Real-time acquisition and remote monitoring platform for supply chain multi-source data based on knowledge graph

Abstract

The application discloses a supply chain multi-source data real-time collection and remote monitoring platform based on a knowledge graph, belongs to the technical field of supply chain monitoring, and comprises the following: a multi-source data collection unit, which is used for collecting original data including transaction time stamps from each node of a supply chain in real time; a time sequence knowledge graph construction unit, which is used for forming a time sequence knowledge graph with double time dimensions and trust degree labels; a graph playback unit, which is used for responding to a playback request, obtaining an abnormal propagation path and an energy accumulation value thereof; a state deduction unit, which is used for predicting entity states at multiple future moments; and a remote monitoring unit, which is used for visually displaying the evolution process of the time sequence knowledge graph, the abnormal propagation path and prediction results with confidence sorting. According to the application, the recursive disturbance back propagation algorithm is used to accurately trace the abnormal propagation path and the root cause, and the intelligent level and the risk management and control efficiency of the supply chain remote monitoring are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of supply chain monitoring technology, specifically a knowledge graph-based platform for real-time acquisition and remote monitoring of multi-source supply chain data. Background Technology

[0002] With the increasing complexity and globalization of global supply chain networks, the data generated by each node in the supply chain exhibits characteristics of being multi-sourced, heterogeneous, massive, real-time, and dynamically evolving. Traditional supply chain monitoring platforms primarily use relational databases to store discrete business data, which has the following significant drawbacks: Low trust in data fusion: Data formats, update frequencies, and reliability vary greatly from data sources. Existing technologies lack a unified trust assessment and fusion mechanism, which can easily lead to data conflicts and decision-making errors.

[0003] Lack of time dimension modeling: Most existing platforms only record transaction timestamps of data, which cannot fully depict the evolution of entity status over time, making it difficult to trace back historical states and analyze the propagation patterns of anomalies.

[0004] Insufficient ability to trace the source of anomalies: When anomalies occur in the supply chain (such as inventory shortages, logistics delays, or production interruptions), existing technologies can only locate the current abnormal node and cannot trace the source and propagation path of the anomaly, resulting in low efficiency in problem solving.

[0005] Low prediction accuracy and confidence: Traditional prediction models are mostly based on single time series data, without considering the complex relationships between supply chain nodes and the impact of anomaly propagation. The prediction results lack reliable confidence assessment and are difficult to support risk warning and decision-making.

[0006] Therefore, those skilled in the art have provided a knowledge graph-based platform for real-time acquisition and remote monitoring of multi-source supply chain data to address the problems mentioned in the background section. Summary of the Invention

[0007] The purpose of this invention is to overcome the problems of low data fusion trust, lack of time dimension modeling, insufficient anomaly tracing ability, low prediction accuracy and confidence, and lack of dynamic feedback mechanism in existing supply chain monitoring platforms, and to provide a knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A knowledge graph-based platform for real-time acquisition and remote monitoring of multi-source supply chain data includes: The multi-source data acquisition unit is used to collect raw data, including transaction timestamps, from each node of the supply chain in real time. The raw data includes inventory status, logistics trajectory, production progress, and transaction records. The temporal knowledge graph construction unit is used to label each collected raw data with a valid time interval and transaction timestamp, and to calculate the multi-source fusion trust degree for each entity's state attribute, forming a temporal knowledge graph with dual time dimensions and trust degree labels. The graph playback unit is used to respond to playback requests, extract historical subgraph sequences from the time-series knowledge graph according to a specified time interval, and calculate the abnormal residual energy of the entity state at each historical moment in reverse along the time axis using a recursive perturbation backpropagation algorithm to obtain the abnormal propagation path and its energy accumulation value. The state inference unit is used to predict the entity state at multiple future times based on the real-time collected current state data and the abnormal propagation path output by the graph playback unit, by using the abnormal residual energy as a weighted encoding input to the time-series graph neural network, and dynamically calculating the comprehensive confidence of each prediction result based on the confidence component, and outputting it after sorting by confidence. The remote monitoring unit is used to visualize the evolution process, anomaly propagation path, and prediction results with confidence ranking of the time-series knowledge graph.

[0009] As a further aspect of the present invention: the specific method for calculating the multi-source fusion trust degree for each entity's state attribute in the temporal knowledge graph construction unit is as follows: For entity state attributes, when they receive observations from multiple heterogeneous data sources and their corresponding source trust levels within the same valid time interval, the pairwise consistency deviation matrix between each observation is calculated. If the absolute value of the deviation between any two observations is less than the preset consistency threshold, the weighted average value with the source trust is used as the fusion state value. The fusion trust is obtained by multiplying the weighted sum of each source trust by the time decay factor. The time decay factor is calculated based on the time difference between the data generation time and the current system fusion processing time, and the preset decay rate is a fixed value. If the deviation between observation pairs exceeds the consistency threshold, a conflict detection mechanism is triggered. By comparing the deviation between the historical fusion value of the entity's state attribute and the current observations, the observation with the smallest deviation from the historical fusion value and a source trust level higher than the dynamic trust elimination threshold is selected as the trusted source. The state value is updated only with this single source data, and the fusion trust level is the product of the single source trust level and the time decay factor.

[0010] As a further aspect of the present invention: the recursive perturbation backpropagation algorithm in the spectrum playback unit specifically includes: For each time slice in the historical subgraph sequence, the abnormal residual energy of each entity node is calculated. The abnormal residual energy is equal to the actual state confidence of the entity node at the current moment minus the expected normal confidence benchmark value of the entity node at the same moment, and the larger value is taken compared with zero. The expected normal confidence benchmark value is obtained by taking the historical confidence moving average of the entity at the same moment in the past statistical period and adjusting it according to the dynamic standard deviation. The time slices are arranged in reverse chronological order, and the abnormal residual energy is propagated from back to front. For the propagation process from the current time to the previous time, the updated abnormal residual energy of the previous time is obtained by summing the original abnormal residual energy of the previous time with the incoming energy from the subsequent time. The incoming energy is the abnormal residual energy of all upstream nodes that have an incoming edge relationship with the current node in the subsequent time, multiplied by the propagation attenuation coefficient on the corresponding edge and then summed. The propagation attenuation coefficient is calculated based on the standard transmission coefficient of the upstream and downstream relationship of the supply chain and the standard deviation of the transportation timeliness corresponding to the relationship at the current time. After all time-propagation calculations are completed, the abnormal residual energy values ​​are sorted from largest to smallest. The nodes whose cumulative energy percentage exceeds the preset energy threshold and their corresponding historical state records are taken as the reasons for the anomalies.

[0011] As a further aspect of the present invention, the propagation attenuation coefficient is dynamically calculated in the following manner: The propagation attenuation coefficient is equal to the basic transmission coefficient from the upstream entity to the downstream entity multiplied by the exponential attenuation value of the product of the propagation attenuation adjustment factor and the standard deviation of transportation timeliness. Specifically, the value of the basic transmission coefficient is pre-calibrated based on the influence of the upstream on the downstream in the supply chain. The calibration method is to take the moving average of the ratio of the covariance of the upstream state change and the downstream state change in the historical data of the upstream and downstream of the supply chain to the variance of the upstream state change. The standard deviation of transportation timeliness refers to the standard deviation of the logistics delivery time or information transmission delay between the upstream and downstream entities within a statistical window of a preset duration before the current moment. In the calculation of the incoming energy, if the same upstream node transmits non-zero energy to the current node at multiple consecutive moments, the incoming energy sequence is smoothed by an exponentially weighted moving average method before being accumulated.

[0012] As a further aspect of the present invention: the temporal graph neural network in the state inference unit adopts a message passing architecture based on anomaly residual weighted coding, and its hidden state update process for nodes at each prediction time step is as follows: The hidden state of the current node at the previous time step and the aggregated neighbor messages are input into the gated recurrent unit to obtain the new hidden state at the current time step. The aggregated neighbor messages are obtained by performing the following operations on each neighbor node and then summing them: first, calculate the attention coefficient between the current node and the neighbor nodes; then, concatenate the hidden state of the neighbor nodes, the current edge feature vector, and the abnormal residual energy encoding vector of the neighbor nodes at the current time step; then multiply by the learnable weight matrix; and finally multiply by the attention coefficient. The abnormal residual energy encoding vector is obtained by normalizing the abnormal residual energy output by the graph replay unit and concatenating it with the historical abnormal energy change rate, where the historical abnormal energy change rate refers to the moving average of the first-order difference values ​​of the abnormal residual energy of the node over multiple consecutive time steps in the past. The prediction output layer maps the final hidden state to the probability distribution of each state category at the next time step.

[0013] As a further aspect of the present invention: the temporal graph neural network introduces a self-supervised training strategy based on time window sliding sampling during the training phase: the historical evolution sequence of the temporal knowledge graph is divided into several training samples according to a fixed-length time window. For the data within each time window, the node states of some time segments within that time window are randomly occluded. The temporal graph neural network is required to predict the occluded state value based on the unoccluded historical information within the time window and the abnormal residual energy before the occluded segment. The training loss function consists of a weighted sum of the prediction cross-entropy loss and the Euclidean distance loss between the predicted state and the true state.

[0014] As a further aspect of the present invention: when the state deduction unit outputs prediction results in order of confidence, the overall confidence of each prediction result is obtained by weighted fusion of the following confidence components: The first confidence component is the normalized complement of the information entropy of the predicted probability distribution output by the time series graph neural network. Specifically, it is obtained by multiplying the ratio of the maximum probability value to the second largest probability value in the probability distribution by a correction factor based on historical prediction deviation after performing a logistic transformation. The correction factor is calculated by comparing the ratio of the average deviation of the model in the most recent predictions to a preset benchmark deviation. The second confidence component is the state stability score of the predicted entity node in the past statistical period, which is obtained by taking the logistic transformation after weighted summation of the node's state change frequency and the reciprocal of the state value variation coefficient. The third confidence component is the model robustness score at the current prediction time step, which is obtained by normalizing the inverse of the average deviation between the prediction results in the prediction windows before and after the current prediction time step and the actual feedback at the corresponding time point, and then taking the moving average. The weighting coefficients of the three confidence components are all preset values ​​and the sum of the three is one.

[0015] As a further aspect of the present invention, it also includes a dynamic feedback calibration unit, used to receive user confirmation or rejection operations from the remote monitoring unit, taking the user-confirmed anomaly cause and the user-accepted high-confidence prediction result as positive feedback samples, and the user-rejected low-confidence prediction result as negative feedback samples, and periodically calibrating and updating the propagation attenuation coefficient in the map playback unit and the confidence weight in the state inference unit online: for the propagation attenuation coefficient, its basic transmission coefficient is updated by comparing the similarity between the user-confirmed anomaly cause and the system's automatic tracing result on the same propagation path; for the confidence weight, the fusion weight of the three confidence components is optimized by gradient ascent by maximizing the ranking score of the user-accepted high-confidence sample in the prediction result ranking list and minimizing the ranking score of the user-rejected sample.

[0016] As a further aspect of the present invention: the remote monitoring unit also includes a multimodal interactive anomaly review interface. When displaying the anomaly propagation path, the multimodal interactive anomaly review interface uses time axis folding technology to merge and display nodes with smooth state changes in multiple adjacent time steps, and only displays the complete time step sequence when the user clicks to expand. When the user selects a root cause node and issues a "simulate mitigation" command, the system calls the state inference unit, corrects the state value of the root cause node to the simulated value specified by the user, and then re-executes the future state prediction, and compares and displays the difference between the original prediction and the simulation-adjusted prediction in the form of a heatmap.

[0017] As a further aspect of the present invention, it also includes a distributed graph storage module. This module stores the time-series knowledge graph using a hybrid sharding strategy of effective time sharding and entity type hashing. A preset natural time unit is used as the first sharding key for effective time. Long-term effective states exceeding the duration of the first sharding key are allocated a separate long-term storage area and stored using the transaction time as the second sharding key. Within each effective time shard, the snapshot index of the shard's start and end times, as well as the differential change logs between each smallest time granularity, are stored simultaneously. The differential change logs are compressed and stored using lossy run-length encoding on the entity state differences and relationship change lists between consecutive time steps. When responding to cross-shard queries, the graph replay unit sequentially reads each involved storage shard, first performing a binary search in the snapshot index of each shard to locate the start time within the target time interval, and then replaying the differential change logs one by one according to the time sequence from the start time to the end time of the target time interval, thus completing the generation of the dynamic evolution sequence.

[0018] Compared with the prior art, the beneficial effects of the present invention are: Trusted fusion of multi-source data: By introducing a multi-source fusion trust calculation mechanism and a conflict detection mechanism, the problem of data conflict from heterogeneous data sources is effectively solved, and the accuracy and reliability of data fusion are improved.

[0019] Complete temporal state modeling: A temporal knowledge graph with dual time dimensions and trust level labels was constructed, which fully depicts the evolution of the state of supply chain entities over time, providing a solid data foundation for historical backtracking and anomaly analysis.

[0020] Precise anomaly tracing: The recursive perturbation backpropagation algorithm is used to calculate the propagation path of the anomaly residual energy in reverse along the time axis, which can quickly locate the source of the anomaly and the propagation chain, greatly improving the efficiency of problem solving.

[0021] High-confidence state prediction: The abnormal residual energy is used as a weighted encoding input to the time-series graph neural network, which fully considers the impact of abnormal propagation on future states. Through the fusion calculation of multi-dimensional confidence components, a reliable confidence assessment is provided for the prediction results.

[0022] Dynamic feedback calibration: By introducing a user feedback mechanism, the propagation attenuation coefficient and confidence weight are calibrated online periodically, enabling the system to adapt to changes in the supply chain environment and maintain high-performance operation.

[0023] High-efficiency storage and query: By adopting a hybrid sharding strategy of effective time sharding and entity type hashing, combined with differential change log compression storage technology, the storage efficiency and cross-shard query speed of time-series knowledge graphs are greatly improved. Attached Figure Description

[0024] Figure 1 This is a structural diagram of a knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The following will describe in detail, with reference to the accompanying drawings, how the solution of this application solves the above-mentioned technical problems.

[0027] Please see Figure 1In this embodiment of the invention, a knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform includes: a multi-source data acquisition unit, used to collect raw data including transaction timestamps from each node of the supply chain in real time; the raw data includes inventory status, logistics trajectory, production progress, and transaction records; a time-series knowledge graph construction unit, used to label each collected raw data with a valid time interval and transaction timestamp, and calculate the multi-source fusion trust degree for each entity state attribute to form a time-series knowledge graph with dual time dimensions and trust degree labels; and a graph playback unit, used to respond to playback requests and extract history from the time-series knowledge graph according to a specified time interval. The system generates a time-series knowledge graph sequence and calculates the abnormal residual energy of the entity state at each historical moment in reverse along the time axis using a recursive perturbation backpropagation algorithm, thus obtaining the abnormal propagation path and its accumulated energy value. A state inference unit, based on real-time acquired current state data and the abnormal propagation path output by the graph playback unit, uses the abnormal residual energy as a weighted encoding input to the time-series graph neural network to predict the entity state at multiple future moments. It also dynamically calculates the comprehensive confidence of each prediction result based on the confidence component and outputs the results sorted by confidence. A remote monitoring unit visualizes the evolution of the time-series knowledge graph, the abnormal propagation path, and the prediction results sorted by confidence. This application achieves reliable fusion of multi-source data and full-link temporal state modeling of the supply chain based on temporal knowledge graphs. It accurately traces the propagation path and root cause of anomalies through a recursive perturbation backpropagation algorithm. Combined with a temporal graph neural network weighted by anomaly residuals, it achieves highly reliable state prediction with multi-dimensional confidence assessment. At the same time, it introduces a user closed-loop feedback mechanism to realize online self-calibration of system parameters. With a multimodal interactive decision support interface and an efficient distributed graph storage architecture, it comprehensively solves the pain points of traditional supply chain monitoring, such as multiple data conflicts, difficulty in anomaly tracing, low prediction accuracy, and easy degradation of system performance. It significantly improves the intelligence level and risk management efficiency of remote supply chain monitoring.

[0028] In this embodiment, the specific method for calculating the multi-source fusion trust level for each entity state attribute in the temporal knowledge graph construction unit is as follows: For an entity state attribute, when it receives observations from multiple heterogeneous data sources and their corresponding source trust levels within the same valid time interval, a pairwise consistency deviation matrix between each observation is calculated; if the absolute value of the deviation between any two observations is less than a preset consistency threshold, the weighted average value with source trust level as the weight is used as the fusion state value. The fusion trust level is obtained by multiplying the weighted sum of each source trust level by a time decay factor, where the time decay factor is calculated based on the time difference between the data generation time and the current system fusion processing time, and the preset decay rate is a fixed value; if the deviation between a pair of observations exceeds the consistency threshold, a conflict detection mechanism is triggered. By comparing the deviation of the historical fusion value of the entity state attribute with the current observations, the observation with the smallest deviation from the historical fusion value and a source trust level higher than the dynamic trust elimination threshold is selected as the trusted source, and the state value is updated only with this single source data. The fusion trust level is the product of the single source trust level and the time decay factor. By designing a scenario-specific multi-source fusion trust calculation mechanism for multi-source heterogeneous observations of the same entity's state attribute within the same valid time interval, a weighted average fusion method with source trust as the weight is adopted when the consistency of each observation is good. A time decay factor based on the time difference between the data generation time and the system fusion processing time is introduced. This can fully utilize the inherent reliability differences of different data sources to improve the accuracy of the fused state value. At the same time, the time decay factor quantifies the timeliness value of the data, avoiding the interference of outdated data on the judgment of the current supply chain state. When there are significant conflicts in the observations, a mechanism based on the deviation of historical fusion values ​​is automatically triggered. The conflict detection and trusted source selection mechanism with dynamic trust elimination threshold can automatically identify and eliminate low-trust data sources without manual intervention. It effectively solves the industry pain points of inconsistent formats, data inconsistencies, frequent conflicts and uneven reliability of multi-source data in the supply chain. It greatly improves the robustness and automation of heterogeneous data fusion. It provides high-quality and high-trust basic data support for the accurate construction of subsequent time-series knowledge graphs, accurate tracing of abnormal events and reliable prediction of supply chain status. It avoids the problems of abnormal misjudgment, risk omission and prediction deviation caused by data conflict or low-quality data from the source.

[0029] Specific examples: Unified preset parameters: consistency threshold T=10, time decay rate λ=0.001 / minute, dynamic trust elimination threshold δ=0.8, entity status attribute is "quantity of electronic component B in East China regional central warehouse A", system fusion processing time is 2026-06-10, 10:15:00.

[0030] Example 1: A scenario where multi-source observations have good consistency; Assume that within the valid time interval [2026-06-10, 09:00:00, 2026-06-10, 10:00:00], the system receives synchronous observations from three heterogeneous data sources and pre-calibrated source trust levels: Data source 1 (Warehouse WMS system): Observation s1 = 1520 items, source trust level w1 = 0.92; Data source 2 (RFID automatic inventory system): Observation value s2=1516 items, source trust level w2=0.88; Data source 3 (shelf IoT weight sensor): Observation value s3=1523 pieces, source trust level w3=0.82; Calculate the pairwise consistency deviation matrix: |s1-s2|=4<10, |s1-s3|=3<10, |s2-s3|=7<10. All deviations are less than the preset consistency threshold. Data fusion is performed using a weighted average method.

[0031] Formula for calculating fusion state value: ; Substituting the values, we get: S = (0.92 × 1520 + 0.88 × 1516 + 0.82 × 1523) / (0.92 + 0.88 + 0.82) ≈ 1519.78 pieces; The median time of data generation, 09:30:00 on June 10, 2026, is taken. The time difference between this and the system fusion processing time is Δt = 45 minutes. The time decay factor is calculated using the following formula: α=e -λΔt Substituting the numerical values, we get: α = e -0.001×45 ≈0.9560; Formula for calculating fusion trust level (weighted average of source trust levels multiplied by time decay factor): Substituting the numerical values, we get: ; Example 2: Scenario where conflicting observations exist from multiple sources; Assume that within the same valid time interval, the system receives three observations and their source confidence levels are as follows: Data source 1 (Warehouse WMS system): Observation s1 = 1520 items, source trust level w1 = 0.92; Data source 2 (RFID automatic inventory system): Observation value s2=1515 items, source trust level w2=0.88; Data source 3 (manual temporary inventory): Observations s3 = 1480 items, source confidence level w3 = 0.75; Calculate the pairwise consistency deviation matrix: |s1-s2|=5<10, |s1-s3|=40>10, |s2-s3|=35>10. If the deviation exceeds the threshold, the conflict detection mechanism is automatically triggered.

[0032] Given that the previous period's historical fusion value of the entity's state attribute is H = 1518, calculate the deviations of each observation value from the historical fusion value: |s1-H| = 2, |s2-H| = 3, |s3-H| = 38.

[0033] The source trust level of data source 3, w3=0.75<δ=0.8, is lower than the dynamic trust elimination threshold and is directly eliminated; data source 1, which has the smallest deviation from the historical fusion value, is selected as the only trusted source, and the fusion status value S=1520.

[0034] Formula for calculating trust level: C = w1 × α; Substituting the values, we get: C = 0.92 × 0.9560 ≈ 0.880.

[0035] In this embodiment, the recursive perturbation backpropagation algorithm in the graph playback unit specifically includes: for each time slice in the historical subgraph sequence, calculating the abnormal residual energy of each entity node, whereby the abnormal residual energy is equal to the actual state confidence of the entity node at the current moment minus the expected normal confidence benchmark value of the entity node at the same moment, and taking the larger value compared with zero, wherein the expected normal confidence benchmark value is obtained by taking the historical confidence moving average of the entity at the same moment in the past statistical period, and adjusting it according to the dynamic standard deviation; arranging each time slice in reverse chronological order, and propagating the abnormal residual energy sequentially from back to front, for the propagation from the current moment to the previous moment... The process involves updating the abnormal residual energy of the previous time step by summing the original abnormal residual energy of the previous time step with the incoming energy from subsequent time steps. The incoming energy is obtained by multiplying the abnormal residual energy of all upstream nodes with incoming edge relationships with the current node in subsequent time steps by the propagation attenuation coefficient on the corresponding edge and then summing them. The propagation attenuation coefficient is calculated based on the standard transmission coefficient of the upstream and downstream relationship of the supply chain and the standard deviation of the transportation timeliness corresponding to the relationship at the current time step. After all time steps have been propagated, the abnormal residual energy values ​​are sorted from largest to smallest. The nodes whose cumulative energy ratio exceeds the preset energy threshold and their corresponding historical state records are taken as the abnormal reasons. By designing a recursive perturbation backpropagation algorithm, precise source tracing of supply chain anomalies is achieved. First, based on historical statistical data, the expected normal confidence benchmark value of each entity node is dynamically calculated. Combined with the actual state confidence value, the anomaly residual energy is obtained, which can effectively distinguish between normal fluctuations and real anomalies in the supply chain, significantly reducing the anomaly misjudgment rate. Second, a time-reverse energy propagation method is adopted to trace the propagation path of the anomaly in reverse along the upstream and downstream relationships of the supply chain. At the same time, a propagation attenuation coefficient is introduced to quantify the degree of anomaly impact between upstream and downstream nodes, which can accurately restore the complete transmission process of the anomaly from its source to the final manifest node, avoiding the shortcomings of traditional source tracing methods that can only locate the current anomaly node but cannot trace the root cause. Finally, the cause of the anomaly is screened by the cumulative percentage threshold of the anomaly residual energy, which can quantify the contribution of different nodes to the anomaly and prioritize the core anomaly node with the greatest impact, providing accurate root cause basis for subsequent anomaly handling and state prediction. The entire algorithm can automatically complete anomaly source tracing without manual intervention, significantly improving the response speed and handling efficiency of supply chain anomalies. At the same time, it provides accurate anomaly propagation path and energy data support for the state inference unit, ensuring the accuracy and pertinence of subsequent state prediction.

[0036] Specific examples: Unified preset parameters: The statistical period is the past 30 days, the preset energy threshold P=80%, the dynamic standard deviation adjustment coefficient k=1.2, the supply chain link is supplier E→manufacturer D→warehouse A→distributor C, and the propagation attenuation coefficient at each link is: from supplier E to manufacturer Manufacturer D to warehouse From warehouse A to distributor The time slice interval is 2 hours, and the target playback time range is [2026-06-10, 10:00:00, 2026-06-10, 18:00:00], with a total of 5 time slices t0=10:00, t1=12:00, t2=14:00, t3=16:00, t4=18:00.

[0037] Example 1: Calculation of initial anomaly residual energy; Formula for calculating abnormal residual energy: ; The formula for calculating the expected normal confidence baseline value is as follows: ; in Let i be the historical moving average of confidence scores for entity i at any given moment over the past 30 days. This represents the standard deviation corresponding to the historical confidence level.

[0038] Taking the stockout anomaly of distributor C at time t4=18:00 as an example, the confidence level of its actual status is known. Historical confidence moving average for the same time over the past 30 days Historical confidence level standard deviation The expected normal confidence baseline value is: ; Initial anomaly residual energy: ; The residual energy of the remaining nodes due to initial anomalies (assuming no obvious initial anomalies): , , , The initial energy of all other nodes is 0 at all other times.

[0039] Example 2: Recursive perturbation backpropagation process; The propagation process proceeds in reverse time order: t4→t3→t2→t1→t0. Update formula: ; in, The abnormal residual energy is the energy updated at node i in the previous time step. This represents the abnormal residual energy of downstream node j at the next time step. Let be the propagation attenuation coefficient from node i to node j.

[0040] Step 1: t4→t3 (18:00→16:00); Warehouse A is the upstream node of distributor C. The energy input to warehouse A at time t3 is: ; The abnormal residual energy of repository A after the update: ; Step 2: t3→t2 (16:00→14:00); Producer D is the upstream node of warehouse A. The energy input to producer D at time t2 is: ; The abnormal residual energy of producer D after the update: ; Step 3: t2→t1 (14:00→12:00); Supplier E is an upstream node of producer D. The energy supplied by supplier E at time t1 is: ; The abnormal residual energy of supplier E after the update: ; Step 4: t1→t0 (12:00→10:00); Supplier E has no upstream node at time t0, the incoming energy is 0, and the abnormal residual energy remains at its initial value. constant.

[0041] In this embodiment, the propagation attenuation coefficient is dynamically calculated as follows: the propagation attenuation coefficient is equal to the basic transmission coefficient from the upstream entity to the downstream entity multiplied by the exponential attenuation value of the product of the propagation attenuation adjustment factor and the standard deviation of transportation timeliness. Specifically, the value of the basic transmission coefficient is pre-calibrated based on the influence of the upstream to the downstream in the supply chain. The calibration method is to obtain the ratio of the covariance of the upstream state change and the downstream state change in the historical data of the upstream and downstream of the supply chain to the variance of the upstream state change. The standard deviation of transportation timeliness refers to the standard deviation of the logistics delivery time or information transmission delay between the upstream entity and the downstream entity within a statistical window of a preset duration before the current moment. In the calculation of the incoming energy, if the same upstream node transmits non-zero energy to the current node at multiple consecutive moments, the incoming energy sequence is smoothed by an exponentially weighted moving average method before being accumulated. By dynamically calculating the propagation attenuation coefficient by decomposing it into two parts: a basic transmission coefficient and an exponential attenuation adjustment factor, the basic transmission coefficient is calibrated based on the moving average of the ratio of the covariance of historical state changes in the upstream and downstream of the supply chain to the variance of upstream state changes. This objectively quantifies the inherent causal influence between upstream and downstream nodes, avoiding the subjective errors caused by manual static calibration. At the same time, an exponential attenuation adjustment factor based on the standard deviation of transportation timeliness is introduced, which can reflect the impact of dynamic factors such as logistics and distribution fluctuations and information transmission delays on the intensity of anomaly propagation in real time. This solves the core problem that traditional static propagation coefficients cannot adapt to the dynamic changes in the supply chain environment. In addition, for scenarios where the same upstream node transmits non-zero energy to the current node at multiple consecutive moments, an exponentially weighted moving average is used to smooth the transmitted energy sequence. This effectively filters random noise and instantaneous abnormal fluctuations in the energy propagation process, preventing deviations in tracing results and misjudgments of root causes caused by single abnormal data. This significantly improves the accuracy, robustness, and environmental adaptability of the recursive perturbation backpropagation algorithm for anomaly tracing. It also provides more accurate anomaly propagation intensity and path data for the state inference unit, further ensuring the reliability and relevance of subsequent supply chain state predictions.

[0042] Specific examples: Unified preset parameters: the basic transmission coefficient calibration window is the past 30 days, the transportation timeliness statistics window is the past 7 days, the propagation attenuation adjustment factor γ=0.05 / hour, the exponential weighted moving average smoothing coefficient η=0.6, the supply chain link is supplier E→manufacturer D→warehouse A→distributor C, and the time slice interval is 2 hours.

[0043] Example 1: Dynamic calculation of propagation attenuation coefficient; The general formula for calculating the propagation attenuation coefficient is as follows: ; in Let i be the basic transfer coefficient from upstream node i to downstream node j. Let be the standard deviation of transportation time from i to j.

[0044] Step 1: Calibration of basic transfer coefficient; Formula for calculating basic transfer coefficient: ; in, This is the daily output / shipment status change sequence of upstream node i. This is the daily receiving volume / inventory volume status change sequence for downstream node j. Let the covariance of the two be , This represents the variance of the upstream state changes.

[0045] Taking supplier E to manufacturer D as an example, the variance of the daily shipment volume change series of supplier E over the past 30 days: Covariance between supplier E's shipments and manufacturer D's receipts ; Base transfer coefficient after moving average: ; Similarly, the calculation yields: Basic transfer coefficient from manufacturer D to warehouse A ; Basic transfer coefficient from warehouse A to distributor C ; Step 2: Calculate the standard deviation of transportation time; taking supplier E to manufacturer D as an example, the actual transportation time (in hours) for the past 7 days is: 12, 14, 11, 13, 15, 12, 13; calculate the average: Hours; Standard deviation of delivery time: Hour Similarly, the calculation yields: Standard deviation of transportation time from manufacturer D to warehouse A Hour; Standard deviation of delivery time from warehouse A to distributor C Hour; Step 3: Final Propagation Attenuation Coefficient Calculation from Supplier E to Manufacturer D: ; From manufacturer D to warehouse A: ; From warehouse A to distributor C: ; Example 2: Exponentially weighted moving average smoothing of continuously supplied energy; When the same upstream node receives non-zero energy at multiple consecutive moments, the smoothing formula is as follows: ; in, Let be the smoothed input energy at time t. The initial value is the original input energy at time t. .

[0046] Assume that the initial energy received by warehouse A from producer D at three consecutive times t1, t2, and t3 are respectively , , : Smoothed energy at time t1: ; Smoothed energy at time t2: ; Smoothed energy at time t3: ; The unsmoothed cumulative input energy is 0.32+0.55+0.41=1.28, and the smoothed cumulative energy is 0.192+0.4068+0.4087=1.0075. This effectively filters out the instantaneous energy peak at time t2 and avoids the abnormal source tracing results from shifting excessively to a single time.

[0047] In this embodiment, the temporal graph neural network in the state inference unit adopts a message-passing architecture based on anomaly residual weighted coding. The process of updating the hidden state of a node at each prediction time step is as follows: the hidden state of the current node at the previous time step and the aggregated neighbor messages are input into the gated recurrent unit to obtain the new hidden state at the current time step. The aggregated neighbor messages are obtained by performing the following operations on each neighbor node and then summing them: first, the attention coefficient between the current node and the neighbor nodes is calculated; then, the hidden state of the neighbor node, the current edge feature vector, and the anomaly residual energy encoding vector of the neighbor node at the current time step are concatenated; then, the result is multiplied by a learnable weight matrix; and finally, the result is multiplied by the attention coefficient. The anomaly residual energy encoding vector is obtained by normalizing the anomaly residual energy output by the graph replay unit and concatenating it with the historical anomaly energy change rate. The historical anomaly energy change rate refers to the moving average of the first-order difference values ​​of the anomaly residual energy of the node over multiple consecutive time steps in the past. The prediction output layer maps the final hidden state to the probability distribution of each state category at the next time step. By designing a time-series graph neural network message-passing architecture based on anomaly residual weighted encoding, the anomaly residual energy encoding vector is deeply integrated into the neighbor message aggregation process. This enables the model to proactively perceive anomalies that have occurred in the supply chain and their propagation intensity, addressing the core deficiency of traditional time-series forecasting models that rely solely on historical state data and cannot capture the impact of anomaly propagation on future states. Furthermore, an attention mechanism is introduced to dynamically calculate the influence weights between nodes, adaptively distinguishing the importance of different upstream and downstream nodes to the target node, avoiding prediction bias caused by fixed weights. Gated recurrent units are used to implement the time-series update of node hidden states, effectively preserving the long-term nature of the supply chain state. The architecture establishes a time-series dependency relationship and concatenates the normalized value of the abnormal residual energy with the historical abnormal energy change rate to generate an encoding vector. This not only quantifies the severity of the current abnormality but also reflects its development trend, further enhancing the model's ability to model the abnormal evolution process. The prediction output layer maps the final hidden state to a state category probability distribution, providing a direct probabilistic basis for subsequent confidence calculations. The entire architecture achieves a deep integration of supply chain topology, time-series evolution patterns, and abnormal propagation characteristics, significantly improving the accuracy of future entity state predictions and the sensitivity to abnormal events. This provides reliable technical support for supply chain risk early warning and decision-making.

[0048] Specific examples: Unified preset parameters: Hidden layer dimension d=64, edge feature vector dimension d e =16, dimension d of the abnormal residual energy encoding vector v =8, the attention mechanism can learn parameter vectors. Learnable weight matrix The input dimension of the gated recurrent unit (GRU) is d, the hidden layer dimension is d, the current time step is t, the target node is warehouse A (node ​​i), and its neighboring nodes include supplier E (node ​​j1), manufacturer D (node ​​j2) and distributor C (node ​​j3). The minimum global abnormal residual energy is min(E)=0, and the maximum value is max(E)=0.61.

[0049] Example 1: Generation of anomaly residual energy encoding vectors; The anomaly residual energy encoding vector is formed by concatenating the normalized current anomaly residual energy with the historical anomaly energy change rate. The calculation formula is as follows: ; in, Let be the minimum-maximum normalized value of the abnormal residual energy at node j at time t. It is the moving average of the first difference values ​​of the abnormal residual energy of node j over the past three consecutive time steps (historical abnormal energy change rate).

[0050] Known abnormal energy data for each node: Node j1 (Supplier E): Historical anomalous energy sequence [0.02, 0.03, 0.40], first difference value [0.01, 0.37], moving average ; Node j2 (Producer D): Historical anomalous energy sequence [0.04, 0.04, 0.49], first difference value [0.00, 0.45], moving average ; Node j3 (Reseller C): Historical anomalous energy sequence [0.00, 0.00, 0.00], first difference value [0.00, 0.00], moving average ; Calculate the normalized value: , , ; Generate encoded vectors (simplified here as a 2D concatenation example, actually a vector expansion of the corresponding dimension): .

[0051] Example 2: Neighbor message aggregation process; For each neighbor node j, first calculate the attention coefficient between it and the target node i. Calculation formula: ; in, Let i and j be the hidden states of nodes i and j in the previous time step, respectively. Let (i,j) be the eigenvector of edge (i,j). Let i be the set of neighbors of node i, and || denote the vector concatenation operation.

[0052] Assume that the attention coefficients of each neighboring node are obtained through calculation: ; Then calculate the messages that each neighbor node passes to the target node. : ; Suppose that the average values ​​of the vector elements obtained after the concatenated feature vectors of each neighbor node are transformed by W, respectively: , , The average value of the elements passed between the neighbors is: ; Aggregated neighbor messages The average value of elements: ; Example 3: Node hidden state update and predicted output; The hidden state at a given time step of the target node. and aggregated neighbor messages Input GRU to get the new hidden state at the current time step. GRU core update formula: ; ; in, To update the door, To reset the door, is the sigmoid activation function, and ⊙ is the element-wise product.

[0053] Assumption: The hidden state in the previous time step average of elements Update Gate average of elements Candidate hidden state average of elements ; The average value of the hidden elements at the current time step: ; Ultimately, the prediction output layer will The probability distribution of the inventory status of warehouse A at the next time step (t+1) is mapped as follows: normal inventory (0.72), insufficient inventory (0.25), and stockout (0.03).

[0054] In this embodiment, the temporal graph neural network introduces a self-supervised training strategy based on time window sliding sampling during the training phase: the historical evolution sequence of the temporal knowledge graph is divided into several training samples according to a fixed-length time window. For the data in each time window, the node states of some time segments within that time window are randomly occluded. The temporal graph neural network is required to predict the occluded state value based on the unoccluded historical information within the time window and the abnormal residual energy before the occluded segment. The training loss function consists of a weighted sum of the prediction cross-entropy loss and the Euclidean distance loss between the predicted state and the true state. This approach proposes a self-supervised training strategy based on sliding time window sampling. It can train a time-series graph neural network without relying on a large number of manually labeled anomaly samples, effectively addressing the industry pain points of poor model generalization caused by the scarcity of anomalies, high labeling costs, and insufficient labeled data in the supply chain field. By dividing the historical evolution sequence of the time-series knowledge graph into training samples using a fixed time window, it can fully utilize massive amounts of unlabeled historical operational data to maximize the mining of hidden patterns in the time-series data. Simultaneously, it introduces the residual energy of anomalies before occlusion segments as input features during training, forcing the model to learn the propagation patterns of anomalies and their impact on... The influence mechanism of future states enables the model to have anomaly perception and prediction capabilities during the training phase, rather than just learning the temporal patterns of normal states. In addition, the combined loss function of weighted summation of prediction cross-entropy loss and Euclidean distance loss is adopted, which takes into account the accuracy requirements of discrete state classification prediction and continuous state numerical regression. This can effectively improve the model's performance in different types of state prediction tasks. The entire training strategy can be completed automatically without manual intervention, which greatly reduces the cost and cycle of model training, while significantly improving the model's generalization ability and the sensitivity to predicting abnormal events. It provides high-precision and high-robust prediction model support for the state inference unit.

[0055] Specific examples: Unified preset parameters: The total duration of the historical sequence of the time-series knowledge graph is 90 days, and the time window length is... Hours, time step Hours, each time window contains Each time step, sliding step size Hourly random occlusion ratio p=0.25, cross-entropy loss weights Euclidean distance loss weight The node status includes discrete categories (normal=0, warning=1, abnormal=2) and continuous values ​​(inventory quantity). The global abnormal residual energy range is [0,0.61]. The target training node is warehouse A (node ​​i).

[0056] Example 1: Time window sliding sampling and training sample construction; Core formula: ; in, For the time window interval corresponding to the k-th training sample, Let k be the starting time of the k-th time window, and let the starting time of the historical sequence be the starting time. Based on the reference, increment by sliding step size, This is the end time of the k-th time window, with a fixed time window length between it and the start time. ; This is a fixed length for a single time window; in this example, it is 24 hours. This represents the sliding step size between adjacent time windows; in this example, the value is 6 hours. The total duration of the historical sequence of the time-series knowledge graph is given; in this example, 90 days corresponds to 2160 hours. K represents the total number of training samples generated, calculated by subtracting the window length from the total duration, dividing by the sliding step size, and then rounding down. (take no more than) The largest integer is calculated and then 1 is added to it.

[0057] Total number of training samples: ; The first three time windows generated: First sample: S1=[0,24] hours (from 00:00 to 24:00 on day 1); Second sample: S2=[6,30] hours (06:00 on day 1 to 06:00 on day 2); The third sample: S3 = [12, 36] hours (from 12:00 on day 1 to 12:00 on day 2); Example 2: Random occlusion strategy and input feature construction; Core formula: ; Where M is the set of occluded time step numbers, contained within a single time window. Randomly select a specified number of elements in each time step; X represents the total number of time steps within a single time window. In the example, there is one time step every 2 hours, for a total of 12 time steps in 24 hours. p is the random occlusion ratio, i.e., the proportion of occluded time steps out of the total number of time steps. In the example, the value is 0.25, which is calculated by rounding down to obtain 3 occluded time steps for each sample. 输入 The input feature set for the model consists of two parts: unoccluded features and pre-occlusion anomaly energy, where H...未遮挡 E represents the node states, neighbor node states, and graph structure features of all unoccluded time steps within the time window. 遮挡前 This is the sequence of anomalous residual energy for all time steps prior to the start of the masked segment.

[0058] Example calculation process: For the first time window S1=[0,24] hours (time step number 1-12), randomly select the set of time steps to be blocked: M={4,7,10}; that is, block the inventory status of warehouse A in the three time steps of the 8th hour, the 14th hour and the 20th hour.

[0059] The model input features include: Inventory status, neighbor node status, and graph structure characteristics at unoccluded time steps (1-3, 5-6, 8-9, 11-12); The anomalous residual energy sequence of all time steps (1-3) before the start time of the occlusion segment (8th hour, time step 4): [0.01, 0.02, 0.03]; Example 3: Calculation of the combined loss function; Core formula: ; ; in, The total loss value for a single training sample is obtained by weighted summation of cross-entropy loss and mean squared error loss; and These are the weighting coefficients for the cross-entropy loss and the Euclidean distance loss, respectively, and their sum is 1. In this example, they are 0.7 and 0.3, respectively. Cross-entropy loss for discrete state classification is used to measure the accuracy of discrete state prediction. The mean squared error loss for continuous numerical values ​​is equivalent to the average of the squared Euclidean distances and is used to measure the accuracy of continuous numerical predictions. C represents the total number of time steps that are occluded, which is 3 in the example; C represents the total number of discrete state categories, which includes 3 categories in the example: normal, warning, and abnormal. For the m-th occlusion time step, the c-th class is the true label, using one-hot encoding, with the position corresponding to the true class being 1 and the rest being 0; Let be the probability of class c at the m-th occlusion time step predicted by the model; log is the natural logarithm function. This represents the actual continuous state value at the m-th occlusion time step; in the example, it represents the inventory quantity. These are the continuous state values ​​predicted by the model at the m-th occlusion time step.

[0060] Example calculation process: Calculate cross-entropy loss Assuming the true discrete labels for the three occlusion time steps are normal, warning, and normal, the corresponding one-hot encoding is: , , The model predicts the probabilities as follows: , , ; ; Calculate the mean square error loss : The mean square error of the original inventory quantity is: ; To standardize the units of measurement, the inventory quantity was normalized to the [0,1] interval before the normalized mean square error loss was calculated. .

[0061] Calculate total loss : ; The model minimizes the total loss function using the gradient descent algorithm, continuously updating the network parameters until the loss converges.

[0062] In this embodiment, when the state inference unit outputs prediction results in order of confidence, the comprehensive confidence of each prediction result is obtained by weighted fusion of the following confidence components: The first confidence component is the normalized complement of the information entropy of the prediction probability distribution output by the time-series graph neural network, which is obtained by multiplying the ratio of the maximum probability value to the second largest probability value in the probability distribution by a correction factor based on historical prediction deviation. The correction factor is calculated by comparing the ratio of the average deviation of the model in the most recent predictions to a preset benchmark deviation. The second confidence component is the state stability score of the predicted entity node in the past statistical period, which is obtained by weighted summation of the node's state change frequency and the reciprocal of the state value variation coefficient, followed by a logistic transformation. The third confidence component is the model robustness score at the current prediction time step, which is obtained by normalizing the reciprocal of the average deviation between the prediction results in the prediction windows before and after the current prediction time step and the actual feedback at the corresponding time point, followed by a moving average. The weighting coefficients of the three confidence components are all preset values, and the sum of the three is one. By designing a multi-dimensional confidence component weighted fusion mechanism, the reliability of prediction results is comprehensively evaluated from three core dimensions: model output uncertainty, entity historical state stability, and model's own prediction robustness. This addresses the deficiency of traditional single confidence indicators in fully reflecting the reliability of prediction results. The first confidence component, based on the ratio of the maximum to the second-highest probability value of the prediction probability distribution and incorporating a historical prediction bias correction factor, quantifies the uncertainty of the model output itself and corrects for potential systematic biases that may occur during long-term model operation. The second confidence component calculates a state stability score based on the entity's state change frequency and coefficient of variation, reflecting the inherent stability of the predicted entity. The volatility characteristic avoids over-reliance on high-probability predictions of unstable entities. The third confidence component calculates the model robustness score based on the deviation between recent prediction results and actual feedback, reflecting the model's actual prediction performance in the current time window. The weighting coefficients of the three components are preset and sum to 1, allowing for flexible adjustment of the weight ratio of each dimension according to the needs of different supply chain scenarios. The final comprehensive confidence score provides an objective and unified standard for ranking prediction results, enabling managers to prioritize high-confidence prediction results, significantly improving the accuracy of risk warnings and decision-making efficiency. At the same time, it provides a clear confidence assessment basis for subsequent dynamic feedback calibration units.

[0063] In this embodiment, a dynamic feedback calibration unit is also included, which receives user confirmation or rejection operations from the remote monitoring unit. The unit uses the user-confirmed anomaly cause and the user-approved high-confidence prediction result as positive feedback samples, and the user-rejected low-confidence prediction result as negative feedback samples. It periodically performs online calibration and updates on the propagation attenuation coefficient in the map playback unit and the confidence weight in the state deduction unit: for the propagation attenuation coefficient, its basic transmission coefficient is updated by comparing the similarity between the user-confirmed anomaly cause and the system's automatic tracing result on the same propagation path; for the confidence weight, the fusion weight of the three confidence components is optimized by gradient ascent, aiming to maximize the ranking score of the user-approved high-confidence sample in the prediction result ranking list and minimize the ranking score of the user-rejected sample. By setting up a dynamic feedback calibration unit, a closed-loop feedback mechanism between user operations and system model parameters is established. This eliminates the need for additional manual annotation costs by generating positive and negative feedback samples from user confirmations or rejections during daily use. The propagation attenuation coefficient in the graph playback unit and the confidence weight in the state deduction unit are periodically calibrated and updated online. This effectively solves the core problem of traditional supply chain monitoring platforms where static model parameters are fixed and unable to adapt to dynamic changes in the supply chain environment, leading to a gradual decline in system performance over time. Specifically, regarding the propagation attenuation coefficient, the basic transmission coefficient is updated by comparing the similarity between user-confirmed anomaly causes and the system's automatic tracing results along the same propagation path. This continuously optimizes the quantitative accuracy of anomaly propagation intensity and improves the accuracy and root cause determination of anomalies. The system's positioning capability, specifically targeting confidence weights, employs gradient ascent optimization by maximizing the ranking score of high-confidence samples accepted by users in the prediction result ranking list and minimizing the ranking score of user-rejected samples. This allows for dynamic adjustment of the weight ratios of the three confidence components, making the calculation of the overall confidence score more aligned with actual business scenarios and user judgment standards. This ensures that highly reliable prediction results are prioritized for display, significantly reducing information screening costs and decision-making errors for management personnel. Furthermore, the entire calibration process does not interrupt normal system operation, enabling continuous self-evolution and iterative optimization of system performance. As usage time increases, the system's anomaly tracing accuracy and state prediction reliability continuously improve, significantly extending the platform's effective lifespan and providing continuous and reliable technical support for the long-term stable operation of the supply chain.

[0064] In this embodiment, the remote monitoring unit also includes a multimodal interactive anomaly review interface. When displaying anomaly propagation paths, this interface uses timeline folding technology to merge nodes with gradual state changes in multiple adjacent time steps, displaying the complete time step sequence only when the user clicks to expand. When the user selects a root cause node and issues a "simulate mitigation" command, the system calls the state inference unit to correct the state value of the root cause node to the simulated value specified by the user, then re-executes future state prediction, and displays the difference between the original prediction and the adjusted simulation prediction in the form of a heatmap. By integrating the multimodal interactive anomaly review interface into the remote monitoring unit, the entire interactive design enhances the collaboration between the user and the system, making the monitoring platform not only a data display tool but also a core platform supporting supply chain risk management decisions.

[0065] In this embodiment, a distributed graph storage module is also included. The distributed graph storage module stores the time-series knowledge graph according to a hybrid sharding strategy of effective time sharding and entity type hashing: a preset natural time unit is used as the first sharding key for effective time. For long-term effective states that exceed the duration of the first sharding key, a separate long-term storage area is allocated and stored using the transaction time as the second sharding key. Within each effective time shard, the snapshot index of the start and end times of the shard and the differential change log between each smallest time granularity are stored simultaneously. The differential change log is compressed and stored by using lossy run-length encoding to compress the entity state difference and relationship change list between consecutive time steps. When responding to cross-shard queries, the graph replay unit reads each involved storage shard in sequence. First, it performs a binary search in the snapshot index of each shard to locate the start time of the target time interval. Then, it replays the differential change log one by one according to the time order between the start time and the end time of the target time interval to complete the generation of the dynamic evolution sequence. By designing a distributed graph storage strategy that combines effective time sharding with entity type hash sharding, the entire storage architecture achieves efficient storage, fast retrieval, and low-cost expansion of time-series knowledge graphs. It provides high-performance data support for historical subgraph extraction in the graph playback unit, anomaly tracing analysis, and real-time data acquisition in the state inference unit, ensuring the stable operation and rapid response capability of the entire platform under massive time-series data.

[0066] This invention achieves reliable fusion of multi-source data and full-link temporal state modeling of the supply chain based on temporal knowledge graphs. It accurately traces the propagation path and root cause of anomalies through a recursive perturbation backpropagation algorithm. Combined with a temporal graph neural network weighted by anomaly residuals, it achieves highly reliable state prediction with multi-dimensional confidence assessment. At the same time, it introduces a user closed-loop feedback mechanism to realize online self-calibration of system parameters. With a multimodal interactive decision support interface and an efficient distributed graph storage architecture, it comprehensively solves the pain points of traditional supply chain monitoring, such as multiple data conflicts, difficulty in anomaly tracing, low prediction accuracy, and easy degradation of system performance. It significantly improves the intelligence level and risk management efficiency of remote supply chain monitoring.

[0067] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0068] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based real-time data acquisition and remote monitoring platform for multi-source supply chain data, characterized in that: include: The multi-source data acquisition unit is used to collect raw data, including transaction timestamps, from each node of the supply chain in real time. The raw data includes inventory status, logistics trajectory, production progress, and transaction records. The temporal knowledge graph construction unit is used to label each collected raw data with a valid time interval and transaction timestamp, and to calculate the multi-source fusion trust degree for each entity's state attribute, forming a temporal knowledge graph with dual time dimensions and trust degree labels; The graph playback unit is used to respond to playback requests, extract historical subgraph sequences from the time-series knowledge graph according to a specified time interval, and calculate the abnormal residual energy of the entity state at each historical moment in reverse along the time axis using a recursive perturbation backpropagation algorithm to obtain the abnormal propagation path and its energy accumulation value. The state inference unit is used to predict the entity state at multiple future times based on the real-time collected current state data and the abnormal propagation path output by the graph playback unit, by using the abnormal residual energy as a weighted encoding input to the time-series graph neural network, and dynamically calculating the comprehensive confidence of each prediction result based on the confidence component, and outputting it after sorting by confidence. The remote monitoring unit is used to visualize the evolution process, anomaly propagation path, and prediction results with confidence ranking of the time-series knowledge graph.

2. The knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform according to claim 1, characterized in that, The specific method for calculating the multi-source fusion trust level for each entity's state attribute in the temporal knowledge graph construction unit is as follows: For entity state attributes, when they receive observations from multiple heterogeneous data sources and their corresponding source trust levels within the same valid time interval, the pairwise consistency deviation matrix between each observation is calculated. If the absolute value of the deviation between any two observations is less than the preset consistency threshold, the weighted average value with the source trust is used as the fusion state value. The fusion trust is obtained by multiplying the weighted sum of each source trust by the time decay factor. The time decay factor is calculated based on the time difference between the data generation time and the current system fusion processing time, and the preset decay rate is a fixed value. If the deviation between observation pairs exceeds the consistency threshold, a conflict detection mechanism is triggered. By comparing the deviation between the historical fusion value of the entity's state attribute and the current observations, the observation with the smallest deviation from the historical fusion value and a source trust level higher than the dynamic trust elimination threshold is selected as the trusted source. The state value is updated only with this single source data, and the fusion trust level is the product of the single source trust level and the time decay factor.

3. The knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform according to claim 2, characterized in that, The recursive perturbation backpropagation algorithm in the spectrum playback unit specifically includes: For each time slice in the historical subgraph sequence, the abnormal residual energy of each entity node is calculated. The abnormal residual energy is equal to the actual state confidence of the entity node at the current moment minus the expected normal confidence benchmark value of the entity node at the same moment, and the larger value is taken compared with zero. The expected normal confidence benchmark value is obtained by taking the historical confidence moving average of the entity at the same moment in the past statistical period and adjusting it according to the dynamic standard deviation. The time slices are arranged in reverse chronological order, and the abnormal residual energy is propagated from back to front. For the propagation process from the current time to the previous time, the updated abnormal residual energy of the previous time is obtained by summing the original abnormal residual energy of the previous time with the incoming energy from the subsequent time. The incoming energy is the abnormal residual energy of all upstream nodes that have an incoming edge relationship with the current node in the subsequent time, multiplied by the propagation attenuation coefficient on the corresponding edge and then summed. The propagation attenuation coefficient is calculated based on the standard transmission coefficient of the upstream and downstream relationship of the supply chain and the standard deviation of the transportation timeliness corresponding to the relationship at the current time. After all time-propagation calculations are completed, the abnormal residual energy values ​​are sorted from largest to smallest. The nodes whose cumulative energy percentage exceeds the preset energy threshold and their corresponding historical state records are taken as the reasons for the anomalies.

4. The knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform according to claim 3, characterized in that, The propagation attenuation coefficient is dynamically calculated in the following way: The propagation attenuation coefficient is equal to the basic transmission coefficient from the upstream entity to the downstream entity multiplied by the exponential attenuation value of the product of the propagation attenuation adjustment factor and the standard deviation of transportation timeliness. Specifically, the value of the basic transmission coefficient is pre-calibrated based on the influence of the upstream on the downstream in the supply chain. The calibration method is to take the moving average of the ratio of the covariance of the upstream state change and the downstream state change in the historical data of the upstream and downstream of the supply chain to the variance of the upstream state change. The standard deviation of transportation timeliness refers to the standard deviation of the logistics delivery time or information transmission delay between the upstream and downstream entities within a statistical window of a preset duration before the current moment. In the calculation of the incoming energy, if the same upstream node transmits non-zero energy to the current node at multiple consecutive moments, the incoming energy sequence is smoothed by an exponentially weighted moving average method before being accumulated.

5. The knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform according to claim 4, characterized in that, The temporal graph neural network in the state inference unit adopts a message-passing architecture based on anomaly residual weighted coding. Its hidden state update process for nodes at each prediction time step is as follows: The hidden state of the current node at the previous time step and the aggregated neighbor messages are input into the gated recurrent unit to obtain the new hidden state at the current time step. The aggregated neighbor messages are obtained by performing the following operations on each neighbor node and then summing them: first, calculate the attention coefficient between the current node and the neighbor nodes; then, concatenate the hidden state of the neighbor nodes, the current edge feature vector, and the abnormal residual energy encoding vector of the neighbor nodes at the current time step; then multiply by the learnable weight matrix; and finally multiply by the attention coefficient. The abnormal residual energy encoding vector is obtained by normalizing the abnormal residual energy output by the graph replay unit and concatenating it with the historical abnormal energy change rate, where the historical abnormal energy change rate refers to the moving average of the first-order difference values ​​of the abnormal residual energy of the node over multiple consecutive time steps in the past. The prediction output layer maps the final hidden state to the probability distribution of each state category at the next time step.

6. The knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform according to claim 5, characterized in that, The time-series graph neural network introduces a self-supervised training strategy based on time window sliding sampling during the training phase: the historical evolution sequence of the time-series knowledge graph is divided into several training samples according to a fixed-length time window. For the data in each time window, the node states of some time segments within that time window are randomly occluded. The time-series graph neural network is required to predict the occluded state value based on the unoccluded historical information within the time window and the abnormal residual energy before the occluded segment. The training loss function consists of a weighted sum of the prediction cross-entropy loss and the Euclidean distance loss between the predicted state and the true state.

7. The knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform according to claim 6, characterized in that, When the state deduction unit outputs prediction results in order of confidence, the overall confidence of each prediction result is obtained by weighted fusion of the following confidence components: The first confidence component is the normalized complement of the information entropy of the predicted probability distribution output by the time series graph neural network. Specifically, it is obtained by multiplying the ratio of the maximum probability value to the second largest probability value in the probability distribution by a correction factor based on historical prediction deviation after performing a logistic transformation. The correction factor is calculated by comparing the ratio of the average deviation of the model in the most recent predictions to a preset benchmark deviation. The second confidence component is the state stability score of the predicted entity node in the past statistical period, which is obtained by taking the logistic transformation after weighted summation of the node's state change frequency and the reciprocal of the state value variation coefficient. The third confidence component is the model robustness score at the current prediction time step, which is obtained by normalizing the inverse of the average deviation between the prediction results in the prediction windows before and after the current prediction time step and the actual feedback at the corresponding time point, and then taking the moving average. The weighting coefficients of the three confidence components are all preset values ​​and the sum of the three is one.

8. The knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform according to claim 7, characterized in that, It also includes a dynamic feedback calibration unit, which is used to receive user confirmation or rejection operations from the remote monitoring unit, and use the user-confirmed anomaly reasons and the user-accepted high-confidence prediction results as positive feedback samples, and the user-rejected low-confidence prediction results as negative feedback samples. It periodically performs online calibration and updates on the propagation attenuation coefficient in the map playback unit and the confidence weight in the state inference unit: for the propagation attenuation coefficient, its basic transmission coefficient is updated by comparing the similarity between the user-confirmed anomaly reasons and the system's automatic tracing results on the same propagation path. For the confidence weights, the fusion weights of the three confidence components are optimized by gradient ascent, with the optimization objective of maximizing the ranking score of high-confidence samples recognized by users in the ranking list of prediction results and minimizing the ranking score of samples rejected by users.

9. The knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform according to claim 8, characterized in that, The remote monitoring unit also includes a multimodal interactive anomaly review interface. When displaying the anomaly propagation path, the multimodal interactive anomaly review interface uses time axis folding technology to merge nodes with smooth state changes in multiple adjacent time steps, and only displays the complete time step sequence when the user clicks to expand. When the user selects a root cause node and issues a "simulate mitigation" command, the system calls the state inference unit, corrects the state value of the root cause node to the simulated value specified by the user, and then re-executes the future state prediction, and compares and displays the difference between the original prediction and the simulation-adjusted prediction in the form of a heatmap.

10. The knowledge graph-based supply chain multi-source data real-time acquisition and remote monitoring platform according to claim 9, characterized in that, It also includes a distributed graph storage module, which stores the time-series knowledge graph according to a hybrid sharding strategy of effective time sharding and entity type hash: the preset natural time unit is used as the first sharding key for effective time, and long-term effective states that exceed the duration of the first sharding key are allocated a long-term storage area and stored with the transaction time as the second sharding key. Within each valid time slice, the snapshot index of the start and end times of the slice and the differential change log between each smallest time granularity are stored simultaneously. The differential change log is compressed and stored by using lossy run-length encoding to compress the entity state difference and relationship change list between consecutive time steps. When responding to a cross-shard query, the map playback unit reads each involved storage shard in sequence, first performs a binary search in the snapshot index of each shard to locate the start time based on the target time interval, and then replays the differential change log one by one according to the time order between the start time and the end time of the target time interval to complete the generation of the dynamic evolution sequence.

Images