Supply chain whole-link data tracing method and system based on multi-modal learning

Abstract

The present application relates to the technical field of data tracing, and in particular to a supply chain full-link data tracing method and system based on multi-modal learning, which solves the technical problem that the existing technology causes link misjudgment or root misjudgment in the tracing process due to the fact that the meta-path model only relies on single data. The method comprises the following steps: obtaining relative position parameters of multi-modal observation data in historical observation data distribution; analyzing the difference degree between the relative position parameters to determine an internal consistency index; analyzing the consistency of the change direction of the relative position parameters before and after the flow to determine a transition consistency index; analyzing the group consensus deviation degree of the relative position parameters to determine a credible weight, and correcting the internal consistency index and the transition consistency index to construct a path trust degree; determining a transition probability according to the path trust degree accumulated on the tracing path from the starting point to the upstream node, and determining a source node in combination with the frequency of the upstream node being accessed.

Description

Technical Field

[0001] This invention relates to the field of data traceability technology, specifically to a supply chain end-to-end data traceability method and system based on multimodal learning. Background Technology

[0002] End-to-end supply chain data traceability is a technical process that starts from the quality defects or credit default events of end products and traces the source of the problem step by step from distributors, core enterprises, first-tier suppliers to upstream levels. It plays a key role in cutting off the transmission of risks in a timely manner and maintaining the security of the supply chain.

[0003] Currently, the system mainly relies on transaction documents recorded in enterprise resource planning (ERP) and warehouse management systems. By constructing a directed graph of the supply chain and using a meta-path-based random walk model for risk path reasoning, the system uses the proportion of transaction amount as the edge weight to achieve probabilistic inference of upstream risk sources.

[0004] However, the true commercial dependencies in the supply chain are not only reflected in transaction amounts, but are also deeply embedded in factors such as technological exclusivity, implicit legal constraints, and logistical timeliness. These factors exist in a multimodal and heterogeneous manner, including images, time-series sensor signals, and free text. Existing models rely solely on single transaction amounts, leading to the systematic underestimation or neglect of exclusive supplier paths with small but deeply interdependent transaction amounts, while generic supplier paths with large but easily replaceable transaction amounts are highly likely to dominate the tracing direction, resulting in incorrect ordering of tracing paths and misjudgment of root causes. Summary of the Invention

[0005] To address the technical problem of existing technologies relying solely on single data points in their path-tracking models, leading to missed links or misjudgments of the root cause during the traceability process, this invention aims to provide a supply chain end-to-end data traceability method and system based on multimodal learning. The specific technical solution adopted is as follows: Firstly, a supply chain end-to-end data traceability method based on multimodal learning is provided, including: acquiring multimodal observation data corresponding to materials at each node of the supply chain, determining the relative position parameters of the observation data of each modality in the distribution of historical observation data; analyzing the degree of difference between the relative position parameters of the multimodality at each node, and determining the internal consistency index of each node; analyzing the consistency of the change direction of the relative position parameters of the common modality of the two ends of each material flow edge before and after the flow, and determining the transition consistency index of each material flow edge; analyzing the degree of deviation of the group consensus of the relative position parameters of each modality at each node, determining the credibility weight of each modality at each node, and correcting the internal consistency index and transition consistency index according to the credibility weight to construct the path trust degree of the walk path between nodes; performing a random walk upstream of the supply chain starting from the terminal node, determining the transition probability according to the path trust degree accumulated on the traceability path from the starting point to the upstream node, and determining the source node by combining the frequency of the upstream node being visited.

[0006] Based on the above technical solution, in the supply chain end-to-end data traceability method based on multimodal learning provided by this invention, the multimodal observation data of each node in the supply chain is uniformly transformed into parameter representations reflecting their historical relative positions, eliminating the obstacle of differences in physical dimensions and numerical ranges of different modalities to cross-modal comparisons. On this basis, cross-validation is performed from two dimensions: the internal consistency of each modal record within a node and the consistency of the transition direction of multimodal changes before and after material flow between adjacent nodes. A trustworthy weight is adaptively assigned to each modality by measuring the deviation of each modal record from the group consensus. This trustworthy weight is used to correct the indicators of nodes and edges, constructing the path trust degree of the traceability path. Finally, the path trust degree replaces the traditional transaction amount as the transition probability driving factor for random walks. This achieves multimodal data consideration, and through the built-in cross-modal cross-validation and adaptive weight allocation mechanism, the accuracy of the traceability direction can still be maintained by relying on the consensus of trustworthy modalities even when some modal data is tampered with or malfunctions, thus achieving precise positioning of the source nodes of the entire supply chain.

[0007] In conjunction with the first aspect above, in one possible implementation, the method for analyzing the consistency of the relative position parameters of the common modes of the two end nodes of each material flow edge before and after the flow, and determining the transition consistency index of each material flow edge, specifically includes: statistically analyzing the differences between the downstream end nodes and the upstream end nodes in each common mode of the target material flow edge, and determining the transition direction of the target material flow edge in each common mode based on the magnitude of the differences; analyzing the consistency quantification value of the transition directions in multiple common modes as the transition consistency index of the target material flow edge.

[0008] In conjunction with the first aspect above, in one possible implementation, the method for determining the transition direction of the target material flow edge in each common mode based on the magnitude of the difference between the two ends specifically includes: determining the equilibrium transition range of the relative position parameters in the target common mode based on the distribution of the difference between the downstream end node and the upstream end node in the target common mode from the perspective of historical observation data; if the difference between the two ends of the target common mode is greater than the upper limit of the equilibrium transition range, the transition direction of the target material flow edge in the target common mode is determined to be a positive transition; if the difference between the two ends of the target common mode is less than the equilibrium transition range... The lower limit value is used to determine that the transition direction of the target material flow edge in the target common mode is a negative transition; if the difference between the two ends of the target common mode is within the range of level transitions, the transition direction of the target material flow edge in the target common mode is determined to be a level transition; the method of analyzing the consistency quantification value of the transition direction in multiple common modes as the transition consistency index of the target material flow edge specifically includes: identifying the common modes corresponding to the transition directions that occur less frequently than level transitions in multiple common modes as rebellious modes; and performing a negative correlation mapping on the proportion of rebellious modes in multiple common modes to obtain the transition consistency index of the target material flow edge.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the method for analyzing the degree of difference between the relative position parameters of multimodalities on each node and determining the internal consistency index of each node specifically includes: calculating the absolute value of the difference between the relative position parameters of any two modalities on the same node to obtain the difference quantification value; taking the average value of the difference quantification values ​​between each pair of multimodalities on the same node as the internal contradiction index, and performing a negative correlation mapping on the internal contradiction index to obtain the internal consistency index.

[0010] In conjunction with the first aspect above, in one possible implementation, the method for analyzing the degree of deviation of the relative position parameters of each modality on each node from the group consensus and determining the trust weight of each modality on each node specifically includes: analyzing the deviation of the relative position parameters of each modality on the same node from the group distribution of the relative position parameters of other modalities; and obtaining the trust weight of the target modality on the target node based on the relative position of the deviation of the target modality on the target node in the deviation distribution of multiple nodes and multimodalities in the supply chain.

[0011] In conjunction with the first aspect above, in one possible implementation, the method for analyzing the deviation of the relative position parameters of each modality at the same node from the group distribution of relative position parameters of other modalities specifically includes: taking the median of the relative position parameters of other modalities (excluding the target modality) at the same node as the group position parameters; calculating the absolute value of the difference between the relative position parameters of the target modality and the group position parameters, and dividing it by the larger value between the relative position parameters of the target modality and the group position parameters to obtain the deviation of the target modality; the method for obtaining the credible weight of the target modality at the target node based on the relative position of the deviation of the target modality at the target node in the deviation distribution of multiple nodes in the supply chain specifically includes: taking the median of the deviation of multiple nodes in the supply chain as the deviation threshold; if the deviation of the target modality at the target node is less than the deviation threshold, then assigning a first credible weight to the target modality at the target node; otherwise, assigning a second credible weight; the first credible weight is greater than the second credible weight.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the method for constructing the path trust degree of the inter-node traversal path by modifying the internal consistency index and the transition consistency index based on the trust weight specifically includes: calculating the geometric mean of the trust weights of any two modes at each node; using the average of the geometric mean of the pairwise geometric mean of the multiple modes at each node as the node weight; weighting the internal consistency index of each node to obtain the weighted internal consistency index; weighting the transition consistency index based on the average of the trust weights of the common modes on the upstream nodes in each material flow edge to obtain the weighted transition consistency index; and multiplying the weighted internal consistency index of each node traversed sequentially by the inter-node traversal path and the weighted transition consistency index of each material flow edge after normalization to obtain the path trust degree of the inter-node traversal path.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, if the observation data of the first mode in the multimodal analysis is numerical data, the method for determining the relative position parameters of the observation data of the first mode in the distribution of historical observation data specifically includes: performing a unified mapping of the anomaly representation direction on the original observation data of the numerical data to obtain the mapped observation data; collecting the mapped historical observation data of the first mode on similar nodes in the supply chain; similar nodes are nodes in the supply chain that perform the same type of business function and have deployed the same type of sensors; and statistically analyzing the mapped observation data of the first mode on each node in the mapped historical observation data. The quantiles are used as the relative position parameters of the first mode at each node. If the observation data of the second mode in the multimodal data is non-numerical data, the method for determining the relative position parameters of the observation data of the second mode in the distribution of historical observation data specifically includes: collecting historical observation data of the second mode at similar nodes in the supply chain; extracting state features from the observation data of the second mode to obtain feature quantification indicators with the same anomaly representation direction after mapping with numerical data; and calculating the quantiles of the feature quantification indicators of the second mode at each node in the feature quantification indicators of historical observation data, which are used as the relative position parameters of the second mode at each node.

[0014] In conjunction with the first aspect mentioned above, in one possible implementation, the method further includes: negatively mapping the transition consistency index of the material flow edge to obtain the transition contradiction index; summing the internal contradiction index of the nodes on the traceability path and the transition contradiction index of the material flow edge, and dividing the sum by the number of nodes on the traceability path and the number of material flow edges to obtain the average suspicion density of the traceability path; arranging multiple traceability paths in descending order of average suspicion density, and outputting the path suspicion ranking result; the path suspicion ranking result is used to indicate the priority order of on-site inspection.

[0015] Secondly, a supply chain end-to-end data traceability system based on multimodal learning is provided, comprising: a data acquisition module for acquiring multimodal observation data corresponding to materials at each node of the supply chain and determining the relative position parameters of the observation data of each modality in the distribution of historical observation data; a node analysis module for analyzing the degree of difference between the relative position parameters of the multimodality at each node and determining the internal consistency index of each node; an edge analysis module for analyzing the consistency of the change direction of the relative position parameters of the common modality of the two ends of each material flow edge before and after the flow, and determining the transition consistency index of each material flow edge; a trust construction module for analyzing the degree of deviation of the group consensus of the relative position parameters of each modality at each node, determining the trust weight of each modality at each node, and correcting the internal consistency index and transition consistency index according to the trust weight to construct the path trust of the inter-node walking path; and a walking traceability module for performing random walks upstream of the supply chain starting from the terminal node, determining the transition probability based on the path trust accumulated on the traceability path from the starting point to the upstream node, and determining the source node by combining the frequency of the upstream node being visited.

[0016] The present invention has the following beneficial effects: By unifying multimodal observation data from each node in the supply chain into parameter representations reflecting their historical relative positions, the obstacles to cross-modal comparisons caused by differences in physical dimensions and numerical ranges of different modalities are eliminated. Based on this, cross-validation is performed from two dimensions: the internal consistency of each modal record within a node and the consistency of transitions in the direction of multimodal changes before and after material flow between adjacent nodes. Furthermore, a trust weight is adaptively assigned to each modality by measuring the deviation of each modal record from the group consensus. This trust weight is used to correct the indicators of nodes and edges, constructing a path trust score for the traceability path. Finally, the path trust score replaces the traditional transaction amount as the transition probability driver for random walks. This approach considers multimodal data and, through built-in cross-validation and adaptive weight allocation mechanisms, maintains the accuracy of the traceability direction based on the consensus of trustworthy modalities even when some modal data is tampered with or malfunctions, achieving precise positioning of the source nodes throughout the entire supply chain. Attached Figure Description

[0017] Figure 1 The following is a system architecture diagram of a supply chain end-to-end data traceability system based on multimodal learning, provided as an embodiment of the present invention; Figure 2 This is a flowchart illustrating a supply chain end-to-end data traceability method based on multimodal learning, as provided in one embodiment of the present invention. Detailed Implementation

[0018] The following description, in conjunction with the accompanying drawings, details the specific scheme of the supply chain end-to-end data traceability method and system provided by the present invention.

[0019] Please see Figure 1 The diagram illustrates a system architecture of a supply chain end-to-end data traceability system based on multimodal learning, provided by an embodiment of the present invention. This supply chain end-to-end data traceability system based on multimodal learning is a digital intelligent system for material quality traceability and risk positioning in all aspects of the supply chain. Relying on the core capabilities of multimodal data acquisition, cross-modal analysis, reliable quantification, and intelligent traceability, it breaks through the limitations of traditional traceability that relies on single structured data, and achieves accurate locking of the source nodes of the supply chain. It includes: data acquisition module 1, node analysis module 2, edge analysis module 3, trust building module 4, and wandering traceability module 5.

[0020] The data acquisition module 1 is the system's data source support module. Its core responsibility is to collect multimodal observation data corresponding to materials at each node of the supply chain, normalize heterogeneous data, and determine the relative position parameters of each modality's observation data within the historical observation data distribution. This module can acquire data through physical sensing devices such as industrial cameras, temperature and humidity sensors, triaxial vibration sensors, gas concentration detectors, vehicle positioning terminals, electronic locks, cold chain temperature probes, and loading / unloading area cameras, as well as the business servers corresponding to the Manufacturing Execution System, Warehouse Management System, and Quality Management System. Simultaneously, it utilizes an edge computing gateway to perform preliminary data cleaning, binding heterogeneous data such as images, time series, text, and trajectories of materials from the same batch using unified material batch codes and timestamps. Then, relying on the built-in quantile calculation unit, it performs unified mapping of anomaly representation directions for multimodal data of different dimensions and transforms them into relative position parameters of a unified scale. The relative position parameters of each node and modality output by the data acquisition module 1 are synchronously transmitted to the node analysis module 2, edge analysis module 3, and trust building module 4, providing basic data support for subsequent cross-modal analysis and trust quantification.

[0021] Node Analysis Module 2 is the core module for node data verification in the system. It is responsible for analyzing the differences in multimodal relative position parameters within a single node of the supply chain and determining the internal consistency index for each node. Deployed on a high-performance data processing server cluster or distributed computing workstation, this module receives the node multimodal relative position parameters transmitted by Data Acquisition Module 1. It then iterates through the relative position parameters of any two modalities within the same node, calculates the absolute value of their difference to obtain a quantified difference value, and averages all quantified difference values ​​to generate an internal consistency index characterizing the logical self-consistency of the multimodal data within the node. The internal consistency indices output by Node Analysis Module 2 are directly input into Trust Building Module 4 for subsequent weighted correction of node trustworthiness and path trust building.

[0022] Edge Analysis Module 3 is the core module for link data verification in the system. Its core responsibility is to analyze the consistency of the relative position parameters of the common modes of the nodes at both ends of the supply chain material flow edge before and after the flow, determining the transition consistency index for each material flow edge. This module relies on distributed computing nodes and a data comparison engine. After receiving the relative position parameters of the common modes of the nodes at both ends of the flow edge from Data Acquisition Module 1, it first calculates the difference between the modal parameters of the downstream node and the upstream node, and then determines the range of balanced transitions by combining the historical data difference distribution, thereby determining the transition direction of each mode. It then identifies the less frequently occurring defective modes and generates a transition consistency index characterizing the logical consistency of link data changes through a negative correlation mapping of the proportion of defective modes. The transition consistency indices of each material flow edge output by Edge Analysis Module 3 are synchronously pushed to Trust Building Module 4, providing crucial evidence for link trust correction and path trust calculation.

[0023] Trust Construction Module 4 is the core module for trust measurement in the system. It is responsible for analyzing the deviation of the relative position parameters of each modality within a node from the collective consensus, determining the modal trust weights, correcting node and link indicators, and constructing the path trust of the traversal paths between nodes. This module is deployed on a high-performance computing server equipped with a weight calculation engine and data caching units. After integrating the relative position parameters of each modality from Data Acquisition Module 1, the internal consistency indicators from Node Analysis Module 2, and the transition consistency indicators from Edge Analysis Module 3, it first calculates the deviation of each modality from the collective consensus and assigns high- and low-priority trust weights based on the global deviation distribution. Then, it corrects the internal consistency indicators using the geometric mean of the modal trust weights and the transition consistency indicators using the average of the upstream modal trust weights. Finally, it multiplies the weighted internal consistency indicators and the weighted transition consistency indicators within the path to generate the path trust. The path trust of each traceability path output by Trust Construction Module 4 is the core decision-making basis for the intelligent traceability implementation by the traversal traceability module 5.

[0024] The random walk traceability module 5 is the core module for risk tracing in the system. It is responsible for conducting random walks upstream of the supply chain, starting from terminal nodes. Based on the accumulated trust level of the path, it determines the transfer probability and, combined with the access frequency of upstream nodes, identifies the source node. This module integrates a random walk calculation engine, a traceability result output terminal, and a visualization display device. After receiving the path trust level of each path from the trust level construction module 4, it initiates multiple rounds of random walks from the abnormal terminal node, determining the transfer probability between nodes based on the normalized accumulated trust level of the path. Simultaneously, it counts the cumulative access frequency of each upstream node during the walk, identifying the upstream node with the highest frequency as the source node of the supply chain, and outputting the node information and suspicious link segment information of the source node. The traceability results generated by the random walk traceability module 5 can be directly connected to the on-site inspection terminal, providing precise guidance for supply chain risk verification and achieving the system's final traceability goal.

[0025] Please see Figure 2 The diagram illustrates a flowchart of a supply chain end-to-end data traceability method based on multimodal learning, according to an embodiment of the present invention. This multimodal learning-based supply chain end-to-end data traceability method includes: S1. Obtain multimodal observation data for materials at each node of the supply chain and determine the relative position parameters of the observation data for each modality in the distribution of historical observation data.

[0026] Multimodal observation data refers to heterogeneous data generated by devices with different acquisition principles during the material's residence time at supply chain nodes. This data reflects the state of materials or the environment of the nodes, encompassing both numerical and non-numerical data, such as temperature readings, vibration RMS values, images, and text. The dimensions and numerical ranges of different modal data vary significantly, making direct cross-modal comparisons impossible. Therefore, it is necessary to convert the data into a parameter form with a unified scale. The relative position parameter of each modality's observation data within the historical observation data distribution can be used to characterize the relative level of the current observation data with respect to the historical state of similar nodes, eliminating dimensional differences between different modes and enabling multimodal data to possess comparable state description capabilities.

[0027] Furthermore, numerical data includes positive anomaly data, negative anomaly data, and interval anomaly data. Positive anomaly data is characterized by larger values ​​indicating a higher degree of anomaly, negative anomaly data is characterized by smaller values ​​indicating a higher degree of anomaly, and interval anomaly data is characterized by deviations from the normal range (using industry-standard ranges). This invention requires a unified correspondence between the numerical value of the indicator and the degree of anomaly at supply chain nodes or materials (i.e., the direction of anomaly representation), eliminating cross-modal comparison conflicts caused by inconsistent anomaly directions, and ensuring that the relative position parameters of different modalities have logical comparability. This embodiment uses the anomaly representation direction where larger mapped values ​​indicate a higher degree of anomaly (positive anomaly) for illustration.

[0028] When collecting historical observation data, data from a single node is incomplete, and distribution discrepancies can easily arise between multiple nodes due to differences in business operations or equipment. Therefore, nodes of the same type are used. These nodes refer to nodes in the supply chain that perform the same type of business function and are equipped with the same type of sensors, such as nodes that are both warehouse nodes and are equipped with temperature and humidity sensors. This is used to ensure the consistency of the distribution of historical observation data.

[0029] The relative position parameter can be represented by quantiles, with values ​​ranging from [0, 1]. This invention uses the lower quantile, which can directly correspond to the actual state of the data. A lower quantile value close to 0 indicates that the current observation is at an abnormally low level in history, a value close to 1 indicates that it is abnormally high, and a value close to 0.5 indicates that it is in a typical intermediate state.

[0030] In some implementations, if the observed data of the first mode in a multimodal approach is numerical (such as transaction amount or material quantity), the original numerical data is first uniformly mapped to indicate the direction of anomalies, resulting in mapped observed data. Specifically, for positive anomalies, the original value is directly reused; for negative anomalies, a preset benchmark value (such as the upper limit of the historical normal range for similar nodes) is subtracted from the original value to generate the mapped value; and for interval anomalies, the absolute value of the original value minus the median of the normal range is used to generate the mapped value. Then, historical observed data of the first mode mapped to similar nodes in the supply chain is collected. The lower quantile of the mapped observed data of the first mode at each node is calculated within the mapped historical observed data, serving as the relative position parameter of the first mode at each node.

[0031] If the observation data of the second modality in a multimodal study is non-numerical, historical observation data of the second modality at similar nodes in the supply chain are collected. Since there is no direct way to represent the distribution of non-numerical data, a pre-defined state feature extraction algorithm is first used to extract state features from the observation data of the second modality, obtaining feature quantification indicators with the same anomaly representation direction as the numerical data. The state feature extraction algorithm is adapted to the modality type, ensuring that the feature quantification indicators are consistent with the anomaly representation direction of the numerical data. For example, for image modalities (such as images of material packaging, appearance of key components, and quality inspection labels), a preset image similarity algorithm is used to extract image difference indices. For instance, using a preset image of qualified materials as a standard template, the structural similarity (SSIM) between the currently captured image and the template is calculated, and then the image difference index is obtained by subtracting the structural similarity from 1. For text modalities (such as quality inspection reports entered into the production system, abnormal entry notes in the warehousing system, and problem records in the handover process), a preset word frequency statistics algorithm is used to extract abnormal word frequency indices (i.e., the frequency of abnormal keywords appearing in the text). The algorithm's setting rule is to extract quantifiable indicators that can stably reflect the degree of data abnormality, transforming unstructured, non-numerical data into a statistically significant numerical form. Then, the lower quantile of the feature quantification index of the second modality at each node is calculated in the feature quantification index of historical observation data, serving as the relative position parameter of the second modality at each node. This achieves a normalized representation of non-numerical multimodal data, eliminating the obstacle that unstructured data cannot directly participate in cross-modal analysis.

[0032] S2. Analyze the degree of difference between the relative position parameters of the multimodal modes on each node, and determine the internal consistency index of each node.

[0033] The material state at the same node is a single objective state. Under normal circumstances, the judgment of abnormal states of different modes should be consistent. Therefore, by analyzing the degree of difference between the relative position parameters of multiple modes, we can obtain internal contradiction indicators, which can directly reveal whether there are logical conflicts caused by equipment failure, transient anomalies or human tampering in the node data. Then, we can perform negative correlation mapping on the internal contradiction indicators to obtain internal consistency indicators.

[0034] In some implementations, the absolute value of the difference between the relative position parameters of any two modes on the same node is calculated to eliminate the influence of the positive and negative directions of the difference, retaining only the quantified result of the degree of divergence. This ensures that the contribution of state deviations in different directions to the degree of divergence is consistent, resulting in a difference quantification value. This difference quantification value is used to quantify the level of divergence between the two modes in their state judgment at the current moment, with a value range of [0, 1]. The larger the value, the greater the difference in the state judgments of the two modes. To further eliminate the influence of the difference in the number of modes connected to the node on the results and ensure the comparability of the index results between different nodes, the average of the difference quantification values ​​between each pair of multiple modes on the same node is used as the internal contradiction index.

[0035] For example, when the material does deteriorate, causing the relative position parameters of modes such as temperature, vibration, and image difference to be concentrated in the range close to 1, the difference quantification value of each pair of modes is close to 0. The internal contradiction index obtained by averaging is close to 0, indicating that the multimodal data within the node is highly self-consistent. When the node has contradictory combinations such as an intact image but a severely excessive temperature, or stable vibration but a red mark on the quality inspection report, the relative position parameters of different modes will deviate significantly, the difference quantification value is close to 1, and the internal contradiction index increases significantly after averaging, with the maximum value approaching 1. This directly reflects that there is a significant logical conflict in the data records within the node. The high value of the internal contradiction index can be used to indicate that the node has partial modal distortion caused by equipment failure, transient anomalies that are only captured by some sensors, or even cases of deliberate tampering with data of a certain mode.

[0036] Internal consistency indices for each node are determined by negatively correlated mapping of internal contradiction indices. These internal consistency indices are positive, representing the degree of agreement between multimodal data within a node; a higher internal contradiction index corresponds to a lower internal consistency index. The negative correlation mapping rule is that the difference between 1 and the internal contradiction index is used as the internal consistency index.

[0037] Specifically, if a node has only a single modality, then the internal contradiction index of that node is set to 0 and the internal consistency index is set to 1, indicating that in the absence of multimodal cross-validation, the node is assumed to have no data self-consistency conflict.

[0038] S3. Analyze the consistency of the relative position parameters of the common modes of the two ends of each material flow edge before and after the flow, and determine the transition consistency index of each material flow edge.

[0039] A material flow edge is a link in the supply chain where materials are transferred from upstream to downstream nodes. Adjacent upstream and downstream nodes are the two ends of this flow edge. Since modes existing only at a single end node cannot determine the direction of state changes before and after the flow, a common mode with complete records deployed at both ends is used for analysis. The direction of change of relative position parameters refers to the trend of change of relative position parameters within the same mode during the material flow from upstream to downstream, directly reflecting the drift of the material's state. By quantifying the consistency of the change direction of the common mode, a transition consistency index for the material flow edge is obtained. This transition consistency index complements the node's internal consistency index. The internal consistency index measures the spatial consistency between modes within a node, while the transition consistency index measures the temporal transition consistency between modes between nodes, thus locating the record reliability of each link in the supply chain from both static and dynamic dimensions.

[0040] In some implementations, the relative position parameters of the downstream node relative to the upstream node in each common mode of the target material flow edge are statistically analyzed (i.e., the relative position parameters of the downstream node minus the relative position parameters of the upstream node). The relative state drift of the mode from upstream to downstream is quantified, and the transition direction of the target material flow edge in each common mode is determined based on the magnitude of the difference between the two ends.

[0041] Specifically, based on historical observation data, the distribution of the difference between the downstream nodes and the upstream nodes in the target common mode is used to determine the range of balanced transitions in the target common mode. This embodiment uses the median of the absolute values ​​of the differences in the relative position parameters of all adjacent nodes in the entire supply chain as the threshold, setting the balanced transition range to [-threshold, threshold]. This setting is based on the historical fluctuation characteristics of the mode itself, avoiding the use of a uniform fixed threshold to determine the transitions of different modes, and adapting to the natural fluctuation amplitude differences of different modes. The median of the absolute values ​​of historical differences is chosen as the threshold because the differences in the modes of the supply chain flow edge exhibit a long-tailed skewed distribution and contain a few extreme outliers. The average value is easily skewed, leading to misjudgments. The median, as a statistical measure of the distribution center position, is not affected by extreme values, has strong robustness, and can stably reflect the typical fluctuation levels of most normal flows. It can also adaptively adapt to the fluctuation characteristics of different modes, without requiring manual preset hyperparameters, avoiding threshold distortion, and accurately distinguishing between normal fluctuations and significant transitions.

[0042] If the difference between the two ends of the target common mode is greater than the upper limit of the equilibrium transition range, the transition direction of the target material flow edge in the target common mode is determined to be a positive transition, indicating that the material state changes in the direction of increasing abnormality during the flow process; if the difference between the two ends of the target common mode is less than the lower limit of the equilibrium transition range, the transition direction of the target material flow edge in the target common mode is determined to be a negative transition, indicating that the material state changes in the direction of decreasing abnormality during the flow process; if the difference between the two ends of the target common mode is within the equilibrium transition range, the transition direction of the target material flow edge in the target common mode is determined to be an equilibrium transition, indicating that the material state does not exceed the normal fluctuation range and there is no significant change.

[0043] The consistency quantification value of transition directions in multiple common modes is analyzed and used as a transition consistency index for the target material flow edge. By statistically analyzing the distribution of transition directions in different modes, the degree of conflict of multimodal change directions within the flow edge is quantified, generating a core index of link reliability.

[0044] Specifically, common modes with fewer occurrences (excluding level transitions) among multiple common modes are identified as "rebellious modes." These modes contradict the generally accepted direction of change in most modes, and their presence lowers the credibility of the flow edge. This indicates that significant changes in this mode may be caused by sensor malfunction, data tampering, or localized, non-representative measurements, rather than a true change in the overall material. The proportion of rebellious modes among multiple common modes is negatively correlated. The rule for this negative correlation is to subtract the proportion of rebellious modes from 1, converting it into a quantitative value of consistency. This yields the transition consistency index of the target material flow edge, with a value range of [0, 1]. The closer to 1, the more consistent the multi-modal transition directions, and the more reliable the connection between records before and after the link. The closer to 0, the more severe the directional conflict, and the flow edge should be marked as a highly suspicious transition segment.

[0045] Specifically, if the frequency of positive and negative transitions is exactly the same across multiple shared modes, and there is no dominant transition trend, then all non-equilibrium transition modes are uniformly classified as rebellious modes. This reflects the completely chaotic and disordered nature of the internal state change trends within the flow link. If all shared modes are determined to be equilibrium transitions, with no positive or negative transitions, then it is directly determined that there are no rebellious modes within the current material flow edge. This indicates that the various states of the material maintain a normal fluctuation range during the flow process, and no substantial state change has occurred. The corresponding transition consistency index is 1. This processing method can supplement the judgment logic in extreme scenarios, avoid the problem of statistical interruption due to the lack of a dominant transition direction, and ensure that rebellious mode identification and transition consistency index calculation can be successfully completed in various flow scenarios, achieving comprehensive coverage of link consistency evaluation.

[0046] If the target material flow edge has no common mode, the transition consistency index of the target material flow edge is directly set to a preset fixed value that is lower than the normal value. The fixed value is used to characterize a conservative estimate of consistency that cannot be verified due to lack of evidence. It is greater than 0 and significantly lower than the typical value of normal consistency, so that the flow edge without common mode can still be visited in the random walk, but its transition probability is greatly weakened compared with the flow edge with positive consistency evidence, thereby balancing recall and precision. For example, it can be set to 0.1 or 0.2.

[0047] S4. Analyze the degree of deviation of the group consensus of the relative position parameters of each modality on each node, determine the trust weight of each modality on each node, and adjust the internal consistency index and transition consistency index according to the trust weight to construct the path trust degree of the traversal path between nodes.

[0048] The degree of deviation of group consensus refers to the magnitude of the deviation between the relative position parameters of a single modality and the group consensus state formed by all other modalities within the same node. It reflects the degree of conformity between the modality and the overall judgment of the multimodal system. The higher the value, the more isolated the modality is, and the more likely it is to have data anomalies caused by single-type data acquisition distortion, sensor operation failure, local acquisition environment interference, or single modality input error. It does not directly represent that there is a real anomaly in the node. Therefore, by quantifying the credibility weight of the corresponding modality, the consistency index of the node and the link is corrected, the interference of isolated low-credibility modalities on the overall trust judgment of the tracing path is weakened, and the acquisition failure noise and real risks are distinguished.

[0049] In some implementations, the deviation of the relative position parameters of each modality on the same node from the distribution of the relative position parameters of other modalities is analyzed to quantify the degree of deviation of a single modality from the multimodal consensus state within the node and to identify isolated abnormal modalities.

[0050] Specifically, the median of the relative position parameters of all modes other than the target mode at the same node is used as the group position parameter. When the number of remaining modes is odd, the value in the middle position after sorting is taken as the group position parameter; when the number of remaining modes is even, the average of the two middle values ​​after sorting is taken as the group position parameter. This group position parameter represents the consensus judgment of the multimodal groups within the node on the material state and is not affected by extreme abnormal modes. The absolute value of the difference between the relative position parameter of the target mode and the group position parameter is calculated to quantify the degree of relative deviation. Then, using the state with a higher level of abnormality in the scenario as a unified reference scale, the absolute value of the difference is divided by the larger value between the relative position parameter of the target mode and the group position parameter to obtain the deviation of the target mode. For example, if there are two sets of relative position parameters and group position parameters for the target modes, the first set being 0.5 and 0.3, and the second set being 0.3 and 0.1, with an absolute difference of 0.2 for both, the deviation for the first set is calculated as 0.2 divided by 0.5, and the deviation for the second set is calculated as 0.2 divided by 0.3, resulting in different deviation values. The first set of 0.5 represents a relatively high abnormal state level. A deviation of 0.2 at this level is only a small fluctuation under a high abnormal baseline, indicating that the disagreement between modes has a weak impact on the overall consensus, hence the smaller deviation value. The second set of 0.3 represents a relatively low normal state level. A deviation of 0.2 at a low abnormal baseline where the overall situation is relatively normal indicates a significant divergence in the two modes' judgments of the material's normal state. The data discrepancy accounts for a higher proportion of their own baseline state, and the mode's deviation from the group consensus is more pronounced, thus warranting a larger deviation value.

[0051] Specifically, when the larger value of the denominator is equal to zero, it means that the relative position parameter of the target mode is completely consistent with the group position parameter and is zero. That is, the group consensus data formed by the target mode and the other modes in the node is uniformly at the lowest level of the historical observation data distribution. The judgment results of all modes in the node on the material state are completely consistent without any discrepancy. At this time, the deviation corresponding to the target mode is directly fixed and assigned to zero.

[0052] Based on the relative position of the deviation of the target mode at the target node in the deviation distribution of multiple nodes in the supply chain, the credibility weight of the target mode at the target node is obtained, thereby amplifying the credibility mode and suppressing the isolated mode.

[0053] Specifically, the median deviation of multiple modalities across various nodes in the supply chain is used as the deviation threshold. This threshold is unaffected by extreme outliers and can stably reflect the typical cohesion level of most modalities. If the deviation of the target modality at the target node is less than the deviation threshold, it indicates that the relative position parameter of this modality is relatively close to the consensus state formed by the other modalities within the node. The data of this modality is highly consistent with the judgment results of the mainstream state within the node, and the data collection status is stable. In this case, a first credibility weight (a gain-type weight with a value greater than 1) is assigned to the target modality at the target node. Otherwise, it means that the modality deviates significantly from the group consensus, the data is highly likely to be distorted, and the reference value is low. A second credibility weight (a decay-type weight with a value less than 1) is assigned to it, and the first credibility weight is greater than the second credibility weight. The first credibility weight can amplify the role of high-credibility modalities in the calculation of internal consistency indicators and transition consistency indicators, and strengthen the dominant role of mainstream compliant and valid data in the evaluation results. Typical values ​​can be set to 1.5, 2.0, etc. The second credibility weight can compress the influence of modal data that deviates from the consensus and is prone to distortion, and reduce the interference of abnormal noise data on the overall evaluation results. Typical values ​​can be set to 0.3, 0.5, etc.

[0054] Specifically, if the target node has no other modalities besides the target modality, meaning the node has only a single modality, then a preset standard confidence weight of 1.0 is directly assigned to this modality. This indicates that the original evaluation validity of the modality data is maintained in the absence of a group reference, and the deviation calculation step for the current node is skipped. For single-modality nodes directly assigned a standard confidence weight, this confidence weight does not participate in subsequent node weight amplification or attenuation based on the geometric mean or arithmetic mean; it is only used to weight the internal consistency index in its original form. Since the internal consistency index of this node is also directly set to 1, the weighted result is still 1, thus ensuring that the node remains neutral due to the lack of cross-validation data and is not incorrectly evaluated as a high-confidence or low-confidence node.

[0055] In some implementations, the geometric mean (i.e., the square root of the product) of the confidence weights of any two modalities at each node is used as the comprehensive reliability assessment of the pairing of two modalities. The geometric mean can reflect the synergistic constraint relationship of the two weights. When the confidence weight of any pair of modalities is low, the overall confidence of the combination of the two will be reduced accordingly, which accurately matches the logic of joint judgment of the paired data. Compared with the arithmetic mean, which only performs linear numerical balancing and easily smooths out the difference in reference value brought about by the pairing of high and low weights, the use of the geometric mean can make the comprehensive reliability assessment of the pairing of modalities more in line with the actual needs of multimodal joint verification. It can distinguish the actual effectiveness of the three modal combinations of high confidence, high and low confidence, and low confidence.

[0056] By comprehensively considering the data reliability matching degree of any two modalities within a node, integrating the combined reliability levels among all modalities to form a unified overall weight, and using the average of the geometric mean between each pair of multiple modalities on each node as the node weight, the overall data reliability of the node can be balanced when correcting the internal consistency index. This amplifies the positive judgment effect brought by the collaboration of multiple high-reliability modalities, while smoothly diluting the data interference caused by scattered low-reliability modalities, making the overall reliability assessment result of the node more objective and fair.

[0057] The internal consistency index of each node is weighted according to the node weight, and the internal consistency index is multiplied by the node weight to obtain the weighted internal consistency index. This achieves the overall correction of the nodes, and the operation is simple and efficient, greatly reducing the difficulty of engineering deployment and real-time calculation.

[0058] In other implementations, the negative correlation mapping value (i.e., 1 minus the absolute value of the difference) of the relative position parameters can be weighted by the geometric mean of the confidence weights of any two modes on the node. The weighted negative correlation mapping values ​​between each pair of multimodal modes are then averaged to obtain the weighted internal consistency index. This amplifies the divergence of high-confidence modes in the sum, while weakening the divergence of low-confidence modes, which can improve the discrimination accuracy. However, this method has a relatively large computational cost.

[0059] Materials flow from upstream nodes to downstream nodes. The original data benchmark and initial state determination of the state transition changes of the flow edge are all based on the upstream nodes. The reliability of the state comparison before and after the link is dominated by the quality of the upstream data source. Therefore, the weight of the upstream node is used as the correction basis first. The transition consistency index is weighted according to the average value of the reliability weight of the common mode on the upstream node in each material flow edge. The transition consistency index is multiplied by the average value of the reliability weight of the common mode of the upstream to obtain the weighted transition consistency index. This weakens the impact of the low reliability mode of the upstream on the link consistency judgment, avoids the pollution of the judgment of the downstream node by the falsified data of the upstream node through cross-node comparison, and prevents the chain risk of the credibility of the entire link being distorted by a single falsification. The higher the value, the higher the consistency of the multimodal transition direction of the nodes at both ends of the link, and the more reliable the link record.

[0060] Because a gain-type credibility weight greater than 1 is set, the values ​​of the weighted internal consistency index and the weighted transition consistency index are not constrained by [0, 1]. Directly multiplying them together will result in an exponential increase in the path credibility after multiple high-credibility links are multiplied consecutively. This leads to excessively large differences in the magnitude of values ​​between traceability paths of different lengths and combinations, making it impossible to fairly judge the overall reliability of the path based on the magnitude of the path credibility value. Therefore, a normalization step is required. The weighted internal consistency index of each node traversed by the inter-node walking path and the weighted transition consistency index of each material flow edge are normalized and then multiplied together to obtain the path credibility of the inter-node walking path.

[0061] The normalization method uses the maximum-minimum linear normalization method, which pre-traverses all nodes and material flow links in the entire supply chain to obtain the maximum and minimum values ​​of the internal consistency index after global weighting, as well as the maximum and minimum values ​​of the global weighted transition consistency index.

[0062] After global linear normalization, both the weighted internal consistency index and the weighted transition consistency index converge to the [0, 1] interval. If any index is exactly 0 after normalization, direct multiplication will cause the entire path's trust level to drop to zero, making it impossible to distinguish the trust level differences between different severely anomalous paths and losing the ability to make detailed distinctions. After normalization and before the multiplication operation begins, a fixed minimum positive number is superimposed on all normalized indices, typically set to 10. -6 This value is extremely low and will not change the original order of the indicators or their relative reliability. It only serves as a safety net to prevent the product from reaching zero.

[0063] In the chain reaction process, the sum of the normalized weighted internal consistency index of each node along the material flow path and a fixed minimum positive number is multiplied by the sum of the normalized weighted transition consistency index from that node to the next node and a fixed minimum positive number. This results in the cumulative trust value of the entire path, i.e., the path trust level. This chain reaction structure implies that the path trust level is jointly determined by the node self-consistency and edge consistency of each link in the path. The longer the path and the more contradictory points, the lower the cumulative path trust level, thus truly reflecting the overall credibility of the entire traceability path.

[0064] S5. Starting from the terminal node, perform a random walk upstream in the supply chain. Determine the transfer probability based on the accumulated path trust along the traceability path from the starting point to the upstream node, and combine this with the frequency of upstream node visits to determine the source node.

[0065] Terminal nodes are the points in the downstream supply chain where abnormal events such as quality defects and credit defaults are reported, serving as the starting point for risk tracing. In the directed graph of the supply chain, the path search process, starting from a terminal node and proceeding along the material flow edge towards upstream supplier nodes, uses path sampling by a large number of random walks to probabilistically locate the source node. The transition probability refers to the probability that a random walk will choose the next upstream node as its next hop, determined by the accumulated trust level of the path; the probability is positively correlated with the path trust level. After a large number of random walks complete path sampling, the total number of times each upstream node is visited, and the frequency, are positively correlated with the probability that the node is the source node, allowing for the identification of the source node.

[0066] Specifically, a directed graph of the supply chain is pre-constructed, where nodes represent the various links in the supply chain, and directed edges represent material flow edges. The existence of edges is based on transaction relationships or material flow records. Then, a preset number of random walks are initiated. The preset number of walks is determined based on the scale of the supply chain nodes; for example, if the node scale is 1000, the number of walks is set to 10,000 to ensure the statistical significance of the sampling results. A convergence condition is set; for example, when the rate of change in the node visit frequency distribution in multiple consecutive walks is lower than a preset threshold (e.g., 1%), the walk is considered converged, and sampling stops.

[0067] Each wandering node starts at the terminal node and moves upstream along the material flow edges of the directed graph of the supply chain. At each step, when selecting the next upstream node, the path trust score for each upstream node is calculated based on the complete path from the starting point to the upstream node. In each wandering step, a candidate node is randomly selected from all upstream nodes with equal probability. A pseudo-random number is generated using a pseudo-random number generator, uniformly distributed on the [0, 1] axis. If the path trust score of the candidate node is greater than the pseudo-random number, the wandering node accepts the candidate node and completes the transfer; otherwise, the wandering node remains at the current node and terminates the current wandering. When the next wandering node begins its wandering, a new candidate node is selected, a pseudo-random number is generated, and the path trust score is determined to achieve either a transfer or termination. This mechanism uses the accumulated path trust score directly as the probability of the wandering node moving forward, allowing it to avoid path segments with sharp internal modal contradictions, intense directional conflicts at junctions, or highly isolated and suspected tampering modalities. Instead, it tends to continuously wander along path directions with highly consistent multimodal records and reliable data self-consistency.

[0068] After a large number of random walks converge, the frequency of each upstream supplier node being visited by the walker is counted. The node with the highest frequency is the most likely source node jointly locked by the multimodal evidence system.

[0069] Simultaneously, the specific locations where path trust levels experience a sharp drop are statistically analyzed along the walk path. These locations correspond to sections with sharp internal modal contradictions, intense directional conflicts at junctions, or highly isolated sections suspected of tampering, and are thus identified as high-suspicious verification locations. The method for determining a sharp drop in path trust level includes, for example, calculating the attenuation of path trust levels between adjacent nodes. This is calculated as the difference between the path trust level from the previous node and the path trust level from the current node, divided by the previous node's path trust level. This reflects the relative change in path trust level during path progression; a larger value indicates a more severe attenuation. A threshold for determining a sharp drop is set, for example, by iterating through the attenuation of path trust levels between adjacent links in all random walk paths and taking the 95th percentile of its distribution as the threshold. This threshold represents the upper limit of normal fluctuations in path trust level; attenuation exceeding this threshold is considered a sharp drop, marking the node as the sharp drop location. The corresponding node and material flow edge are then identified as high-suspicious verification sections.

[0070] In some implementations, the transition consistency index of the material flow edge is negatively correlated to obtain the transition contradiction index. The negative correlation mapping method is 1 minus the transition consistency index, which represents the degree of inconsistency in the multimodal transition direction between the two nodes of the material flow edge. It is consistent with the node internal contradiction index in terms of dimension and meaning. The larger the value, the more severe and suspicious the multimodal direction conflict of the flow edge is.

[0071] The total contradiction quantification index of the path is obtained by summing the internal contradiction index of nodes on the traceability path and the transition contradiction index of material flow edges. Simultaneously, the total number of evaluation objects is obtained by summing the number of nodes on the traceability path and the number of material flow edges. Dividing the total contradiction quantification index by the total number of evaluation objects yields the average suspicion density of the traceability path, representing the average contradiction level of each link on the path. The value range is fixed in the interval [0, 1]. A value closer to 1 indicates a denser distribution of contradictions in the link and a higher systemic risk; a value closer to 0 indicates better self-consistency of the link data and higher credibility.

[0072] By normalizing the total contradiction quantification index according to the total number of assessment objects, the ranking bias caused by differences in assessment objects across paths of different lengths is eliminated, enabling comparison of suspicion levels for paths of varying lengths on the same scale. All traceability paths corresponding to terminal nodes are arranged in descending order of average suspicion density, outputting the path suspicion ranking results to indicate the priority of on-site inspections and guide traceability personnel to prioritize node-by-node verification along highly suspicious paths. Furthermore, given limited computing resources, after obtaining the internal contradiction index and transition consistency index, steps S4 and S5 can prioritize calculating the nodes and material flow edges on traceability paths with higher average suspicion density.

[0073] Based on the above technical solution, by unifying the multimodal observation data at each node of the supply chain into parameter representations reflecting their historical relative positions, the obstacles to cross-modal comparisons caused by differences in the physical dimensions and numerical ranges of different modalities are eliminated. On this basis, cross-validation is performed from two dimensions: the internal consistency of each modal record within a node and the consistency of the transition direction of multimodal changes before and after material flow between adjacent nodes. Furthermore, a trust weight is adaptively assigned to each modality by measuring the deviation of each modal record from the group consensus. This trust weight is used to correct the indicators of nodes and edges, constructing a path trust degree for the traceability path. Finally, the path trust degree replaces the traditional transaction amount as the transition probability driving factor for random walks. This achieves multimodal data consideration, and through the built-in cross-validation and adaptive weight allocation mechanism, the accuracy of the traceability direction can still be maintained by relying on the consensus of trustworthy modalities even when some modal data is tampered with or malfunctions, thus achieving precise positioning of the source nodes throughout the entire supply chain.

Claims

1. A supply chain end-to-end data traceability method based on multimodal learning, characterized in that, include: Acquire multimodal observation data for materials at each node of the supply chain, and determine the relative position parameters of the observation data for each mode in the distribution of historical observation data; Analyze the degree of difference between the relative position parameters of the multimodal modes on each node, and determine the internal consistency index of each node; Analyze the consistency of the relative position parameters of the common modes of the two ends of each material flow edge before and after the flow, and determine the transition consistency index of each material flow edge. Analyze the degree of group consensus deviation of the relative position parameters of each modality on each node, determine the trust weight of each modality on each node, and correct the internal consistency index and the transition consistency index according to the trust weight to construct the path trust degree of the inter-node traversal path. Starting from the terminal node, a random walk is performed upstream in the supply chain. The transfer probability is determined based on the accumulated path trust along the traceability path from the starting point to the upstream node. The source node is determined by combining the frequency of upstream node visits.

2. The supply chain end-to-end data traceability method based on multimodal learning according to claim 1, characterized in that, Analyze the consistency of the relative position parameters of the common modes of the two endpoints of each material flow edge before and after the flow, and determine the transition consistency index of each material flow edge, including: The differences between the relative position parameters of the downstream node and the upstream node in each common mode of the target material flow edge are statistically analyzed, and the transition direction of the target material flow edge in each common mode is determined based on the magnitude of the differences between the two ends. The consistency quantification value of the transition direction in multiple common modes is analyzed and used as the transition consistency index of the target material flow edge.

3. The supply chain end-to-end data traceability method based on multimodal learning according to claim 2, characterized in that, The transition direction of the target material flow edge in each common mode is determined based on the magnitude of the difference between the two ends, including: Based on historical observation data, the distribution of the difference between the downstream end nodes and the upstream end nodes in the target common mode of multiple material flow edges is used to determine the range of level transitions of the relative position parameters in the target common mode. If the difference between the two ends of the target common mode is greater than the upper limit of the flat transition range, the transition direction of the target material flow edge in the target common mode is determined to be a positive transition. If the difference between the two ends of the target common mode is less than the lower limit of the equal transition range, the transition direction of the target material flow edge in the target common mode is determined to be a negative transition. If the difference between the two ends of the target common mode is within the range of level transition, the transition direction of the target material flow edge in the target common mode is determined to be a level transition. The consistency quantification value of transition directions in multiple common modes is analyzed and used as a transition consistency index for the target material flow edge, including: The common modes that occur less frequently, other than the level transitions, among multiple common modes are identified as rebellious modes; By negatively mapping the proportion of the rebellious mode among multiple common modes, a transition consistency index for the target material flow edge is obtained.

4. The supply chain end-to-end data traceability method based on multimodal learning according to claim 1, characterized in that, Analyze the degree of difference between the relative position parameters of the multimodal modes at each node, and determine the internal consistency index of each node, including: The absolute value of the difference between the relative position parameters of any two modes at the same node is used to obtain the difference quantification value; The average of the quantified differences between each pair of multimodal features at the same node is used as the internal contradiction index, and the internal contradiction index is negatively correlated to obtain the internal consistency index.

5. The supply chain end-to-end data traceability method based on multimodal learning according to claim 1, characterized in that, Analyze the degree of group consensus deviation of the relative position parameters of each modality on each node, and determine the credibility weight of each modality on each node, including: Analyze the deviation of the relative position parameters of each mode at the same node from the group distribution of the relative position parameters of other modes; The confidence weight of the target mode at the target node is obtained based on its relative position in the deviation distribution of multiple nodes and multimodal deviations in the supply chain.

6. The supply chain end-to-end data traceability method based on multimodal learning according to claim 5, characterized in that, Analyze the deviation of the relative position parameters of each mode at the same node from the group distribution of relative position parameters of other modes, including: The median of the relative position parameters of all modes other than the target mode at the same node is taken as the group position parameter; The deviation of the target mode is obtained by dividing the absolute value of the difference between the relative position parameter of the target mode and the group position parameter by the larger value between the relative position parameter of the target mode and the group position parameter. Based on the relative position of the deviation of the target mode at the target node in the deviation distribution of multiple nodes and multimodals in the supply chain, the credibility weight of the target mode at the target node is obtained, including: The median of the multimodal deviations across multiple nodes in the supply chain is used as the deviation threshold. If the deviation of the target mode on the target node is less than the deviation threshold, then a first confidence weight is assigned to the target mode on the target node; otherwise, a second confidence weight is assigned. The first confidence weight is greater than the second confidence weight.

7. The supply chain end-to-end data traceability method based on multimodal learning according to claim 1, characterized in that, Based on the trusted weights, the internal consistency index and the transition consistency index are adjusted to construct the path trust degree of the inter-node traversal path, including: The geometric mean of the confidence weights of any two modalities on each node is calculated. The average of the geometric mean of the pairwise geometric mean of the multimodalities on each node is used as the node weight. The internal consistency index of each node is weighted to obtain the weighted internal consistency index. The transition consistency index is weighted by the average value of the credible weights of the common modes on the upstream nodes of each material flow edge to obtain the weighted transition consistency index. The path trust degree of the inter-node traversal path is obtained by multiplying the weighted internal consistency index of each node passed through in sequence and the weighted transition consistency index of each material flow edge after normalization.

8. The supply chain end-to-end data traceability method based on multimodal learning according to claim 1, characterized in that, If the observation data of the first mode in a multimodal study are numerical data, determine the relative position parameters of the observation data of the first mode in the distribution of historical observation data, including: The original observation data of the numerical data are subjected to a unified mapping of anomaly representation directions to obtain the mapped observation data; Collect historical observation data of the first mode mapped to similar nodes in the supply chain; the similar nodes are nodes in the supply chain that perform the same type of business function and are equipped with the same type of sensors; The quantiles of the first mode's mapped observation data at each node are calculated in the mapped historical observation data and used as the relative position parameter of the first mode at each node. If the observation data of the second mode in a multimodal study are non-numerical data, determine the relative position parameters of the second mode's observation data within the historical observation data distribution, including: Collect historical observation data of the second modality at similar nodes in the supply chain; State features are extracted from the observation data of the second mode to obtain feature quantification indicators with the same anomaly representation direction after mapping with the numerical data; The quantile of the feature quantification index of the second mode at each node is calculated in the feature quantification index of the historical observation data, and used as the relative position parameter of the second mode at each node.

9. The supply chain end-to-end data traceability method based on multimodal learning according to claim 4, characterized in that, Also includes: The transition consistency index of the material flow edge is negatively correlated to obtain the transition contradiction index. The sum of the internal contradiction index of the node on the traceability path and the transition contradiction index of the material flow edge is calculated and divided by the sum of the number of nodes on the traceability path and the number of material flow edges to obtain the average suspicion density of the traceability path. Multiple tracing paths are arranged in descending order of average suspicion density, and the path suspicion ranking result is output; the path suspicion ranking result is used to indicate the priority order of on-site inspection.

10. A supply chain end-to-end data traceability system based on multimodal learning, characterized in that, include: The data acquisition module is used to acquire multimodal observation data corresponding to materials at each node of the supply chain and determine the relative position parameters of the observation data of each mode in the distribution of historical observation data. The node analysis module is used to analyze the degree of difference between the relative position parameters of the multimodal modes on each node and determine the internal consistency index of each node. The edge analysis module is used to analyze the consistency of the relative position parameters of the common modes of the two end nodes of each material flow edge before and after the flow, and to determine the transition consistency index of each material flow edge. The trust construction module is used to analyze the degree of deviation of the group consensus of the relative position parameters of each modality on each node, determine the trust weight of each modality on each node, and correct the internal consistency index and the transition consistency index according to the trust weight to construct the path trust of the traversal path between nodes. The walk-and-trace module is used to perform random walks from the terminal node to the upstream of the supply chain. It determines the transfer probability based on the path trust accumulated on the traceability path from the starting point to the upstream node, and determines the source node by combining the frequency of the upstream node being visited.

Images