real-time heterogeneous data stream: a short window is used for high-frequency data and a long window is used for low-frequency data; when the timeliness difference of data within the time window exceeds a preset threshold, the time window is automatically split into multiple sub-windows, and each sub-window corresponds to a data source type.

[0013] The spatial window is dynamically adjusted, with a preset 6-bit Geohash encoding. The system calculates the entity density and real-time heterogeneous data stream update frequency within the geographic grid cell in real time: when the entity density exceeds the preset density threshold or the update frequency is greater than the preset update frequency threshold, the geographic grid is expanded to 7-bit precision; when the entity density is less than 20% of the preset density threshold, the geographic grid is reduced to 5-bit precision.

[0014] The spatiotemporal normalization process specifically includes:

[0015] Time dimension: The timestamps of all access devices are calibrated through the network time protocol to eliminate time asynchrony problems caused by device clock deviation and ensure that the order of events is accurate.

[0016] Spatial dimension: The geographic coordinate system is uniformly converted to Geohash encoding; for encrypted or biased coordinates, compensation and correction are performed to ensure the consistency and accuracy of spatial location.

[0017] The set of three-dimensional spatiotemporal data tensors specifically includes:

[0018] In terms of entity type, each channel represents a specific supply chain entity type, which is coded and divided according to predefined entity types. The entity types come from the accessed real-time heterogeneous data stream. The specific channel design is as follows: Channel 1 represents the inventory of all nodes, Channel 2 represents the location of transport vehicles in transit, and Channel 3 represents the intensity of public opinion on port congestion extracted from news.

[0019] In terms of time window dimension, the timestamps of all real-time heterogeneous data streams are first calibrated through the network time protocol to eliminate device clock deviation; the dynamically adjusted time window is divided into a fixed number of continuous time slices, each time slice representing a time snapshot, which is used to capture the evolution pattern of entity state in time series.

[0020] Spatial grid dimension: The spatial grid dimension is divided according to Geohash encoding. By dynamically adjusting the Geohash precision, fine-grained monitoring of key areas and coarse-grained monitoring of non-critical areas can be achieved.

[0021] Each element value in a three-dimensional spatiotemporal data tensor Recorded at The specific state of the i-th entity type within the k-th time slice and the k-th spatial grid cell.

[0022] The construction of a dynamic knowledge graph with spatiotemporal attributes specifically includes:

[0023] Through prompting engineering, the pre-trained large language model is guided to understand the intrinsic meaning of different dimensions of data in the three-dimensional spatiotemporal data tensor and to perform entity-relation-time triple extraction task.

[0024] Entity node: Each entity includes inherent attributes and time-varying attributes, wherein the time-varying attributes are dynamically associated with the latest timestamp and spatial coordinates;

[0025] Relationship edge: Each relationship edge is associated with a timestamp, indicating the time or time period when the relationship edge was established, as well as the geographical location where the relationship occurred.

[0026] The prompt project is a predefined structured prompt template; the structured prompt template contains the following instructions:

[0027] Semantic parsing instructions for the three dimensions of entity type, time slice, and spatial grid in a three-dimensional spatiotemporal data tensor;

[0028] The pre-trained large language model is required to analyze data according to the reasoning steps of the thought chain to identify the instructions of entities and relationships;

[0029] The constraint output is a standardized JSON format instruction containing the subject, relation, object, relation timestamp, and relation spatial coordinates.

[0030] The generation of the incremental map version specifically includes:

[0031] When a pre-trained large language model recognizes an entity state change from a three-dimensional spatiotemporal data tensor generated by a newly accessed real-time heterogeneous data stream, it does not modify the current knowledge graph version, but instead generates a lightweight difference file.

[0032] The difference file records the changes, including: the unique ID of the object being changed, the changed attribute key-value pairs, the change timestamp, and the event type that triggered the change;

[0033] Monitor the event occurrence rate at the entity level. When the event occurrence rate of an entity exceeds a dynamic threshold calculated based on its historical behavior within a preset time window, aggregate multiple change events of that entity within that window to generate a summary change record.

[0034] The version backtracking function of the temporal graph database is specifically implemented as follows:

[0035] Storage and Indexing: Each difference file is associated with a unique time range and records the parent-child dependencies between it and adjacent difference files, forming a version chain;

[0036] Time travel query: When a risk event is detected, analysts input the time and location of the event; by tracing back along the version chain, a dynamic knowledge graph snapshot of the spatiotemporal attributes at the moment the risk occurred is reconstructed;

[0037] Risk diffusion path construction: Based on the backtracking knowledge graph snapshot, and combined with relationship changes and spatial displacements in the change log, a topological diffusion path graph of risk events is constructed.

[0038] The topological diffusion path diagram of the risk event is constructed as follows:

[0039] Starting with the risk source points in the knowledge graph snapshot obtained through backtracking, scan the graph change log;

[0040] Based on the chronological order of event timestamps and the spatial proximity of the event locations recorded in the map change log, multiple discrete entity state change events are connected into a propagation chain with causal and spatiotemporal order.

[0041] The propagation chain is visualized on a graph structure to generate a topological diffusion path diagram of the risk event.

[0042] The anomaly detection specifically includes:

[0043] If the timestamp is abnormal, analyze whether the interval of the entity status update timestamp meets the preset business time interval threshold, or whether the timeliness has expired.

[0044] If spatial coordinates are abnormal, check whether the coordinates of the transportation path deviate from the preset geographic grid, or whether the entity density within a unit grid exceeds the preset density threshold.

[0045] Map change anomalies are identified by counting the frequency of attribute changes for the same entity in consecutive incremental map versions. When the frequency of attribute changes exceeds a preset multiple K of the historical average frequency, or exceeds an adaptive threshold calculated based on the historical fluctuation range, it is marked as an anomaly.

[0046] If the timestamp and spatial coordinate attributes conform to the preset rules, and there are no abnormal events in the map change record, then no risk label will be output.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] By employing a dynamic sliding window mechanism and 3D tensor construction, this approach addresses the challenges of spatiotemporal alignment and real-time fusion of multi-source heterogeneous data using traditional methods, effectively responding to sudden fluctuations in data flow and changes in spatial density. Combining large language model semantic understanding with an incremental knowledge graph update mechanism, it overcomes the limitations of traditional static knowledge graph construction and long update cycles, constructing a dynamic knowledge graph with spatiotemporal attributes and version backtracking capabilities. Based on the incremental knowledge graph and multi-dimensional anomaly detection rules, it improves the accuracy and real-time performance of risk detection, achieving structured expression of risk labels and closed-loop optimization of detection, feedback, and tuning. Through dynamic threshold adjustment and domain adaptation mechanisms, it enhances the system's adaptability in complex scenarios, expanding its application scope to multilingual and multi-business supply chain scenarios. This solution addresses core issues of traditional technologies, such as inefficient multi-source data fusion, lagging risk detection, and rigid knowledge updates, achieving real-time, accurate, and adaptive optimization capabilities for logistics supply chain risk identification, providing technical support for the stable operation of complex supply chain networks. Attached Figure Description

[0049] Figure 1 This is an architecture diagram of the dynamic risk identification method for logistics supply chain based on a large language model used in embodiments of the present invention;

[0050] Figure 2 This is a system flowchart of the dynamic risk identification method for logistics supply chain based on a large language model used in an embodiment of the present invention.

[0051] Figure 3 The flowchart illustrates the dynamic risk anomaly detection and precise location process used in embodiments of the present invention. Detailed Implementation

[0052] To make the technical solution, objectives, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the provided embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0053] The following is a detailed description of a dynamic risk identification method for logistics supply chain based on a large language model, which aims to help understand the technical solution and operation steps of this invention.

[0054] See Figure 1 The method includes the following steps:

[0055] S110. Access real-time heterogeneous data streams covering the entire logistics supply chain, and perform spatiotemporal normalization processing on the real-time heterogeneous data streams through a dynamic sliding window mechanism to construct a three-dimensional spatiotemporal data tensor set with entity type, timestamp, and spatial grid encoding as dimensions.

[0056] S120. Input the three-dimensional spatiotemporal data tensor set into the pre-trained large language model. Through the entity-relationship-time triplet extraction module, identify dynamic entities and their relationships in the logistics supply chain, construct a dynamic knowledge graph with spatiotemporal attributes, and generate incremental graph versions based on entity state change events. The corresponding graph change logs are stored in the temporal graph database. Dynamic entities include transport vehicles and warehouse nodes; relationships include "transport vehicle A is located in warehouse B".

[0057] S130. Based on incremental graph versions, multi-dimensional anomaly detection is performed in conjunction with the corresponding graph change logs to generate structured risk labels. When a risk event occurs, the spatiotemporal coordinates of the root cause of the risk event are accurately located through the version backtracking function of the temporal graph database.

[0058] In this embodiment, the detailed implementation steps of step S110 include:

[0059] S111, Data Access: Access real-time data sources across the entire logistics supply chain through standardized data interfaces. The entity types of the real-time data sources include: IoT sensor time-series data, structured data from enterprise resource planning systems, spatial path coordinates and market sentiment text data from transportation management systems, market sentiment monitoring platform data, operational indicator data of supply chain node enterprises, order status, warehouse inventory, transportation path coordinates, and environmental events. Specifically, the system collects real-time sensor data such as temperature, humidity, and vibration via the MQTT protocol, with timestamps accurate to the millisecond level, and the data format is JSON. Order status, inventory, and supplier information are synchronized via a RESTful API, with the data format being XML. The real-time location of transportation vehicles is recorded in WGS-84 latitude and longitude format, updated every 10 seconds. Text data from social media and news platforms is crawled using the Scrapy framework, and sentiment analysis and keyword extraction are performed using the BERT model to output structured data. Information such as natural disaster warnings is obtained by accessing the meteorological bureau's API, in GeoJSON format, including event type, impact range, and occurrence time. All data sources are transmitted in real-time via an Apache Kafka message queue, with 8 partitions and a replication factor of 3 to ensure a high throughput of ≥100,000 messages / second and a low latency of ≤100ms. Data is stored categorized by topic.

[0060] S112. Time-Sensitive Implementation: Time Dimension Adjustment, assuming the data source update frequency is... The unit is times per second, and the time window length is... and sliding step size The dynamic calculation formula is:

[0061]

[0062]

[0063] For high-frequency data with an update frequency of ≥1 time / second, a short window with a window length of 300 seconds is set to ensure real-time performance, and a sliding step of 60 seconds is used to avoid data overlap and redundancy; for low-frequency data with an update frequency of ≤1 time / hour, the window length is extended to 7200 seconds, and the sliding step is 1440 seconds to balance computational efficiency and data coverage; a coefficient of 0.2 ensures that 80% of the historical data continuity is retained when the window slides.

[0064] Spatial dimension adjustment is achieved using 6-bit Geohash encoding, corresponding to a geographic grid cell with an accuracy of approximately 1.2km × 0.6km, dynamically calculating the entity density within the grid. The formula for calculating the entity density is:

[0065]

[0066] like If the threshold is reached, the window is narrowed down to a 3-bit Geohash code, corresponding to a grid cell with a precision of approximately 156km × 156km; if If the threshold is exceeded, the grid size is expanded to approximately 4.8m × 4.8m, corresponding to a 9-bit Geohash precision.

[0067] Spatiotemporal normalization is performed, and timestamp discrepancies between multi-source data are aligned using a linear interpolation algorithm. The calculation formula for the linear interpolation algorithm is as follows:

[0068]

[0069] Data loss caused by transmission delay is filled by linear interpolation between adjacent time points;

[0070] Affine transformation matrix is ​​used to correct coordinate system deviations, eliminate coordinate offsets from different data sources, and ensure spatial consistency. The calculation formula is as follows:

[0071]

[0072] The parameters a, b, c, d, e, and f are calculated by fitting the coordinates of the reference point using the least squares method. This method can dynamically optimize the window parameters according to the data characteristics, reduce resource consumption, and ensure the consistency of multi-source data through spatiotemporal alignment, providing reliable input for subsequent risk identification. It is the key technological foundation for achieving real-time and accurate risk identification in this invention.

[0073] S113, 3D Tensor Generation: Encode the preprocessed data into 3D tensors according to the following dimensions. The entity type dimension (C) divides channels according to entity type encoding, with each channel corresponding to one entity category, using One-Hot encoding; the time window dimension (T) divides a predetermined number of continuous time slices according to the time series slices within the sliding window; the spatial grid dimension (S) maps the geographical area covered by Geohash encoding to spatial grid cells; the tensor element filling rule is: for numerical parameters, their original values ​​are directly recorded; if outliers are detected, Otherwise, it is 0.

[0074] See Figure 2 In this embodiment, the detailed implementation steps of step S120 include:

[0075] S121. Triple Extraction Driven by Large Language Model: The preprocessed 3D tensor set is input into a pre-trained Large Language Model (LLM). This embodiment uses a pre-trained model based on the Transformer architecture, leveraging its semantic understanding capabilities to parse entities, relations, and spatiotemporal attributes within the tensor. Specifically, a multi-head attention mechanism is used to capture potential associations between entities, calculated using the following formula:

[0076]

[0077] Q, K, and V represent the query, key, and value matrices, respectively, and their dimensions are... =64; Used to scale the dot product and prevent gradient vanishing.

[0078] Then, by combining timestamps and spatial grid encoding, entity, relation, and time triples are generated.

[0079] S122. Construction of a Spatiotemporally Sensitive Knowledge Graph: A dynamic knowledge graph is constructed based on extracted triples. Each entity node and relation edge is embedded with a timestamp and spatial coordinates. The graph is stored using a temporal graph database and supports version snapshot functionality. When a node state change is detected, the system automatically generates an incremental graph version, retains the old version, and records a change tracing log, including the trigger event type, scope of impact, data source hash value, and spatial displacement.

[0080] S123. Continuous Learning Framework Design: To balance new and old knowledge, the general semantic representation layer of the large language model is frozen, and only the parameter update permissions of the supply chain adaptation layer are opened. The optimizer adopts AdamW, and the learning rate is set to... The weight decay coefficient is 0.01.

[0081] The joint loss function is designed and its calculation formula is as follows:

[0082]

[0083] in, Cross-entropy loss is used to calculate the difference between the predicted risk label and the true label. The calculation formula is as follows:

[0084]

[0085] in, For real labels, To predict probabilities;

[0086] Used to measure parameters of historical tasks The importance of this is emphasized to ensure that important parameters are not overwritten during iteration. The calculation formula is as follows:

[0087]

[0088] This represents the number of historical task samples.

[0089] Let be the entropy weights of the current task parameters. The uncertainty of the current task parameters is estimated through Monte Carlo sampling, and its calculation formula is as follows:

[0090]

[0091] This is the task matching coefficient; This represents the dynamic elasticity coefficient that decays over time. The Fisher information matrix of historical tasks is updated every 24 hours. Write to disk and update as regularization constraint parameters during training for new tasks to prevent catastrophic forgetting.

[0092] See Figure 3 In this embodiment, the detailed implementation steps of step S130 include:

[0093] S131, Multi-dimensional anomaly detection:

[0094] Timestamp anomaly detection: Define the threshold for entity state update interval The calculation formula is as follows:

[0095]

[0096] based on The principle is to cover 99.7% of normally distributed data to reduce false alarms; if If so, it is marked as "information update lag";

[0097] Spatial coordinate anomaly detection: Calculate the Haversine distance between the transport path and the preset geofence, using the following formula:

[0098]

[0099] in, =6371km is the Earth's radius; , Both are the preset center coordinates of the fence. Indicates the latitude of the center of the fence. Indicates the longitude of the center of the fence; Both are the real-time coordinates of the transportation vehicle. Indicates the real-time latitude of the transportation vehicle. Indicates the real-time longitude of the means of transport; if If the distance is greater than 500m, an alarm will be triggered;

[0100] Anomaly detection of map changes: Count the number of changes to entity e within the time window [tk,t]. ,like,

[0101]

[0102] If it is a "abnormal fluctuation", it is marked as "abnormal". If at least two abnormal conditions are met at the same time, it is marked as "compound rule abnormality".

[0103] S132, Feedback-driven closed-loop optimization:

[0104] Based on the false alarm rate during manual review The timestamp threshold is dynamically adjusted, and the calculation formula is as follows:

[0105]

[0106] in, =0.1 is the adjustment coefficient. If This reduces detection sensitivity; when a risk label is manually rejected, adversarial examples are generated using the fast gradient sign method, calculated as follows:

[0107]

[0108] =0.01 is the perturbation amplitude, used to control the visibility of adversarial examples; The gradient of the loss function with respect to the input data;

[0109] If the number of supply chain nodes If the growth exceeds 20%, the version retention period will be adjusted. Adjustments will be made to balance the costs of tracing and storing historical data. The adjustment formula is as follows:

[0110]

[0111] An implementation device for a dynamic risk identification method in logistics supply chain based on a large language model includes the following three modules:

[0112] Data access and preprocessing module: Accesses real-time data sources across the entire logistics supply chain through standardized data interfaces. Based on a dynamic sliding window mechanism, it dynamically adjusts the window size and sliding step according to the timestamp interval and spatial distribution density of the data stream to perform spatiotemporal alignment, and performs structured processing on heterogeneous data streams to construct a three-dimensional tensor set containing entity type, timestamp, and spatial grid encoding.

[0113] Large Language Model (LLM) and Knowledge Graph Construction Module: Input the pre-processed three-dimensional tensor set into the pre-trained large language model, and identify dynamic entities and their relationships in the supply chain through the entity-relationship-time triplet extraction module. Construct a spatiotemporally sensitive knowledge graph, and generate incremental graph versions based on entity state change events. Design a continuous learning model fine-tuning framework to achieve adaptive balance between new and old knowledge.

[0114] Risk identification and optimization module: Anomaly detection is performed based on the incremental map version generated according to entity state change events. Risk labels are generated based on the anomaly detection results. The model loss function weights and map update thresholds are dynamically adjusted through a feedback-driven mechanism to form a closed-loop optimization system.

[0115] In the above technical solution, the dynamic sliding window mechanism in the data access and preprocessing module specifically includes:

[0116] In terms of time, the time window range is adaptively narrowed or expanded based on the update frequency of the data source;

[0117] In terms of spatial dimension, the preset level is 6-bit Geohash encoding, corresponding to a geographic grid unit of approximately 1.2km × 0.6km. The number of entities and data update frequency within a unit grid are calculated. If the preset density threshold is exceeded, the window range is reduced; otherwise, the window is expanded.

[0118] Spatiotemporal normalization within a sliding window eliminates temporal bias and spatial mapping errors in multi-source data;

[0119] The construction of the 3D tensor set, which includes entity type, timestamp, and spatial mesh encoding, specifically includes:

[0120] In the entity type dimension, channels are divided according to the entity type code, and each channel corresponds to an entity category;

[0121] The time window dimension divides a predetermined number of continuous time slices according to the time series slices within the sliding window;

[0122] Spatial grid dimension, mapped to spatial grid units based on the geographical area covered by Geohash encoding;

[0123] Tensor element values ​​record numerical state parameters or binary existence identifiers of the corresponding entity type within a specified time slice or spatial grid cell.

[0124] In the above technical solution, the specific implementation of constructing a spatiotemporally sensitive knowledge graph includes:

[0125] Assign timestamps and spatial coordinates to each entity node and relation edge in the knowledge graph. The timestamp format adopts the ISO 8601 standard, and the spatial coordinates adopt geographic grid coding or latitude and longitude coordinates.

[0126] When a node's state changes, an incremental graph version is generated, and the type of the triggered event, the scope of its impact, the hash value of the associated data source, and the spatial displacement are recorded in the change tracing log.

[0127] Based on the version snapshot function of the temporal graph database, it supports the retrospective of historical graph status by risk event time window, realizes risk root cause analysis and spatial impact range tracing, and ensures spatiotemporal data continuity.

[0128] The anomaly detection specifically includes: timestamp anomalies, analyzing whether the interval of entity state update timestamps meets a preset threshold, or whether the timeliness has expired; spatial coordinate anomalies, detecting whether the transportation path coordinates deviate from a preset geographic grid, or whether the entity density within a unit grid exceeds a dynamically adjusted threshold; graph change anomalies, combining node attribute change logs and version traceability information in the knowledge graph to detect whether the change frequency of the same entity in consecutive graph versions exceeds the historical average; if the spatiotemporal attributes meet preset rules and there are no abnormal events in the graph version change records, no risk label will be output.

[0129] In the above technical solution, the entity types of the real-time data sources for the entire logistics supply chain include: IoT sensor time-series data, structured data from enterprise resource planning systems, spatial path coordinates and market sentiment text data from transportation management systems, market sentiment monitoring platform data, operational indicator data of supply chain node enterprises, order status, warehouse inventory, transportation path coordinates, and environmental events.

[0130] The risk label specifically includes: defining the structure of the risk label as anomaly type and risk type; based on the anomaly detection results, the anomaly types include: timestamp anomaly, spatial coordinate anomaly, map change anomaly, and composite rule anomaly; specific risk types include: transportation delay, information update lag, timeliness failure, path deviation, regional congestion, cargo delay, geofence crossing, abnormal inventory fluctuation, supplier delivery interruption, node attribute tampering, relationship chain breakage, impact of natural disasters, impact of supply chain disruption, and systemic security incidents, etc.

[0131] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Other embodiments obtained by those skilled in the art based on the above description of the present invention without departing from the principles of the present invention should all fall within the scope of protection of the present invention.

Claims

1. A method for dynamic risk identification in logistics supply chains based on a large language model, characterized in that, Includes the following steps: Access real-time heterogeneous data streams covering the entire logistics supply chain, and perform spatiotemporal normalization processing on the real-time heterogeneous data streams through a dynamic sliding window mechanism to construct a three-dimensional spatiotemporal data tensor set with entity type, timestamp, and spatial grid encoding as dimensions; The three-dimensional spatiotemporal data tensor set is input into a pre-trained large language model. Through the entity-relationship-time triple extraction module, dynamic entities and their relationships in the logistics supply chain are identified, a dynamic knowledge graph with spatiotemporal attributes is constructed, and an incremental graph version is generated based on the entity state change events. The corresponding graph change log is stored in the temporal graph database. Based on incremental map versions, combined with their corresponding map change logs, multi-dimensional anomaly detection is performed to generate structured risk labels. When a risk event occurs, the spatiotemporal coordinates of the root cause of the risk event can be accurately located through the version backtracking function of the temporal graph database.

2. The method for dynamic risk identification in logistics supply chains based on a large language model according to claim 1, characterized in that, The dynamic sliding window mechanism specifically includes: The time window is dynamically adjusted according to the update frequency of the real-time heterogeneous data stream: a short window is used for high-frequency data and a long window is used for low-frequency data; when the timeliness difference of data within the time window exceeds a preset threshold, the time window is automatically split into multiple sub-windows, and each sub-window corresponds to a data source type. The spatial window is dynamically adjusted, with a preset 6-bit Geohash encoding. The system calculates the entity density and real-time heterogeneous data stream update frequency within the geographic grid cell in real time: when the entity density exceeds the preset density threshold or the update frequency is greater than the preset update frequency threshold, the geographic grid is expanded to 7-bit precision; when the entity density is less than 20% of the preset density threshold, the geographic grid is reduced to 5-bit precision.

3. The method for dynamic risk identification in logistics supply chains based on a large language model according to claim 2, characterized in that, The spatiotemporal normalization process specifically includes: Time dimension: The timestamps of all access devices are calibrated through the network time protocol to eliminate time asynchrony problems caused by device clock deviation and ensure that the order of events is accurate. Spatial dimension: The geographic coordinate system is uniformly converted to Geohash encoding; for encrypted or biased coordinates, compensation and correction are performed to ensure the consistency and accuracy of spatial location.

4. The method for dynamic risk identification in logistics supply chains based on a large language model according to claim 3, characterized in that, The set of three-dimensional spatiotemporal data tensors specifically includes: In terms of entity type, each channel represents a specific supply chain entity type, which is coded and divided according to predefined entity types. The entity types come from the accessed real-time heterogeneous data stream. The specific channel design is as follows: Channel 1 represents the inventory of all nodes, Channel 2 represents the location of transport vehicles in transit, and Channel 3 represents the intensity of public opinion on port congestion extracted from news. In terms of time window dimension, the timestamps of all real-time heterogeneous data streams are first calibrated through the network time protocol to eliminate device clock deviation; the dynamically adjusted time window is divided into a fixed number of continuous time slices, each time slice representing a time snapshot, which is used to capture the evolution pattern of entity state in time series. Spatial grid dimension: The spatial grid dimension is divided according to Geohash encoding. By dynamically adjusting the Geohash precision, fine-grained monitoring of key areas and coarse-grained monitoring of non-critical areas can be achieved. Each element value in a three-dimensional spatiotemporal data tensor Record the specific state of the i-th entity type within the j-th time slice and the k-th spatial grid cell.

5. The method for dynamic risk identification in logistics supply chains based on a large language model according to claim 1, characterized in that, The construction of a dynamic knowledge graph with spatiotemporal attributes specifically includes: Through prompting engineering, the pre-trained large language model is guided to understand the intrinsic meaning of different dimensions of data in the three-dimensional spatiotemporal data tensor and to perform entity-relation-time triple extraction task. Entity node: Each entity includes inherent attributes and time-varying attributes, wherein the time-varying attributes are dynamically associated with the latest timestamp and spatial coordinates; Relationship edge: Each relationship edge is associated with a timestamp, indicating the time or time period when the relationship edge was established, as well as the geographical location where the relationship occurred.

6. The method for dynamic risk identification in logistics supply chain based on a large language model according to claim 5, characterized in that, The prompt project is a predefined structured prompt template; the structured prompt template contains the following instructions: Semantic parsing instructions for the three dimensions of entity type, time slice, and spatial grid in a three-dimensional spatiotemporal data tensor; The pre-trained large language model is required to analyze data according to the reasoning steps of the thought chain to identify the instructions of entities and relationships; The constraint output is a standardized JSON format instruction containing the subject, relation, object, relation timestamp, and relation spatial coordinates.

7. The method for dynamic risk identification in logistics supply chains based on a large language model according to claim 1, characterized in that, The generation of the incremental map version specifically includes: When a pre-trained large language model recognizes an entity state change from a three-dimensional spatiotemporal data tensor generated by a newly accessed real-time heterogeneous data stream, it does not modify the current knowledge graph version, but instead generates a lightweight difference file. The difference file records the changes, including: the unique ID of the object being changed, the changed attribute key-value pairs, the change timestamp, and the event type that triggered the change; Monitor the event occurrence rate at the entity level. When the event occurrence rate of an entity exceeds a dynamic threshold calculated based on its historical behavior within a preset time window, aggregate multiple change events of that entity within that window to generate a summary change record.

8. The method for dynamic risk identification in logistics supply chains based on a large language model according to claim 7, characterized in that, The version backtracking function of the temporal graph database is specifically implemented as follows: Storage and Indexing: Each difference file is associated with a unique time range and records the parent-child dependencies between it and adjacent difference files, forming a version chain; Time travel query: When a risk event is detected, analysts input the time and location of the event; by tracing back along the version chain, a dynamic knowledge graph snapshot of the spatiotemporal attributes at the moment the risk occurred is reconstructed; Risk diffusion path construction: Based on the backtracking knowledge graph snapshot, and combined with relationship changes and spatial displacements in the change log, a topological diffusion path graph of risk events is constructed.

9. The method for dynamic risk identification in logistics supply chains based on a large language model according to claim 8, characterized in that, The topological diffusion path diagram of the risk event is constructed as follows: Starting with the risk source points in the knowledge graph snapshot obtained through backtracking, scan the graph change log; Based on the chronological order of event timestamps and the spatial proximity of the event locations recorded in the map change log, multiple discrete entity state change events are connected into a propagation chain with causal and spatiotemporal order. The propagation chain is visualized on a graph structure to generate a topological diffusion path diagram of the risk event.

10. The method for dynamic risk identification in logistics supply chains based on a large language model according to claim 1, characterized in that, The anomaly detection specifically includes: If the timestamp is abnormal, analyze whether the interval of the entity status update timestamp meets the preset business time interval threshold, or whether the timeliness has expired. If spatial coordinates are abnormal, check whether the coordinates of the transportation path deviate from the preset geographic grid, or whether the entity density within a unit grid exceeds the preset density threshold. Map change anomalies are identified by counting the frequency of attribute changes for the same entity in consecutive incremental map versions. When the frequency of attribute changes exceeds a preset multiple K of the historical average frequency, or exceeds an adaptive threshold calculated based on the historical fluctuation range, it is marked as an anomaly. If the timestamp and spatial coordinate attributes conform to the preset rules, and there are no abnormal events in the map change record, then no risk label will be output.

Images