Enterprise supply chain risk prediction method based on artificial intelligence

Abstract

The invention discloses an enterprise supply chain risk prediction method based on artificial intelligence, and relates to the technical field of supply chain risk management driven by artificial intelligence, and the method comprises the following steps: (a) dynamically obtaining sudden external event information in an internet unstructured data source through a heterogeneous intelligence collection network; (b) an event semantic distillation module is adopted to analyze external event information, and event ontology tags with industry correction parameters are generated; according to the enterprise supply chain risk prediction method based on artificial intelligence, through dynamic bandwidth scheduling and a multi-stage cleaning mechanism of a heterogeneous intelligence acquisition network, capture and high signal-to-noise ratio transmission of sudden unstructured events are realized, and the problem of information lag in the traditional technology is radically solved; and the event semantic distillation module is combined with dynamic correction of an industry sensitivity weight table, so that cross-industry risk misjudgment is eliminated, and the analysis accuracy of scenes such as transportation interruption of chips in the electronic manufacturing industry is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence-driven supply chain risk management, and specifically to an enterprise supply chain risk prediction method based on artificial intelligence. Background Art

[0002] In recent years, artificial intelligence (AI) technologies, particularly machine learning and deep learning, have been introduced into the field of supply chain risk prediction, demonstrating significant potential. These technologies can process massive amounts of structured data, identify complex patterns and hidden correlations, and build predictive models to assess common factors such as supplier financial risk, logistics delay risk, and demand volatility. This has significantly improved the automation and accuracy of forecasting. However, existing AI-based forecasting methods have a significant limitation: they are slow to respond to and inaccurately address risks driven by sudden, unstructured external events. Enterprise supply chains are often subject to external shocks such as sudden natural disasters, geopolitical conflicts, major regulatory and policy changes, regional public health crises, or significant social media outbursts. These events are often unpredictable, rapidly evolving, and require information from dispersed sources, often in the form of unstructured data such as text, news, and social media posts. Existing solutions often rely heavily on internal structured operational data or limited external structured market data sources. They lack the efficient, real-time capabilities to proactively perceive, capture, parse, and quantify these massive unstructured information flows from the open internet, news media, and government announcements. Even if relevant information is obtained, it is difficult to timely and effectively convert event semantics into risk characteristics that the model can understand, and drive the prediction model to make rapid and accurate dynamic adjustments. This results in the system's risk warning signals being seriously delayed and quantitative assessments being inaccurate when major external events occur, making it impossible to provide timely decision-making support for enterprises to mitigate the impact. Therefore, the core technical problem that needs to be urgently solved in the enterprise supply chain risk prediction method based on artificial intelligence is: how to improve the system's real-time perception and accurate analysis capabilities of sudden unstructured external event information, and realize its immediate and accurate feedback and dynamic adjustment of the risk prediction model, thereby significantly enhancing the timeliness and accuracy of predictions in extreme disturbance scenarios. Summary of the Invention

[0003] To achieve the above objectives, the present invention is implemented through the following technical solutions: an enterprise supply chain risk prediction method based on artificial intelligence, comprising the following steps: (a) Dynamically acquire information about sudden external events from unstructured internet data sources through a heterogeneous intelligence collection network. Furthermore, the heterogeneous intelligence collection network covers news, government bulletins, and social media data sources through distributed edge crawler nodes. The crawler initially loads a database of industry keywords and prioritizes collection based on semantic urgency. These priorities include: Highest-level keywords trigger immediate, interruptive collection, forcing idle threads to establish encrypted channels and ensuring transmission to a central scrubbing pool within five seconds; intermediate-level keywords initiate high-frequency polling collection, dynamically compressing transmission time windows based on network status; and basic-level keywords are processed in batches only when the system is idle.

[0004] The cleaning pool performs multi-level filtering, including comparing text hashes and timestamps to eliminate duplicate reports; automatically marking government sources as credible, verifying account authentication status and entity mentions on social media platforms; and stripping out advertising code and content in non-target languages. In the dynamic cache queue, if the same event is reported by multiple independent sources in a short period of time, the event aggregation engine is triggered to consolidate and process the reports. All operations simultaneously record the data lineage map, including the data source site, collection time, and cleaning logs.

[0005] (b) The event semantic distillation module parses external event information and generates event ontology labels with industry correction parameters. Furthermore, the event semantic distillation module performs multi-dimensional analysis on the cleansed text, including identifying subject-action association pairs, such as "port-closure," spatiotemporal tags, such as geographic coordinates and effective time windows, and impact scope modifiers, such as "full" or "partial." After mapping to a predefined supply chain event ontology library to generate initial labels, industry risk correction is initiated, including searching the industry sensitivity weight table based on the company's registered industry code and invoking the industry correction coefficient for the event type.

[0006] When semantic conflicts exist in text, the proportion of negative mentions of the conflicting entity in recent multi-source reporting is examined. If the conflicting entity is overwhelmingly negative, the label is maintained at a high confidence level; if the conflict is evenly matched, the label is downgraded for manual review. Finally, an event label with an industry correction factor and confidence level is output. Feature vectors for unlisted events are automatically extracted and temporarily stored, and initial coefficients are generated using a cross-industry impact simulator.

[0007] (c) The impact quantification module associates event labels with the real-time supply chain topology and outputs node risk values. Furthermore, the impact quantification module injects event label feature vectors into the real-time supply chain topology. First, directly affected nodes are located and the initial impact values ​​are scaled by industry correction coefficients. Risk is propagated layer by layer along the topological edges, including assigning high-intensity propagation weights to direct supply edges and medium-intensity weights to multi-level transit edges. Weights are dynamically adjusted based on real-time logistics punctuality: weights are increased when the path is unobstructed, decreased for persistent delays, and severed for complete disruptions. For unknown event types, event features are extracted and matched to similar cases in the historical database. Impact patterns are then reused based on topological similarity. When no matching cases exist, the events are decomposed into atomic events, quantified separately, and the risk values ​​reconstructed. When multiple events impact the same node, the geometric mean is used for aggregation.

[0008] (d) When the node risk value exceeds a threshold, a neuron-directed update mechanism is initiated to adjust the baseline prediction model parameters. Furthermore, the neuron-directed update mechanism is activated when the node risk value exceeds a dynamic threshold. The sensitivity analysis unit monitors the responses of neurons in each layer of the baseline model, including screening a subset of neurons whose activation intensities consistently exceed historical fluctuations and are strongly correlated with the event type. The gradient reprojection unit generates targeted adversarial examples, specifically by constructing perturbation inputs by fine-tuning event parameters, retaining samples whose risk prediction trends align with the direction of real-time logistics deterioration. Only the target subset of neurons is trained in small batches, with the learning rate increasing with the risk value but subject to a safety upper limit. The shadow model is run synchronously, including the original model and the updated model processing real-time data in parallel. When prediction deviations repeatedly exceed the warning threshold, the learning rate is reduced and training is resumed for the first time. A second time, the process is frozen to diagnose the root cause. A third time, the parameters are forcibly rolled back.

[0009] (e) Generate risk prediction results after confirming the credibility of the data and the validity of the model through a dual verification mechanism. Furthermore, the dual verification mechanism performs data layer and model layer verification simultaneously, including: Data Layer: Event tags for key entities are cross-checked with real-time logistics status. Government data is automatically approved; social media reports must be republished by authoritative media and not officially refuted. When inconsistencies arise, complete inconsistencies interrupt the process, while partial inconsistencies reduce the risk value.

[0010] Model layer: Validate the updated model using a historical extreme event test set. Event types must belong to the same pre-set risk classification cluster, and supply chain topology similarity must exceed a set threshold. Validation is only passed if the prediction direction of key nodes is correct and the comprehensive deviation meets the standard.

[0011] After passing the double-check validation, a report is generated based on the confidence level. If the report fails, data inconsistencies freeze the process, and the model fails, resulting in a fallback to the baseline model. In the event of resource overload, a degradation strategy is activated, namely: validating only the head node, compressing the test set samples, and relaxing the deviation tolerance.

[0012] Preferably, in step (a), the heterogeneous intelligence collection network deploys an edge crawler cluster and adopts a dynamic bandwidth allocation strategy, which adjusts the data collection priority in real time according to the urgency of the preset keyword set.

[0013] Preferably, the event semantic distillation module of step (b) includes: A multi-head cross-attention architecture maps unstructured text to a supply chain event ontology library; The industry risk correction unit adjusts the probability distribution of event ontology labels based on the industry sensitivity weight table. Furthermore, the construction of the industry sensitivity weight table requires the screening of high-quality historical cases, including: clear event type, clear industry classification, and complete loss data. Industry differences are quantified through a multivariate regression model, including: event intensity as the basic independent variable, industry type as the moderating variable, and node loss as the dependent variable. The generated coefficients are double-verified by back-to-back expert reviews and recent event backtesting. When the forecast deviation of an industry continues to deviate, the coefficient review is triggered. The initial coefficients of emerging industries are mapped to adjacent industries, and an independent model is established after accumulating enough events.

[0014] Preferably, the industry sensitivity weight table is constructed through historical event analysis, including: performing regression analysis on actual loss data caused by the same type of external events at different industry supply chain nodes to generate industry-specific correction coefficients.

[0015] Preferably, the impact quantification module of step (c) includes: Graph convolution operation unit, which embeds event ontology labels into node feature vectors of the enterprise supply chain topology structure; The cross-event correlation engine calculates the risk value by matching the node impact patterns of similar events in the historical event library.

[0016] Preferably, the cross-event correlation engine performs the following operations: Extract the feature vector of the current event and calculate its similarity with the feature vector of the historical event; The top K historical events with the highest similarity are selected and weightedly fused according to their actual impact data and topological matching degree.

[0017] Preferably, the neuron-directed updating mechanism of step (d) includes: Sensitivity analysis unit, which identifies the subset of neurons in the baseline prediction model that has the highest response intensity to the current event; The gradient reprojection unit only performs adversarial sample training on the subset of neurons and updates the parameters.

[0018] Preferably, the dual verification mechanism of step (e) includes: Data layer verification: When the event body label is inconsistent with the real-time logistics status data, the processing flow is interrupted; Model-level verification: The updated model is tested on a historical extreme event dataset, and the results are output only when the error meets the standard.

[0019] Preferably, the real-time logistics status data includes at least two of ship positioning information, customs clearance status and warehouse sensor data; the historical extreme event data set must meet the following requirements: the event type belongs to the same preset classification as the current event, and the supply chain topology similarity meets the predetermined topology matching criteria.

[0020] Preferably, the shadow model is run synchronously during the neuron directed update process, and the prediction deviation value between the updated model and the original model is compared in real time; when the deviation value exceeds the warning threshold, the model parameters are rolled back.

[0021] The present invention provides an enterprise supply chain risk prediction method based on artificial intelligence. It has the following beneficial effects: This AI-based enterprise supply chain risk prediction method, through the dynamic bandwidth scheduling and multi-level cleaning mechanism of heterogeneous intelligence collection networks, can capture sudden unstructured events and transmit them with high signal-to-noise ratio, thus eradicating the information lag problem of traditional technologies. The event semantic distillation module, combined with the dynamic correction of the industry sensitivity weight table, eliminates cross-industry risk misjudgments and improves the accuracy of analysis of scenarios such as chip transportation interruptions in the electronics manufacturing industry. The impact quantification module, based on real-time topology propagation control and cross-event pattern migration, overcomes the problem of inaccurate impact of unknown events. The neuron directed update mechanism is more efficient than the traditional full model retraining. The dual verification mechanism outputs the credibility of the prediction results in a hierarchical and quantitative manner through data-model dual-track mutual verification, providing an auditable risk assessment for decision-making.

[0022] This AI-based enterprise supply chain risk prediction method builds a closed-loop system from risk perception to decision support. The industry calibration mechanism enables targeted early warnings for differential risks such as raw material price increases and semiconductor equipment embargoes in the apparel industry. Topological dynamic propagation accurately locates vulnerable nodes in the supply chain, guiding enterprises to prioritize the reinforcement of high-impact paths. Shadow model monitoring and automatic rollback ensure the security of model updates and avoid production line shutdowns caused by erroneous decisions. Resource adaptive strategies ensure the stable output of plans in concurrent crises such as trade wars, assisting managers in initiating alternative procurement or inventory release during prime time windows. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a schematic diagram of the framework of the enterprise supply chain risk prediction method based on artificial intelligence of the present invention; Figure 2 This is a flow chart of the enterprise supply chain risk prediction method based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] See also Figure 1 and Figure 2 The present invention provides a technical solution: an enterprise supply chain risk prediction method based on artificial intelligence, comprising the following steps: (a) Dynamically acquire information about sudden external events from unstructured internet data sources through a heterogeneous intelligence collection network. It should be further explained that, in its implementation, the heterogeneous intelligence collection network covers target data sources, including news portals, government announcement platforms, and social media APIs, through a distributed, lightweight edge crawler node. The crawler node is initialized with a preloaded industry keyword library containing keywords such as strikes, sanctions, earthquakes, and epidemic outbreaks. Keywords are then prioritized based on semantic urgency, namely: First-level keywords with the highest urgency: trigger immediate interruptive collection, for example, keywords related to war and trade ban; Second-level keywords with medium urgency: Start high-frequency polling and collection, for example, keywords such as port closures and logistics strikes; The third-level keywords in basic monitoring are routine scanning and collection, for example, policy revisions and weather warning keywords.

[0026] When the crawler captures text containing first-level keywords, it immediately seizes bandwidth resources and initiates an encrypted data transmission channel, ensuring that the original data is transmitted to the central cleaning pool within a set time period, such as ensuring that the original data is transmitted to the central cleaning pool within 5 seconds. Second- and third-level keyword data is compressed and transmitted in batches according to a preset time window. The cleaning pool uses a multi-level filtering mechanism, including deduplication verification, credibility marking, and noise stripping. Deduplication verification includes comparing text hash values ​​and publication timestamps to eliminate duplicate reports. Credibility marking includes automatically marking government domain sources as highly trustworthy and requiring verification of social media sources' account authentication status. Noise stripping includes removing advertising code, irrelevant hyperlinks, and content not in the target language.

[0027] The cleaned data stream enters a dynamic cache queue. If the same event is reported by more than a set number of independent sources within a set timeframe—for example, if the same event is reported by more than three independent sources within 10 minutes—then the event aggregation engine is triggered to merge the relevant text. If a data source continuously returns invalid responses, its collection weight is automatically reduced until it is temporarily blocked. All processing steps synchronously record the data lineage map, providing traceable event context for subsequent steps.

[0028] (b) The event semantic distillation module parses external event information and generates event ontology labels with industry-corrected parameters. It should be noted that in its implementation, after receiving the cleaned unstructured text, the event semantic distillation module first uses a multi-head cross-attention architecture to parse three semantic dimensions in parallel: Subject-action pairs: identify core entities and their state changes. Core entities include ports, and state changes include strikes and closures. Spatiotemporal tagging: extracting geographic location coordinates and event effective time windows; Impact scope modifiers: capture degree descriptors and chain reaction prompt words. Degree descriptors include comprehensive and partial, and chain reaction prompt words include cause and affect.

[0029] Mapping and matching is performed based on the predefined supply chain event ontology library to generate initial event labels, such as port operation interruption. At this time, the industry risk correction mechanism is activated: Industry sensitivity matching: Based on the target enterprise's industry code, the correction coefficient of the event type in the industry sensitivity weight table is retrieved. The target enterprise's industry code includes the manufacturing classification number. Confidence cross-validation: When contradictory descriptions exist simultaneously in a text, such as when a strike and resumption of work negotiations coexist in contradictory descriptions, an entity relationship graph traversal is initiated. This includes: if the event subject has been mentioned by more than five independent sources in the past week, and the proportion of negative descriptions exceeds 80%, a high-confidence label is adopted; otherwise, the label is downgraded to a pending verification label, triggering the manual review interface.

[0030] The final output is an event ontology label with industry correction parameters in the format of [event type][industry correction coefficient][confidence level]. If a new event type is encountered that is not covered, the ontology library is automatically expanded and the feature vector is recorded. At the same time, a cross-industry impact simulator is launched to estimate the initial sensitivity coefficient.

[0031] (c) The impact quantification module associates the event ontology label with the real-time supply chain topology and outputs the node risk value. It should be further explained that, in the specific implementation process, after receiving the event ontology label with industry correction parameters, the impact quantification module first parses the event type, industry correction coefficient, and confidence level, and simultaneously accesses the enterprise's real-time supply chain topology, including the connection relationships and attribute weights of supplier nodes, logistics path nodes, and warehouse nodes. Perform the following association operations: Dynamic embedding of topological graphs includes: converting event ontology labels into feature vectors and injecting them into the first-level nodes directly affected by the events; and amplifying or reducing the initial impact values ​​of the first-level nodes according to the industry correction coefficient.

[0032] Risk propagation calculation involves spreading the impact layer by layer along the connection direction of the supply chain topology edges, including: direct supply relationship edges, such as from supplier to factory, which are assigned a propagation weight of 0.7; multi-level transit edges, such as from warehouse to regional distribution center, which are assigned a propagation weight of 0.4; each layer of propagation recursively calculates the downstream node risk value according to "node current risk value × propagation weight × path health", and the path health is dynamically adjusted according to the real-time logistics delay rate.

[0033] Cross-event correlation calibration includes: if the current event type does not exist in the historical database, the following operations are performed: extract the event feature vector, including the subject type, impact scope, and industry coefficient; match events in the historical database with a similarity greater than 85%, and weightedly fuse their node impact patterns according to the topological structure similarity; when the difference between the historical event and the current supply chain topology is greater than 30%, start the Monte Carlo simulation to generate the risk distribution interval.

[0034] The final output is the risk value and confidence level of each node, in the format of [node ID][risk value][confidence level]. If the same node is affected by multiple events simultaneously, the geometric mean of the risk values ​​of each event is taken as the aggregation result.

[0035] (d) When the node risk value exceeds the threshold, the neuron-directed update mechanism is activated to adjust the baseline prediction model parameters. It should be further explained that in the specific implementation process, when the risk value of any supply chain node exceeds the preset dynamic threshold, such as the threshold of 0.7 for key supplier nodes and 0.6 for logistics nodes, the neuron-directed update mechanism is triggered. First, the baseline prediction model is loaded, that is, the pre-trained LSTM-Transformer fusion network, and the following operations are performed: Sensitive neuron positioning: The current event feature vector and the topological features of the affected nodes are input into the model to monitor the activation strength of neurons in each hidden layer. Neurons with activation values ​​exceeding 3 standard deviations of the historical mean of the layer and strongly correlated with the event type are screened and marked as highly sensitive subsets.

[0036] Adversarial sample generation: Construct a perturbation sample set based on the current event characteristics, calculate the gradient response of highly sensitive neurons to the perturbation through backpropagation; retain the perturbation samples that make the gradient direction consistent with the risk propagation trend.

[0037] Directed parameter update: Only a subset of highly sensitive neurons is trained in small batches: the loss function is calculated using adversarial examples, and the gradient update range is limited to this subset; the update intensity is regulated by the node risk value: for every 0.1 increase in the risk value, the learning rate is increased by 1.5 times, but the upper limit does not exceed 2 times the baseline learning rate.

[0038] At the same time, shadow model monitoring is started: the original model and the updated model process the same batch of real-time data in parallel; if the prediction deviation of the updated model on 5 consecutive samples exceeds 15% of the shadow model, the update is judged to be invalid, and the parameters of the previous version are automatically rolled back and an alarm is generated.

[0039] (e) Generate risk prediction results after confirming the credibility of the data and the validity of the model through a dual verification mechanism. It should be further explained that in the specific implementation process, the dual verification mechanism simultaneously performs two types of verification before outputting the risk prediction results, including: The first level of data credibility verification: extract key entities from the event ontology label, such as port name and supplier ID, and match them with the real-time logistics status database; if the event label is "port closed", but the AIS trajectory of the corresponding port ship shows that a ship has left the port normally in the past two hours, it is judged as a data contradiction; government-sourced data automatically passes the verification, and social media-sourced events must meet the following requirements: the same event is reprinted by more than five authoritative media within 30 minutes and there is no official rumor-refuting record.

[0040] The second level of model validity verification: load the historical extreme event test set, screen cases of the same type as the current event and with a supply chain topology similarity of not less than 75%; run the updated model on this test set, and calculate the mean absolute error between the predicted risk value and the actual loss value; if the error rate does not exceed 12% and the prediction direction of the key nodes is correct, the model is determined to be valid.

[0041] When both verifications are passed, a final risk prediction report is generated. The report indicates the confidence level, including: Level A: Data verification is consistent and the model error is ≤8%; Level B: Partial data inconsistencies have been manually reviewed or the model error is ≤12%; Level C: The data source is single but the model error meets the standard and needs to be re-verified every hour.

[0042] Initiate emergency response for situations that fail verification, including: when data verification fails, freeze the prediction related to the event for 24 hours and request manual verification of the original source; when model verification exceeds the tolerance, fall back to the baseline model output result and mark it as "unverified prediction".

[0043] In step (a), the heterogeneous intelligence collection network deploys an edge crawler cluster and employs a dynamic bandwidth allocation strategy. This strategy adjusts data collection priorities in real time based on the urgency of a preset keyword set. It should be further explained that, during implementation, the heterogeneous intelligence collection network operates under the control of the dynamic bandwidth allocation strategy, and its edge crawler cluster prioritizes collection tasks based on the urgency of a preset keyword library. When a crawler node detects text containing the most urgent keyword, it immediately interrupts all current low-priority tasks and uses idle thread pool resources to establish an encrypted transmission line, ensuring that event data reaches the central cleaning pool directly. For moderately urgent keyword data, data is aggregated and compressed within a fixed time window and then transmitted. The time window length is dynamically adjusted based on current network latency, shortening the window to increase transmission frequency during network congestion. Basic monitoring keyword data is processed in batches only during system idle periods. Bidirectional traffic monitoring is implemented on all transmission channels. If high-priority data streams continuously occupy bandwidth for longer than a set period, some bandwidth resources are automatically released, reverting to balanced mode. The crawler node has a built-in source site health evaluator. When the response error rate of a specific data source continues to rise, its collection weight is gradually reduced until a temporary block is triggered, and the crawler node switches to the backup mirror site to maintain coverage.

[0044] The cleaning pool performs a multi-stage cleansing operation on the raw text. These operations include: first, filtering out duplicate reports by comparing text hashes and publication timestamps; then, verifying source credibility by directly labeling government domains and authoritative media sources with a high-trustworthiness label; and finally, allowing social media content to be released only after it meets both official account verification status and a threshold for entity mentions. Finally, non-textual noise is removed, retaining the core event descriptions. Data lineage records are generated during this processing, recording key fields such as the data source site, collection timestamp, and cleaning operation logs.

[0045] The event semantic distillation module of step (b) includes: A multi-head cross-attention architecture maps unstructured text to a supply chain event ontology library; The industry risk correction unit adjusts the probability distribution of event ontology labels based on the industry sensitivity weight table.

[0046] It should be further explained that, during its implementation, the event semantic distillation module performs multi-dimensional semantic analysis on the cleaned input text. This includes: using a multi-headed cross-attention architecture to simultaneously identify subject-action pairs, precise spatiotemporal markers, and impact modifiers within the text. For example, when capturing the phrase "a strike by workers at a major port in a certain country lasted three days," the subject-action pair is extracted as "port-strike," the spatiotemporal markers are localized to the country's geographic coordinates and a three-day time window, and the impact modifier is determined to be "major," representing a high level of impact. Mapping and matching is performed based on the supply chain event ontology library, generating the initial event label "port labor disruption."

[0047] The industry risk correction unit is immediately activated, including: searching the industry sensitivity weight table based on the target enterprise's registered industry code. If the enterprise belongs to the electronics manufacturing industry, the correction coefficient for the "port disruption" event is applied to the industry. The correction operation adopts a layered weighting mechanism: the probability of the basic event is multiplied by the industry coefficient, and then regional chain reaction factors are added, such as the proportion of key components handled by the port. When there is a semantic conflict in the text, the entity relationship verification process is initiated. This includes checking the frequency and sentiment of the conflicting entities in recent multi-source reports. When negative mentions dominate, the high confidence label is maintained; if the positive and negative statements are evenly matched, the label is downgraded to pending verification, triggering the manual review interface and suspending downstream processing.

[0048] The final output event ontology tag integrates three key attributes: a standard event type code, an industry-specific correction value, and a confidence level identifier. When encountering an event type not included in the dataset, the event feature vector is automatically extracted and stored in a temporary ontology library. A cross-industry impact simulator is used to generate initial correction coefficients for manual calibration.

[0049] The industry sensitivity weight table is constructed through historical event analysis, including: regression analysis of actual loss data caused by the same type of external events at different industry supply chain nodes to generate industry-specific correction coefficients. It should be further explained that in the specific implementation process, the construction of the industry sensitivity weight table begins with an in-depth analysis of the historical event library, and conducts pattern mining on the actual impact differences caused by the same type of external events, such as port strikes, at different industry supply chain nodes. First, historical cases that meet quality requirements are screened, including: clear event types, quantifiable impact ranges, clear industry classifications, and complete loss records. When the electronics manufacturing industry and the food processing industry encounter the same level of port strikes, core indicators such as the duration of supply chain node interruption, order fulfillment loss value, and recovery cost are extracted respectively.

[0050] A multivariate regression model is used to quantitatively analyze industry differences. This model uses event intensity as the primary independent variable, industry type as the categorical moderating variable, and node loss value as the dependent variable. The model outputs industry-specific correction coefficients. For example, in the case of a port strike, the coefficient for the electronics manufacturing industry is significantly higher than that for the food processing industry, reflecting the much greater sensitivity of chip shipping timelines to agricultural products. Once generated, these coefficients must undergo two validation steps, including: Industry expert review: Organize field experts to conduct back-to-back evaluations of the rationality of the coefficients, and initiate a review if the difference exceeds the threshold; Recent event backtesting: Select new events that occurred within six months to test whether the deviation between the loss value predicted by the weight table and the actual loss value is within an acceptable range.

[0051] The weight table implements a dynamic maintenance mechanism, including triggering a coefficient review process when a particular industry experiences three or more consecutive forecast deviations significantly from historical levels. Initial coefficients for emerging industries are generated using a neighboring industry mapping method, and independent regression models are established after accumulating sufficient event samples. All coefficient updates must be synchronized with notifications to downstream impact quantification modules to load the new version of the weight table.

[0052] The impact quantification module of step (c) includes: Graph convolution operation unit, which embeds event ontology labels into node feature vectors of the enterprise supply chain topology structure; The cross-event correlation engine calculates the risk value by matching the node impact patterns of similar events in the historical event library.

[0053] It is important to further clarify that, during implementation, after the impact quantification module is activated, the feature vector of the event ontology label is first injected into the enterprise's real-time supply chain topology graph. The graph convolution operation unit performs the following operations: locating the first-level nodes directly affected by the event, such as the shipping hub node in a port closure event, concatenating the event feature vector with the node attribute vector, and then propagating the impact layer by layer along the topological edge connection direction. This propagation process follows preset rules, including high-intensity propagation weighting for direct supply relationship edges and medium-intensity propagation weighting for multi-level transit edges. Simultaneously, real-time monitoring of logistics status data dynamically adjusts the weighting. Specifically, if a transportation route's current on-time performance exceeds the industry benchmark, the propagation weight is increased; otherwise, the weight is decreased.

[0054] The cross-event correlation engine is activated for events of unknown types, including: extracting the subject type, geographical scope and industry correction coefficient of the current event to generate a feature vector, and searching for similar cases in the historical event library. Similarity matching uses multi-dimensional measurement, including: prioritizing event type similarity, followed by comparing the scope of impact and industry coefficients. After screening out several historical events with the highest similarity, check the degree of match between their original supply chain topology and the current structure, including: directly reusing the impact model when the overlap of key nodes is high; if there are significant differences, start the impact migration algorithm, retain the risk propagation path of historical events but replace the node attributes with the current topology parameters. When finally outputting the node risk value, add a simulation calculation mark to the events with low historical matching.

[0055] The cross-event correlation engine performs the following operations: Extract the feature vector of the current event and calculate its similarity with the feature vector of the historical event; The top K historical events with the highest similarity are selected and weightedly fused according to their actual impact data and topological matching degree.

[0056] It should be further explained that, during implementation, the neuron-directed update mechanism is activated after a node's risk value exceeds a threshold, initially loading the baseline prediction model's network structure and parameters. The sensitivity analysis unit then inputs the current event feature vector and the affected node topology data into the model, monitoring the activation state of hidden layer neurons in real time. The screening criteria are based on a dual criterion: the intensity of neuron activation exceeds the upper limit of the layer's historical fluctuation range, and the activation pattern is strongly correlated with the event type. After marking a subset of highly sensitive neurons, an adversarial sample set is generated. This involves constructing a perturbation input by fine-tuning key dimensions of the event feature vector, retaining samples that align the risk prediction trend with the deterioration of real-time logistics.

[0057] The gradient reprojection unit targets a subset of highly sensitive neurons and performs mini-batch adversarial training. This involves calculating the loss function for the perturbed samples and restricting backpropagation to target neurons only. The update intensity is dynamically adjusted by the node risk value, increasing the learning rate with increasing risk, but with a safety cap. Shadow model validation is performed simultaneously during training. The original and updated models process the same real-time data stream in parallel. If the updated model continuously produces prediction deviations outside the acceptable range, a parameter rollback mechanism is triggered to restore the model to its pre-update state and generate a system alert.

[0058] The neuron-directed updating mechanism of step (d) includes: Sensitivity analysis unit, which identifies the subset of neurons in the baseline prediction model that has the highest response intensity to the current event; The gradient reprojection unit only performs adversarial sample training on the subset of neurons and updates the parameters.

[0059] It should be further explained that, in the specific implementation process, after the neuron directed update mechanism is activated, the sensitivity analysis unit monitors in real time the response intensity of the neurons in each hidden layer of the baseline prediction model to the input event features. Double verification is performed when selecting highly sensitive neurons, including: the neuron activation value must continuously exceed the historical mean of the layer by more than two standard deviations, and its activation pattern must have a strong statistical correlation with the current event type. For example, in tariff adjustment events, the response intensity of the neurons in the trade policy encoding layer must be significantly higher than that in other areas. After locking the target neuron subset, the gradient reprojection unit generates a set of targeted adversarial samples. That is, by fine-tuning the key parameters of the event to construct a perturbation input combination, only retaining samples that cause the model output trend to be consistent with the real-time risk propagation direction, and eliminating invalid samples that cause the prediction results to deviate from satellite logistics imagery or real-time customs data.

[0060] Parameter updates are strictly scoped, including allowing only a subset of highly sensitive neurons to participate in backpropagation calculations and freezing the parameters of other neurons. Update intensity is dynamically adjusted based on the node's risk level, with a moderate increase in the learning rate as risk climbs, but with a hard cap to prevent over-adjustment. A shadow model validation channel is synchronously initiated after each update, including: the updated model and the original model process the same real-time data stream in parallel. If the updated model continuously produces predictions that exceed a preset deviation threshold, the parameters are immediately rolled back to the pre-update version, triggering a Level 3 alert. Failed update signatures are also recorded for future reference in preventing similar incidents.

[0061] The dual verification mechanism of step (e) includes: Data layer verification: When the event body label is inconsistent with the real-time logistics status data, the processing flow is interrupted; Model-level verification: The updated model is tested on a historical extreme event dataset, and the results are output only when the error meets the standard.

[0062] It's important to note that, during implementation, a dual verification mechanism simultaneously executes validation processes at both the data and model levels. Data-level verification begins by extracting key entity identifiers from the event's tag, which are then cross-checked with the logistics status monitoring system in real time. This includes: if an event tag states "a port is completely closed," but satellite ship positioning data indicates a cargo ship has departed the port within the past six hours, this is considered a hard contradiction. Data from government announcements automatically passes verification; events originating from social media must meet two conditions: reprinted by three or more authoritative media outlets and no official debunking record. Otherwise, they are placed in a pending queue awaiting manual review.

[0063] Model-level validation involves loading a test set of historical extreme events. The screening criteria strictly adhere to a two-pronged approach: the event type must belong to the same risk classification cluster, and the topology of the affected supply chain nodes must be similar to the current one by exceeding a set threshold. After running the updated model on this test set, the average deviation between the node risk prediction and the historical actual loss value is calculated. A verification certificate is issued if the prediction direction of key nodes is correct and the combined deviation is below the maximum allowable error. If the prediction direction of key nodes is reversed or the deviation continuously exceeds the limit, the model is marked as invalid.

[0064] If both levels of verification pass, a final risk report is generated with annotated confidence levels. Highest-level reports are directly pushed to the decision-making system, mid-level reports include additional recommendations for review, and low-level reports trigger an automatic re-verification mechanism. Failures in verification trigger an emergency protocol: data-layer discrepancies freeze the event chain process, and model-layer failures fall back to the baseline model output with an unverified warning. All verification results are logged into an audit trail, supporting full lifecycle traceability.

[0065] The real-time logistics status data includes at least two of the following: ship positioning information, customs clearance status and warehouse sensor data; the historical extreme event data set must meet the following requirements: the event type belongs to the same preset category as the current event, and the supply chain topology similarity meets the predetermined topology matching standard, wherein the predetermined topology matching standard supply chain topology similarity exceeds 70%.

[0066] It should be further explained that, during the specific implementation process, the real-time logistics status data verification link requires cross-verification from at least two independent data sources, including: ship positioning information must be combined with customs clearance status for verification, or jointly analyzed with warehouse sensor data. When obtaining ship positioning information, priority is given to accessing the satellite AIS tracking service. If the signal of the primary service provider is interrupted, it will automatically switch to port radar data for supplementation; customs clearance status is obtained in real time through the official API to obtain the length of the customs clearance queue and the proportion of abnormal declarations; warehouse sensor data focuses on monitoring the displacement rate of the three-dimensional warehouse shelves and temperature and humidity alarm signals. When there is room for interpretation of multi-source data, hierarchical acceptance rules are implemented, including: satellite positioning data is superior to port radar data, official customs data is superior to enterprise self-declared data, and real-time sensor data is superior to manual inventory records.

[0067] The construction of a historical extreme event dataset requires two rigid conditions: the event type must belong to the same pre-defined risk classification cluster as the current emergency, such as earthquakes and hurricanes within the natural disaster classification cluster, and the topological similarity of the affected supply chain nodes must exceed a set threshold. Similarity assessment focuses on key path matching, including comparing the number of core supplier overlaps, the consistency of logistics hub node distribution, and differences in warehouse hierarchy depth. If the current emergency is a new category, a classification cluster expansion mechanism is activated, including back-to-back determination of classification by three domain experts, with at least two reaching consensus before inclusion in the validation set.

[0068] During the neuron-directed update process, the shadow model is run synchronously to compare the predicted deviation between the updated model and the original model in real time; when the deviation exceeds the warning threshold, the model parameters are rolled back. It should be further explained that in the specific implementation process, the shadow model monitoring channel is started synchronously during the neuron-directed update process, and the original model and the updated model process the same real-time data stream in parallel. The deviation monitoring unit calculates the absolute difference in the risk values ​​output by the two sets of models in real time. When the difference value of several consecutive samples exceeds the dynamic warning threshold, a multi-level response mechanism is triggered, including: automatically reducing the learning rate of the updated model when the limit is exceeded for the first time and injecting historical stable period samples for retraining; freezing the parameter update process after the second limit is exceeded, and starting the deviation root cause diagnosis protocol; if the limit is exceeded for the third time, it is forced to roll back to the original model parameter version, and at the same time marking the update as invalid and generating an encrypted alarm log.

[0069] The shadow model monitoring channel is activated during neuron updates. The deviation monitoring unit calculates the difference between the two sets of model outputs in real time. When dynamic alert thresholds are continuously exceeded, the following procedures are implemented: The first violation injects historical stable samples for retraining; the second violation freezes updates and diagnoses the root cause, such as data leakage, adversarial sample anomalies, or neuron omissions; and the third violation forces a rollback of parameters. When resources are limited, only strategic nodes are monitored, using a lightweight difference algorithm. The rollback operation retains a snapshot of features for pre-screening similar events.

[0070] Unified response strategies for extreme working conditions, including: When the data source fails: satellite signal interruption causes switching to base station positioning; customs failure causes activation of the customs broker confirmation channel; When historical cases are insufficient: decompose into atomic event matching or start Monte Carlo simulation; When the server is overloaded: only process the highest-risk nodes, reduce the sample size, and extend the processing interval; When classifying new events: domain experts make back-to-back decisions, and in case of disagreement, international risk classification standards are used as a reference.

[0071] The deviation root cause diagnosis protocol performs a layered investigation, including: first checking the consistency of the input data stream to confirm that the event feature vector transmission process is distortion-free; secondly, verifying whether the adversarial sample generation module introduces abnormal perturbations; and finally, auditing whether the neuron screening logic misses key response units. During a rollback operation, a snapshot of the updated features is retained for reference before subsequent updates of similar events. In the event of resource overload, a streamlined monitoring mode is activated, which includes comparing only the predicted values ​​of the top risk nodes, compressing the historical sample verification scale, and extending the deviation determination interval.

[0072] Through the dynamic bandwidth scheduling and multi-level cleaning mechanism of the heterogeneous intelligence collection network, the capture and high signal-to-noise ratio transmission of sudden unstructured events are achieved, fundamentally solving the information lag problem of traditional technologies; the event semantic distillation module, combined with the dynamic correction of the industry sensitivity weight table, eliminates cross-industry risk misjudgments and improves the analysis accuracy of scenarios such as chip transportation interruptions in the electronics manufacturing industry; the impact quantification module, based on real-time topology propagation control and cross-event pattern migration, overcomes the problem of inaccurate impact of unknown events; the neuron directed update mechanism is more efficient than traditional full-model retraining; the double verification mechanism, through data-model dual-track mutual verification, quantifies the credibility of the prediction results in layers, providing auditable risk assessment for decision-making.

[0073] A closed-loop system from risk perception to decision support is built. The industry calibration mechanism enables targeted early warning of differential risks such as raw material price increases and semiconductor equipment embargoes in the apparel industry. Topological dynamic propagation accurately locates vulnerable nodes in the supply chain and guides companies to prioritize the reinforcement of high-impact paths. Shadow model monitoring and automatic rollback ensure the security of model updates and avoid production line shutdowns caused by erroneous decisions. Resource adaptive strategies ensure the stable output of plans in concurrent crises such as trade wars, assisting managers in initiating alternative procurement or inventory release during prime time windows.

[0074] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0075] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The enterprise supply chain risk prediction method based on artificial intelligence is characterized by: The steps include: (a) Dynamically acquire information about sudden external events from unstructured Internet data sources through a heterogeneous intelligence collection network; (b) Using the event semantic distillation module to parse external event information and generate event ontology labels with industry correction parameters; (c) Associating event ontology labels with real-time supply chain topology through the impact quantification module and outputting node risk values; (d) When the node risk value exceeds the threshold, the neuron directed update mechanism is activated to adjust the baseline prediction model parameters; (e) Generate risk prediction results after confirming the credibility of data and the validity of the model through a double verification mechanism; The impact quantification module of step (c) includes: Graph convolution operation unit, which embeds event ontology labels into node feature vectors of the enterprise supply chain topology structure; The cross-event correlation engine calculates the risk value by matching the node impact patterns of similar events in the historical event library.

2. The enterprise supply chain risk prediction method based on artificial intelligence according to claim 1 is characterized by: In step (a), the heterogeneous intelligence collection network deploys an edge crawler cluster and adopts a dynamic bandwidth allocation strategy. The dynamic bandwidth allocation strategy adjusts the data collection priority in real time according to the urgency of the preset keyword set.

3. The enterprise supply chain risk prediction method based on artificial intelligence according to claim 1 is characterized by: The event semantic distillation module of step (b) includes: A multi-head cross-attention architecture maps unstructured text to a supply chain event ontology library; The industry risk correction unit adjusts the probability distribution of event ontology labels based on the industry sensitivity weight table.

4. The enterprise supply chain risk prediction method based on artificial intelligence according to claim 3 is characterized by: The industry sensitivity weight table is constructed through historical event analysis, including: performing regression analysis on actual loss data caused by the same type of external events at different industry supply chain nodes to generate industry-specific correction coefficients.

5. The enterprise supply chain risk prediction method based on artificial intelligence according to claim 4 is characterized by: The cross-event correlation engine performs the following operations: Extract the feature vector of the current event and calculate its similarity with the feature vector of the historical event; The top K historical events with the highest similarity are selected and weightedly fused according to their actual impact data and topological matching degree.

6. The enterprise supply chain risk prediction method based on artificial intelligence according to claim 1 is characterized by: The neuron-directed updating mechanism of step (d) includes: Sensitivity analysis unit, which identifies the subset of neurons in the baseline prediction model that has the highest response intensity to the current event; The gradient reprojection unit only performs adversarial sample training on the subset of neurons and updates the parameters.

7. The enterprise supply chain risk prediction method based on artificial intelligence according to claim 1 is characterized by: The dual verification mechanism of step (e) includes: Data layer verification: When the event body label is inconsistent with the real-time logistics status data, the processing flow is interrupted; Model-level verification: The updated model is tested on a historical extreme event dataset, and the results are output only when the error meets the standard.

8. The enterprise supply chain risk prediction method based on artificial intelligence according to claim 7 is characterized by: The real-time logistics status data includes at least two of the following: ship positioning information, customs clearance status, and warehouse sensor data; the historical extreme event dataset must meet the following requirements: the event type belongs to the same preset classification as the current event, and the supply chain topology similarity meets the predetermined topology matching criteria.

9. The enterprise supply chain risk prediction method based on artificial intelligence according to claim 6 is characterized by: The shadow model is run synchronously during the neuron directed update process, and the prediction deviation value between the updated model and the original model is compared in real time; when the deviation value exceeds the warning threshold, the model parameters are rolled back.

Images