A multi-modal material fusion prediction method

Abstract

This invention discloses a multimodal material fusion prediction method, comprising the following steps: constructing a multi-source heterogeneous material dataset; performing preprocessing to generate a standardized multimodal material feature matrix; performing event-driven decomposition on each modal feature and constructing a perturbation event sequence; performing correlation analysis to generate an initial perturbation causal relationship graph; performing causal structure learning and path filtering to construct a multimodal perturbation causal propagation network; calculating the impact attenuation and superposition effect of each perturbation during propagation to generate a perturbation propagation feature set; constructing a causal evidence fusion feature representation and generating material prediction results; performing perturbation inversion on the multimodal perturbation causal propagation network to generate perturbation inversion results; and combining the material prediction results and perturbation inversion results to generate multidimensional interpretable prediction results. This invention employs a multimodal fusion and perturbation causal propagation modeling method to achieve material prediction and inversion analysis, possessing the advantages of high prediction accuracy and strong interpretability.

Description

Technical Field

[0001] This invention relates to the field of material forecasting, and more particularly to a multimodal material fusion forecasting method. Background Technology

[0002] In existing technologies, material forecasting methods are mostly based on historical orders, inventory records, or single business data to build statistical models or time-series forecasting models to estimate material demand, inventory levels, and losses. As production, logistics, and the external environment increasingly influence material conditions, some solutions are beginning to incorporate production process data, transportation data, and external event data to improve forecast accuracy, and are attempting to combine machine learning or deep learning methods to achieve multi-source data fusion analysis.

[0003] However, most existing technologies focus on simple data splicing or correlation modeling from multiple sources, making it difficult to characterize the triggering relationships, propagation paths, and cumulative effects of different disturbance factors on changes in material demand, inventory, and losses. This results in insufficient dynamic response capabilities for prediction results in complex scenarios. Furthermore, existing methods generally lack the ability to trace prediction biases back to the source and explain key disturbances, making it difficult to simultaneously meet the application requirements for material forecast accuracy, causal analysis capabilities, and interpretability. Summary of the Invention

[0004] One objective of this invention is to propose a multimodal material fusion prediction method. This invention employs multimodal fusion and perturbation causal propagation modeling methods to achieve material prediction and inversion analysis, and has the advantages of high prediction accuracy and strong interpretability.

[0005] A multimodal material fusion prediction method according to an embodiment of the present invention includes the following steps:

[0006] Collect multimodal data related to the target material and construct a multi-source heterogeneous material data set;

[0007] The dataset of multi-source heterogeneous materials is preprocessed and then fused and stitched together according to a unified time index to generate a standardized multimodal material feature matrix.

[0008] Based on the standardized multimodal material feature matrix, event-driven decomposition is performed on each modal feature to extract the corresponding disturbance trigger signal and construct a disturbance event sequence;

[0009] A correlation analysis is performed on the sequence of disturbance events and the sequence of material state changes to construct a set of candidate causal relationships between disturbance events and changes in material demand, inventory and loss, and an initial disturbance causal relationship diagram is generated.

[0010] The initial perturbation causal relationship graph is subjected to causal structure learning and path filtering to identify perturbation source nodes, propagation paths and terminal nodes, and a multimodal perturbation causal propagation network is constructed.

[0011] Based on a multimodal perturbation causal propagation network, perturbation intensity quantification is performed on each perturbation source node, and the impact attenuation and superposition effect of each perturbation during the propagation process are calculated to generate a perturbation propagation feature set.

[0012] The perturbation propagation feature set is fused with the standardized multimodal material feature matrix to construct a causal evidence fusion feature representation and generate material prediction results;

[0013] Based on the material prediction results and the actual material state, perturbation inversion is performed on the multimodal perturbation causal propagation network to generate perturbation inversion results;

[0014] By combining material prediction results with perturbation inversion results, multidimensional interpretive prediction results are generated.

[0015] Optionally, the multimodal data includes business transaction data, inventory status data, production process data, logistics and transportation data, and external disturbance event data.

[0016] Optionally, the preprocessing includes missing value imputation, outlier removal, time alignment, uniform encoding, and numerical normalization.

[0017] Optionally, the construction of the perturbation event sequence specifically includes:

[0018] Based on the standardized multimodal material feature matrix, business transaction features, inventory status features, production process features, logistics and transportation features, and external disturbance event features are grouped and extracted according to feature modes to form time series feature subsets of the corresponding modes;

[0019] For each modal time series feature subset, sliding segmentation is performed, and the change amplitude, change rate and fluctuation intensity of feature values ​​within each time window are calculated to generate a feature change sequence;

[0020] A mutation detection is performed on the feature change sequence to identify the time point that meets the disturbance triggering condition, and the multi-dimensional vector composed of the corresponding feature change amplitude, change rate and fluctuation intensity is extracted as the disturbance triggering signal;

[0021] The disturbance trigger signals identified in each mode are aggregated in chronological order, and the comprehensive change intensity value and feature type value of the disturbance event segment are determined.

[0022] Based on the start time, end time, duration, comprehensive change intensity, and characteristic type of each disturbance event segment, disturbance event candidates are generated, and the disturbance event candidates are arranged in chronological order to form a disturbance event sequence.

[0023] Optionally, the generation of the initial perturbation causal graph includes:

[0024] Based on the disturbance event sequence, the start time value, end time value, duration value, comprehensive change intensity value and feature type value corresponding to each disturbance event candidate are extracted, and sorted according to a unified time index to form an ordered disturbance event sequence;

[0025] Obtain the material state change sequence within the time range corresponding to the disturbance event sequence, and perform time alignment on the material state change sequence. The material state change sequence includes the material demand change sequence, inventory change sequence, and loss change sequence.

[0026] For each disturbance event candidate, the corresponding candidate influence time window is determined according to the time interval relationship between its end time value and the material state change sequence. The material state change value within the candidate influence time window is extracted to generate event-state aligned data pairs.

[0027] For each event-state aligned data pair, perform change correlation analysis and calculate the correlation strength between the comprehensive change intensity value of the disturbance event and the changes in material demand, inventory, and loss.

[0028] The relationship between disturbance event candidates and material state changes is filtered based on the correlation strength value, and a set of causal candidate relationships between disturbance events and material state changes is constructed.

[0029] The candidate causal relationships are categorized and organized according to time order and feature type values, and initial perturbation causal relationship edges are constructed with perturbation event candidates as the starting point and material state changes as the ending point.

[0030] All initial disturbance causal relationship edges are summarized, and connections are established according to the chronological order of disturbance event candidates to form an initial disturbance causal relationship graph containing disturbance event nodes and material status nodes.

[0031] Optionally, the generation of the multimodal perturbation causal propagation network includes:

[0032] Read each initial disturbance causal relationship edge in the initial disturbance causal relationship graph, and extract the disturbance event candidate, material state change, event end time index value, feature type value and correlation strength value corresponding to each initial disturbance causal relationship edge to form an initial causal relationship edge set;

[0033] Based on the event end time index value corresponding to each initial perturbation causal relationship edge, the event is arranged in chronological order to determine the temporal relationship between each perturbation event candidate and to generate the temporal connection relationship between the perturbation event candidate;

[0034] Connect the perturbation event candidates with temporal connections to construct a set of candidate propagation paths;

[0035] For each candidate propagation path in the candidate propagation path set, the edge weights are accumulated according to the correlation strength values ​​of each initial perturbation causal relationship edge in the path to obtain the path strength value of the corresponding candidate propagation path, and the path length value is calculated according to the number of relation edges contained in the path.

[0036] Based on the path strength and path length values ​​of each candidate propagation path, a set of effective propagation paths is formed;

[0037] For each candidate propagation path in the effective propagation path set, the starting point is identified, the disturbance event candidates with zero incoming edges are extracted as disturbance source nodes, the material status nodes corresponding to the end of the path are extracted as action terminal nodes, and the disturbance event candidates in the middle of the path are extracted as propagation intermediate nodes.

[0038] By summarizing the source nodes of disturbances, intermediate nodes of propagation, terminal nodes of effects, and the set of effective propagation paths, a multimodal disturbance causal propagation network is constructed, which includes node connection relationships, edge weights, and path order.

[0039] Optionally, the generation of the disturbance propagation feature set specifically includes:

[0040] Read the source nodes, intermediate nodes, terminal nodes, and edge weights of each causal connection in the multimodal perturbation causal propagation network.

[0041] For each disturbance source node, the comprehensive change intensity value of the disturbance event candidates corresponding to the disturbance source node is used as the disturbance intensity value of the disturbance source node, and as the initial propagation intensity value of the first hop of the corresponding propagation path;

[0042] Along each propagation path in the multimodal perturbation causal propagation network, the edge weights and event pair propagation strength values ​​in each path are extracted and arranged in the order of the paths to form a sequence of path propagation parameters.

[0043] For each propagation path, the propagation strength values ​​of each edge weight and event pair in the path propagation parameter sequence are read sequentially to generate single-hop propagation strength values ​​and form a path propagation strength sequence.

[0044] The single-hop propagation intensity values ​​in the same propagation path are accumulated to obtain the cumulative path impact value of the corresponding propagation path, and the terminal single-hop propagation intensity value in the path propagation intensity sequence is extracted to obtain the terminal path impact value.

[0045] For multiple propagation paths pointing to the same terminal node, the cumulative impact value and the terminal impact value of each propagation path are read respectively, and the cumulative impact value and the terminal impact value of each path are added together to obtain the superimposed impact value of the terminal node.

[0046] The disturbance intensity value corresponding to each disturbance source node, the cumulative path impact value corresponding to each propagation path, and the superimposed impact value corresponding to each terminal node are summarized to generate a disturbance propagation feature set.

[0047] Optionally, the generation of the material prediction results specifically includes:

[0048] Read the disturbance propagation feature set, and align the disturbance intensity value, path cumulative impact value, and superimposed impact value with time according to a unified time index to obtain the disturbance propagation feature vector;

[0049] Multimodal material feature vectors corresponding to each time index are extracted from the standardized multimodal material feature matrix. The perturbation propagation feature vectors are time-aligned with the multimodal material feature vectors, and then normalized and spliced ​​to generate a fused feature vector, which serves as the fused feature representation of causal evidence.

[0050] The causal evidence fusion feature representations of multiple consecutive time steps are sequentially combined to construct the input feature sequence. The input feature sequence is then input into the time-series fusion Transformer prediction network to obtain the material demand forecast, inventory forecast, and loss forecast under the corresponding time index.

[0051] The material demand forecast, inventory forecast, and loss forecast values ​​under each time index are arranged and summarized in chronological order to generate material forecast results.

[0052] Optionally, the generation of the perturbation inversion result specifically includes:

[0053] Read the material demand forecast, inventory forecast, and loss forecast values ​​corresponding to each time index in the material forecast results, and obtain the actual material demand, inventory, and loss values ​​under the corresponding time index to obtain the material demand deviation, inventory deviation, and loss deviation values.

[0054] Based on a unified time index, material demand deviation, inventory deviation, and loss deviation are combined to construct a material status deviation vector and form a deviation-driven time index set.

[0055] For each time index position in the deviation-driven time index set, the corresponding terminal node is located in the multimodal disturbance causal propagation network, and the superimposed influence value is obtained. The material state deviation vector and the superimposed influence value are combined to obtain the terminal driving value.

[0056] Path activation is performed on each propagation path in the multimodal perturbation causal propagation network to obtain the path activation value corresponding to each propagation path, and the propagation paths are filtered to form a set of activated propagation paths.

[0057] For each propagation path in the set of active propagation paths, read the disturbance source node corresponding to the starting node of the path and its corresponding disturbance intensity value, and combine the disturbance intensity value with the path activation value of the corresponding propagation path to obtain the source driving value corresponding to each disturbance source node;

[0058] The source driving values ​​corresponding to each disturbance source node are sorted, and the source driving values ​​of multiple disturbance source nodes are compared under the same time index. The disturbance source node with the largest source driving value is selected as the dominant disturbance source node, and its corresponding propagation path is determined as the dominant propagation path.

[0059] For disturbance source nodes that are not selected as the dominant disturbance source node, secondary disturbance source nodes are obtained by filtering based on the difference between their source driving value and the source driving value of the dominant disturbance source node, and the corresponding propagation paths are extracted.

[0060] The dominant disturbance source node, secondary disturbance source nodes and their corresponding propagation paths are organized according to a unified time index and associated with the corresponding source driving values ​​to generate disturbance inversion results.

[0061] Optionally, the generation of the multidimensional interpretive prediction results specifically includes:

[0062] Read the material demand forecast, inventory forecast and loss forecast values ​​corresponding to each time index in the material forecast results, and read the dominant disturbance source node, secondary disturbance source node and their corresponding propagation path corresponding to each time index in the disturbance inversion results, and extract the source driving value corresponding to each disturbance source node.

[0063] For each time index position in the deviation-driven time index set, the source driving value and corresponding propagation path of the corresponding dominant disturbance source node are extracted, and the source driving value is associated with the material demand forecast, inventory forecast and loss forecast corresponding to the time index position to obtain the dominant explanatory value.

[0064] For each time index position in the deviation-driven time index set, the source driving value and corresponding propagation path corresponding to each secondary disturbance source node are read, and the source driving value corresponding to each secondary disturbance source node is associated with the material demand forecast value, inventory forecast value and loss forecast value corresponding to the time index position to obtain the secondary explanation value.

[0065] The propagation paths corresponding to the dominant and secondary disturbance source nodes are associated and labeled, and the path activation values ​​corresponding to each propagation path are extracted to generate path interpretation results.

[0066] For each time index position in the deviation-driven time index set, the corresponding material demand forecast, inventory forecast, and loss forecast are combined with the corresponding dominant explanatory value, each secondary explanatory value, and path explanation result to obtain the explanatory forecast result.

[0067] For time index positions that do not belong to the deviation-driven time index set, the corresponding material demand forecast, inventory forecast, and loss forecast are read and associated with the corresponding time index to obtain the forecast results. The explanatory forecast results are then summarized with the forecast results to generate multidimensional explanatory forecast results.

[0068] The beneficial effects of this invention are:

[0069] The multimodal material fusion prediction method provided by this invention addresses the technical problems of "dispersed data sources, complex disturbance effects, and lack of interpretation of prediction results" in the material prediction process. It constructs a complete technical chain consisting of multimodal data acquisition and preprocessing, disturbance event extraction, causal relationship modeling, propagation path analysis, fusion prediction, disturbance inversion, and interpretive output. Compared to existing technologies that only utilize historical business data or employ simple correlation modeling, this invention can align and structure multi-source heterogeneous information such as business transactions, inventory status, production processes, logistics and transportation, and external disturbance events under a unified time index, thereby improving the completeness and temporal consistency of material status representation.

[0070] Furthermore, this invention utilizes an event-driven decomposition mechanism to identify disturbance trigger signals from multimodal features and construct a disturbance event sequence. It then combines changes in material demand, inventory, and losses to establish a set of causal candidate relationships between disturbance events and material state changes, thereby generating an initial disturbance causal relationship graph and a multimodal disturbance causal propagation network. This approach effectively reveals the triggering relationships, transmission order, action paths, and impact endpoints of different disturbance factors on material state changes. It overcomes the shortcomings of existing technologies that can only reflect surface correlations and are unable to describe complex disturbance propagation mechanisms. This elevates material forecasting from static fitting to causal forecasting oriented towards dynamic disturbance propagation processes, which is beneficial for improving the prediction accuracy and robustness in complex business scenarios.

[0071] Meanwhile, this invention quantifies the disturbance intensity of the disturbance source node, the cumulative impact value of the propagation path, and the superimposed impact value of the terminal node, forming a disturbance propagation feature set. This set is then fused with a standardized multimodal material feature matrix to construct a causal evidence fusion feature representation, which is then input into a time-series fusion Transformer prediction network to generate material prediction results. This not only preserves the supporting role of the original multimodal features in the prediction results but also introduces causal evidence information during the disturbance propagation process. This allows the prediction model to simultaneously focus on the original business state and the disturbance evolution trajectory, thereby improving its responsiveness to fluctuations in material demand, inventory changes, and loss changes, and enhancing its adaptability to complex situations such as emergencies, logistics anomalies, and production line fluctuations.

[0072] Furthermore, after obtaining the prediction results, this invention can perform disturbance inversion on the multimodal disturbance causal propagation network in conjunction with the actual material state. By locating the terminal nodes through the deviation-driven time index set, activating the propagation path, and deducing the dominant and secondary disturbance source nodes, it achieves reverse tracing of the source of prediction deviation. Based on this, the predicted values, dominant explanatory values, secondary explanatory values, and path explanation results are further correlated and integrated to generate multidimensional explanatory prediction results. This invention can not only output prediction results for material demand, inventory, and losses, but also explain which disturbance sources, along which propagation paths, and with what intensity of influence the prediction results are generated, thus improving the interpretability, traceability, and decision-making support value of the prediction results, facilitating managers to conduct early warning analysis, cause identification, and scheduling optimization. Attached Figure Description

[0073] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0074] Figure 1 This is a flowchart of a multimodal material fusion prediction method proposed in this invention;

[0075] Figure 2 This is a schematic diagram illustrating the process of generating multidimensional interpretive prediction results for a multimodal material fusion prediction method proposed in this invention. Detailed Implementation

[0076] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0077] refer to Figures 1-2 A multimodal material fusion prediction method includes the following steps:

[0078] Collect multimodal data related to the target material and construct a multi-source heterogeneous material data set;

[0079] The dataset of multi-source heterogeneous materials is preprocessed and then fused and stitched together according to a unified time index to generate a standardized multimodal material feature matrix.

[0080] Based on the standardized multimodal material feature matrix, event-driven decomposition is performed on each modal feature to extract the corresponding disturbance trigger signal and construct a disturbance event sequence;

[0081] A correlation analysis is performed on the sequence of disturbance events and the sequence of material state changes to construct a set of candidate causal relationships between disturbance events and changes in material demand, inventory and loss, and an initial disturbance causal relationship diagram is generated.

[0082] The initial perturbation causal relationship graph is subjected to causal structure learning and path filtering to identify perturbation source nodes, propagation paths and terminal nodes, and a multimodal perturbation causal propagation network is constructed.

[0083] Based on a multimodal perturbation causal propagation network, perturbation intensity quantification is performed on each perturbation source node, and the impact attenuation and superposition effect of each perturbation during the propagation process are calculated to generate a perturbation propagation feature set.

[0084] The perturbation propagation feature set is fused with the standardized multimodal material feature matrix to construct a causal evidence fusion feature representation and generate material prediction results;

[0085] Based on the material prediction results and the actual material state, perturbation inversion is performed on the multimodal perturbation causal propagation network to generate perturbation inversion results;

[0086] By combining material prediction results with perturbation inversion results, multidimensional interpretive prediction results are generated.

[0087] In this embodiment, multimodal data includes business transaction data, inventory status data, production process data, logistics and transportation data, and external disturbance event data. Business transaction data includes order generation time, order quantity, order type, and transaction frequency information. Inventory status data includes inventory quantity, inventory change records, inbound time, outbound time, and inventory turnover information. Production process data includes production cycle time, equipment operating status, and production line load information. Logistics and transportation data includes transportation routes, transportation duration, transportation batches, and transportation delay records. External disturbance event data includes climate and environmental change information and records of emergencies. In business transaction data, order generation time is mapped to a time index sequence, order quantity is converted into order size characteristics, and order type is represented by discrete numerical values. Transaction frequency is calculated based on the number of times orders occur within a unit time window. In inventory status data, inventory quantity is constructed as an inventory level time series, and inventory change records are converted into phase... The inventory difference sequence at adjacent time points maps the inbound and outbound times to inventory inflow and outflow timestamps, and calculates inventory turnover rate characteristics based on inventory consumption and average inventory. Production cycle time in production process data is transformed into unit output time characteristics, equipment operating status is discretized into running, shutdown, and abnormal states, and production line load is normalized by the ratio of actual output to rated capacity. Transportation routes in logistics and transportation data are transformed into path vectors composed of node sequences, transportation duration is constructed as transportation time characteristics, transportation batches are encoded as batch identifier sequences, and transportation delay records are transformed into delay durations. Climate and environmental change information in external disturbance event data is transformed into temperature deviation and precipitation intensity, and sudden event records are represented as event vectors containing event occurrence time, event duration, and event impact intensity. All of the above quantitative features are aligned according to a unified time index and constructed into a structured numerical representation of multimodal material features.

[0088] In this embodiment, preprocessing includes missing value imputation, outlier removal, time alignment, uniform encoding, and numerical normalization.

[0089] In this embodiment, the construction of the perturbation event sequence specifically includes:

[0090] Based on the standardized multimodal material feature matrix, business transaction features, inventory status features, production process features, logistics and transportation features, and external disturbance event features are grouped and extracted according to feature modes to form time series feature subsets of the corresponding modes;

[0091] For each modal time series feature subset, sliding segmentation is performed, and the change amplitude, change rate and fluctuation intensity of feature values ​​within each time window are calculated to generate a feature change sequence. The fluctuation intensity is the degree of discrete change of the same feature value relative to its mean value within a preset time window.

[0092] A mutation detection is performed on the feature change sequence to identify the time point that meets the disturbance triggering condition, and the multi-dimensional vector composed of the corresponding feature change amplitude, change rate and fluctuation intensity is extracted as the disturbance triggering signal;

[0093] The disturbance triggering condition is a comprehensive change index obtained by weighting and fusing the change amplitude, change rate and fluctuation intensity of the same feature between adjacent time windows, and the judgment criteria that are met when the comprehensive change index exceeds a preset threshold.

[0094] The disturbance trigger signals identified in each mode are aggregated in chronological order, and the comprehensive change intensity value and feature type value of the disturbance event segment are determined.

[0095] The disturbance trigger signals identified in each mode are sorted according to a unified time index to obtain the trigger time value corresponding to each disturbance trigger signal. All disturbance trigger signals are arranged in chronological order based on their trigger time values ​​to generate an ordered disturbance trigger signal sequence. For each disturbance trigger signal in the ordered sequence, its corresponding characteristic change amplitude, change rate, and fluctuation intensity are extracted, and these are weighted and summed to obtain a change intensity value. The characteristic mode corresponding to each disturbance trigger signal is identified and encoded to obtain a feature type value. A disturbance trigger signal triplet is constructed using the trigger time value, feature type value, and change intensity value. The time interval between adjacent disturbance trigger signal triplets is calculated sequentially according to the trigger time order. When the time interval is less than a preset time window, the corresponding disturbance trigger signal triplets are merged into the same disturbance event segment. When the time interval is greater than or equal to the preset time window, a new disturbance event segment is generated. The change intensity values ​​corresponding to each disturbance trigger signal triplet contained in each disturbance event segment are extracted and summed in chronological order to obtain the comprehensive change intensity value of the disturbance event segment. The frequency of the feature type values ​​corresponding to each disturbance trigger signal triplet in each disturbance event segment is statistically analyzed, and the feature type value with the most occurrences is selected as the feature type value of the disturbance event segment. When there are multiple feature type values ​​with the same occurrence frequency, the feature type value with the largest average change intensity value is selected as the feature type value of the disturbance event segment.

[0096] Based on the start time, end time, duration, comprehensive change intensity, and characteristic type of each disturbance event segment, disturbance event candidates are generated, and the disturbance event candidates are arranged in chronological order to form a disturbance event sequence.

[0097] The start time value of a disturbance event segment is the trigger time value corresponding to the first disturbance trigger signal triplet arranged in chronological order within the disturbance event segment, and the end time value of a disturbance event segment is the trigger time value corresponding to the last disturbance trigger signal triplet arranged in chronological order within the disturbance event segment.

[0098] In this embodiment, the generation of the initial perturbation causal relationship graph includes:

[0099] Based on the disturbance event sequence, the start time value, end time value, duration value, comprehensive change intensity value and feature type value corresponding to each disturbance event candidate are extracted, and sorted according to a unified time index to form an ordered disturbance event sequence;

[0100] Obtain the material state change sequence within the time range corresponding to the disturbance event sequence, and perform time alignment on the material state change sequence. The material state change sequence includes the material demand change sequence, inventory change sequence, and loss change sequence.

[0101] Material demand change series, inventory change series, and loss change series are time series data formed by calculating the changes in material demand, inventory quantity, and loss between adjacent time points.

[0102] For each disturbance event candidate, the corresponding candidate influence time window is determined according to the time interval relationship between its end time value and the material state change sequence. The material state change value within the candidate influence time window is extracted to generate event-state aligned data pairs.

[0103] The generation of event-state aligned data pairs specifically includes: for each disturbance event candidate, reading its end time value and locating the time position corresponding to the end time value on a unified time index to obtain the event end time index value; expanding backward from the event end time index value to determine the start and end time values ​​of the corresponding candidate influence time window, generating the candidate influence time window range value; extracting the material demand change value, inventory change value, and loss change value within the corresponding time interval from the material demand change sequence, inventory change sequence, and loss change sequence based on the candidate influence time window range value, forming a material state change subsequence; aggregating the material state change subsequences in chronological order, calculating the material demand change value, inventory change value, and loss change value respectively; combining the comprehensive change intensity value of the disturbance event candidate with the material demand change value, inventory change value, and loss change value to construct an event-state feature vector; and associating the event-state feature vector with the corresponding feature type value and the event end time index value to generate event-state aligned data pairs.

[0104] For each event-state aligned data pair, perform change correlation analysis and calculate the correlation strength between the comprehensive change intensity value of the disturbance event and the changes in material demand, inventory, and loss.

[0105] The generation of correlation strength values ​​specifically includes: reading the comprehensive change strength value from each event-state aligned data pair and constructing an event strength sequence value according to a unified time index; reading the material demand change value, inventory change value, and loss change value respectively, and constructing material demand change sequence values, inventory change sequence values, and loss change sequence values ​​according to a unified time index; calculating the cosine similarity between the event strength sequence value and the material demand change sequence value to obtain the correlation strength value between the comprehensive change strength value of the disturbance event and the material demand change value; using the same method, calculating the cosine similarity between the event strength sequence value and the inventory change sequence value, and the cosine similarity between the event strength sequence value and the loss change sequence value, respectively, to obtain the correlation strength value between the comprehensive change strength value of the disturbance event and the inventory change value, and the correlation strength value between the comprehensive change strength value of the disturbance event and the loss change value, respectively.

[0106] The relationship between disturbance event candidates and material state changes is filtered based on the correlation strength value, and a set of causal candidate relationships between disturbance events and material state changes is constructed.

[0107] The specific steps for constructing a set of causal candidate relationships between disturbance events and material state changes include: reading the material demand correlation strength value, inventory correlation strength value, and loss correlation strength value from each event-state aligned data pair, and comparing them with the corresponding preset correlation thresholds. When the corresponding correlation strength value is greater than the corresponding preset correlation threshold, the corresponding correlation judgment value is considered to meet the condition; when the corresponding correlation strength value is less than or equal to the corresponding preset correlation threshold, the corresponding correlation judgment value is considered to not meet the condition. For each event-state aligned data pair, when the material demand correlation judgment value meets the condition, a demand causal candidate relationship pair between the disturbance event candidate and the material demand change is generated; when the inventory correlation judgment value meets the condition, an inventory causal candidate relationship pair between the disturbance event candidate and the inventory change is generated; when the loss correlation judgment value meets the condition, a loss causal candidate relationship pair between the disturbance event candidate and the loss change is generated. The demand causal candidate relationship pairs, inventory causal candidate relationship pairs, and loss causal candidate relationship pairs are then summarized to form a set of causal candidate relationships between disturbance events and material state changes.

[0108] The candidate causal relationships are categorized and organized according to time order and feature type values, and initial perturbation causal relationship edges are constructed with perturbation event candidates as the starting point and material state changes as the ending point.

[0109] Starting with the candidate disturbance event and ending with the change in material state, the candidate disturbance event is taken as the source node of the causal relationship, and the changes in material demand, inventory, or loss affected by it are taken as the target node of the causal relationship.

[0110] All initial disturbance causal relationship edges are summarized, and connections are established according to the chronological order of disturbance event candidates to form an initial disturbance causal relationship graph containing disturbance event nodes and material status nodes;

[0111] The edge weights of the initial disturbance causal relationship graph are set to the correlation strength values ​​of the corresponding causal candidate relationship pairs, which are used to characterize the influence of disturbance event candidates on material state changes. Disturbance event nodes and material state nodes are graph structure nodes in the initial disturbance causal relationship graph used to characterize disturbance event candidates and changes in material demand, inventory, or loss, respectively.

[0112] In this embodiment, the generation of the multimodal perturbation causal propagation network includes:

[0113] Read each initial disturbance causal relationship edge in the initial disturbance causal relationship graph, and extract the disturbance event candidate, material state change, event end time index value, feature type value and correlation strength value corresponding to each initial disturbance causal relationship edge to form an initial causal relationship edge set;

[0114] Based on the event end time index value corresponding to each initial perturbation causal relationship edge, the event is arranged in chronological order to determine the temporal relationship between each perturbation event candidate and to generate the temporal connection relationship between the perturbation event candidate;

[0115] The specific steps for generating temporal connection relationships between disturbance event candidates include: arranging each disturbance event candidate sequentially according to its event end time index value to obtain an ordered disturbance event sequence; for any two adjacent disturbance event candidates in the ordered disturbance event sequence, extracting the end time value of the preceding disturbance event candidate and the start time value of the following disturbance event candidate, and calculating the time interval between them; comparing the time interval value with a preset connection time threshold, generating a connection judgment value of 1 when the time interval value is less than or equal to the preset connection time threshold, and generating a connection judgment value of 0 when the time interval value is greater than the preset connection time threshold; for two adjacent disturbance event candidates with a connection judgment value of 1, extracting the feature type values ​​of the preceding and following disturbance event candidates, and generating corresponding type combination values; and combining the preceding disturbance event candidates, the following disturbance event candidates, the time interval value, and the type combination value to generate temporal connection relationships between the disturbance event candidates.

[0116] Connect the perturbation event candidates with temporal connections to construct a set of candidate propagation paths;

[0117] The path connection of disturbance event candidates with temporal connectivity includes: reading the temporal connectivity between each disturbance event candidate; extracting the preceding and following disturbance event candidates for each temporal connectivity to generate candidate connection event pairs; for each candidate connection event pair, reading the comprehensive change intensity value corresponding to the preceding disturbance event candidate, the comprehensive change intensity value corresponding to the following disturbance event candidate, and the time interval between them; and averaging the comprehensive change intensity values ​​corresponding to the preceding and following disturbance event candidates and dividing by the sum of the time interval and 1 to obtain the propagation intensity value of the event pair; and extracting the feature type value corresponding to the preceding disturbance event candidate and the following disturbance event candidate... The feature type values ​​corresponding to the options are combined in sequence to obtain the event pair type combination value; the preceding perturbation event candidates, the following perturbation event candidates, the event pair propagation strength value, the event pair type combination value, and the time interval value are combined to generate a single-hop propagation path item; the preceding perturbation event candidates of the subsequent single-hop propagation path item are sequentially concatenated with the following perturbation event candidates of the previous single-hop propagation path item to determine the starting perturbation event candidates and the ending perturbation event candidates in the concatenated path, and the propagation strength value, the event pair type combination value, and the time interval value of each hop event pair are summarized to generate a multi-hop propagation path item; the single-hop propagation path items and the multi-hop propagation path items are summarized to form a candidate propagation path set;

[0118] For each candidate propagation path in the candidate propagation path set, the edge weights are accumulated according to the correlation strength values ​​of each initial perturbation causal relationship edge in the path to obtain the path strength value of the corresponding candidate propagation path, and the path length value is calculated according to the number of relation edges contained in the path.

[0119] Based on the path strength value and path length value of each candidate propagation path, the path is filtered, and the candidate propagation paths with a path strength value greater than a preset path threshold and a path length value less than a preset length threshold are retained to form an effective propagation path set;

[0120] For each candidate propagation path in the effective propagation path set, the starting point is identified, the disturbance event candidates with zero incoming edges are extracted as disturbance source nodes, the material status nodes corresponding to the end of the path are extracted as action terminal nodes, and the disturbance event candidates in the middle of the path are extracted as propagation intermediate nodes.

[0121] By summarizing the set of disturbance source nodes, intermediate propagation nodes, terminal nodes, and effective propagation paths, a multimodal disturbance causal propagation network is constructed, which includes node connection relationships, edge weights, and path order. The multimodal disturbance causal propagation network is composed of different disturbance event candidates, material state changes, and their causal connections, and is used to characterize the network structure of the propagation path and propagation intensity of disturbances between nodes.

[0122] In this embodiment, the generation of the disturbance propagation feature set specifically includes:

[0123] Read the source nodes, intermediate nodes, terminal nodes, and edge weights of each causal connection in the multimodal perturbation causal propagation network.

[0124] For each disturbance source node, the comprehensive change intensity value of the disturbance event candidates corresponding to the disturbance source node is used as the disturbance intensity value of the disturbance source node, and as the initial propagation intensity value of the first hop of the corresponding propagation path;

[0125] Along each propagation path in the multimodal perturbation causal propagation network, the edge weights and event pair propagation strength values ​​in each path are extracted and arranged in the order of the paths to form a sequence of path propagation parameters.

[0126] For each propagation path, the propagation strength values ​​of each edge weight and event pair in the path propagation parameter sequence are read sequentially to generate single-hop propagation strength values ​​and form a path propagation strength sequence.

[0127] The generation of the path propagation strength sequence specifically includes: sequentially reading the propagation strength values ​​of each edge weight and event pair in the path propagation parameter sequence; when the current hop is the first hop in the propagation path, reading the initial propagation strength value corresponding to the first hop and using the initial propagation strength value as the input propagation strength value of the current hop; when the current hop is the second hop and subsequent hops in the propagation path, reading the single-hop propagation strength value corresponding to the previous hop and using the single-hop propagation strength value corresponding to the previous hop as the input propagation strength value of the current hop; multiplying the input propagation strength value of the current hop by the corresponding edge weight to obtain the edge weight effect value, and then multiplying the edge weight effect value by the corresponding event pair propagation strength value to obtain the single-hop propagation strength value of the current hop; sequentially calculating the single-hop propagation strength values ​​corresponding to each hop in the propagation path and arranging them according to the path order to form the path propagation strength sequence;

[0128] The single-hop propagation intensity values ​​in the same propagation path are accumulated to obtain the cumulative path impact value of the corresponding propagation path, and the terminal single-hop propagation intensity value in the path propagation intensity sequence is extracted to obtain the terminal path impact value.

[0129] For multiple propagation paths pointing to the same terminal node, the cumulative impact value and the terminal impact value of each propagation path are read respectively, and the cumulative impact value and the terminal impact value of each path are added together to obtain the superimposed impact value of the terminal node.

[0130] The disturbance intensity value corresponding to each disturbance source node, the cumulative path impact value corresponding to each propagation path, and the superimposed impact value corresponding to each terminal node are summarized to generate a disturbance propagation feature set.

[0131] In this embodiment, the generation of material prediction results specifically includes:

[0132] Read the disturbance propagation feature set, and align the disturbance intensity value, path cumulative impact value, and superimposed impact value with time according to a unified time index to obtain the disturbance propagation feature vector;

[0133] Read the disturbance intensity value and the start time value of the corresponding disturbance event candidate for each disturbance source node in the disturbance propagation feature set, and determine the time index position corresponding to each start time value according to the unified time index. Assign the corresponding disturbance intensity value to the corresponding time index position, and assign a value of zero to the unassigned time index position to obtain the disturbance intensity value sequence. Read the path cumulative impact value and the corresponding propagation time interval for each propagation path, and allocate the path cumulative impact value of each propagation path to each time index position corresponding to its propagation time interval one by one, and assign the path cumulative impact value to the corresponding time index position. The impact value is calculated by summing multiple cumulative impact values ​​at the same time index position, and assigning a value of zero to the unassigned time index position to obtain a sequence of cumulative impact values ​​for the path. The superimposed impact value and its corresponding time index value for each affected terminal node are read, and the superimposed impact value is assigned to the corresponding time index position, while a value of zero is assigned to the unassigned time index position to obtain a sequence of superimposed impact values. At each time index position, the corresponding disturbance intensity value, cumulative impact value for the path, and superimposed impact value are extracted sequentially in a fixed order and combined to obtain a disturbance propagation feature vector corresponding to each time index.

[0134] Multimodal material feature vectors corresponding to each time index are extracted from the standardized multimodal material feature matrix. The perturbation propagation feature vectors are time-aligned with the multimodal material feature vectors, and then normalized and spliced ​​to generate a fused feature vector, which serves as the fused feature representation of causal evidence.

[0135] The causal evidence fusion feature representations of multiple consecutive time steps are sequentially combined to construct the input feature sequence. The input feature sequence is then input into the time-series fusion Transformer prediction network to obtain the material demand forecast, inventory forecast, and loss forecast under the corresponding time index.

[0136] The material demand forecast, inventory forecast, and loss forecast values ​​under each time index are arranged and summarized in chronological order to generate material forecast results.

[0137] In this embodiment, the generation of the perturbation inversion result specifically includes:

[0138] Read the material demand forecast, inventory forecast, and loss forecast values ​​corresponding to each time index in the material forecast results, and obtain the actual material demand, inventory, and loss values ​​under the corresponding time index to obtain the material demand deviation, inventory deviation, and loss deviation values.

[0139] Based on a unified time index, material demand deviation, inventory deviation, and loss deviation are combined to construct a material status deviation vector and form a deviation-driven time index set.

[0140] According to a unified time index, the material demand deviation value, inventory deviation value, and loss deviation value are combined to obtain a material status deviation vector corresponding to each time index. The absolute value of the material status deviation vector at each time index position is calculated, and the absolute values ​​of the material demand deviation value, inventory deviation value, and loss deviation value are accumulated to obtain the deviation intensity value corresponding to each time index position. The deviation intensity values ​​corresponding to each time index position are sorted according to their numerical values ​​to obtain a deviation intensity sorting sequence. From the deviation intensity sorting sequence, the time index positions with deviation intensity values ​​greater than a preset deviation threshold are selected, and the time index positions are summarized to obtain a deviation-driven time index set.

[0141] For each time index position in the deviation-driven time index set, the corresponding terminal node is located in the multimodal disturbance causal propagation network, and the superimposed influence value is obtained. The material state deviation vector and the superimposed influence value are combined to obtain the terminal driving value.

[0142] When generating the terminal driving value, each time index value in the deviation driving time index set is read one by one, and the material state node corresponding to the time index value is retrieved in the multimodal disturbance causal propagation network. From the material state nodes, the corresponding material demand change nodes, inventory change nodes, and loss change nodes are extracted respectively, and these nodes are identified as the terminal nodes corresponding to the time index, resulting in a set of terminal nodes. The superimposed influence value corresponding to each terminal node in the set of terminal nodes is read, and multiple superimposed influence values ​​at the same time index position are accumulated to obtain the terminal superimposed influence value corresponding to that time index position. The material state deviation vector corresponding to that time index position is read, and the material demand deviation value, inventory deviation value, and loss deviation value are extracted respectively. After absolute value processing of the material demand deviation value, inventory deviation value, and loss deviation value, they are accumulated to obtain the deviation intensity value corresponding to that time index position. The deviation intensity value is multiplied by the terminal superimposed influence value to obtain the terminal driving value corresponding to that time index position.

[0143] Path activation is performed on each propagation path in the multimodal perturbation causal propagation network to obtain the path activation value corresponding to each propagation path, and the propagation paths are filtered to form a set of activated propagation paths.

[0144] When generating the set of active propagation paths, all propagation paths in the multimodal perturbation causal propagation network are read, and the corresponding cumulative impact value and the propagation intensity value of each event pair in the propagation path are extracted for each propagation path. The path propagation intensity sequence of the propagation path is formed according to the path order. For each propagation path, the terminal driving value at the corresponding time index position is read, and the terminal driving value is used as the initial input value of the propagation path to obtain the initial driving value of the path. Along the path propagation intensity sequence of the propagation path, the propagation intensity value of each event pair is read in sequence, and the current path input value is multiplied by the corresponding event pair propagation intensity value to obtain the current path propagation value. The current path propagation value is then updated to the path input value of the next hop, until the hop-by-hop calculation of the propagation intensity value of all event pairs in the propagation path is completed to obtain the path propagation result value corresponding to the propagation path. The path propagation result value is multiplied by the cumulative impact value of the path corresponding to the propagation path to obtain the path activation value corresponding to the propagation path. The path activation values ​​corresponding to each propagation path are compared with a preset activation threshold. Propagation paths with path activation values ​​greater than the preset activation threshold are retained, and the retained propagation paths are summarized to obtain the set of active propagation paths.

[0145] For each propagation path in the set of active propagation paths, read the disturbance source node corresponding to the starting node of the path and its corresponding disturbance intensity value, and combine the disturbance intensity value with the path activation value of the corresponding propagation path to obtain the source driving value corresponding to each disturbance source node;

[0146] When generating the set of active propagation paths, each propagation path in the set is read one by one, and the starting node of each propagation path is extracted to obtain the path starting node corresponding to each propagation path; the node type of each path starting node is identified, and the corresponding disturbance source node is extracted to obtain the disturbance source node identifier value corresponding to each propagation path; based on the disturbance source node identifier value, the corresponding disturbance intensity value is read from the disturbance propagation feature set to obtain the source node disturbance intensity value corresponding to each propagation path; the path activation value corresponding to each propagation path is read, and the source node disturbance intensity value corresponding to each propagation path is multiplied by the corresponding path activation value to obtain the path source driving value corresponding to each propagation path; the path source driving values ​​corresponding to each propagation path with the same starting node are accumulated to obtain the cumulative source driving value corresponding to each disturbance source node; the identifier value of each disturbance source node is associated with the corresponding cumulative source driving value to obtain the source driving value corresponding to each disturbance source node.

[0147] The source driving values ​​corresponding to each disturbance source node are sorted, and the source driving values ​​of multiple disturbance source nodes are compared under the same time index. The disturbance source node with the largest source driving value is selected as the dominant disturbance source node, and its corresponding propagation path is determined as the dominant propagation path.

[0148] For disturbance source nodes that are not selected as the dominant disturbance source node, secondary disturbance source nodes are obtained by filtering based on the difference between their source driving value and the source driving value of the dominant disturbance source node, and the corresponding propagation paths are extracted.

[0149] Read the source driving value corresponding to each disturbance source node, sort them according to the size of the source driving value, and determine the disturbance source node with the largest source driving value as the dominant disturbance source node, thus obtaining the dominant source driving value; remove the dominant disturbance source node from the disturbance source nodes to obtain a set of candidate disturbance source nodes that were not selected as the dominant disturbance source node; read the source driving value corresponding to each candidate disturbance source node in the candidate disturbance source node set one by one, and calculate the difference between the source driving value corresponding to each candidate disturbance source node and the dominant source driving value to obtain the source driving difference corresponding to each candidate disturbance source node; then, sort the candidate disturbances... The source driver difference corresponding to the source node is compared with a preset difference threshold. Candidate disturbance source nodes whose source driver difference is less than the preset difference threshold are retained to obtain the secondary disturbance source node identification value. Based on the identification value of each secondary disturbance source node, the propagation path starting from the corresponding secondary disturbance source node is retrieved in the active propagation path set, and the retrieved propagation path is extracted to obtain the secondary propagation path set corresponding to each secondary disturbance source node. The identification value of each secondary disturbance source node is associated with the corresponding secondary propagation path set to obtain the secondary disturbance source node and its corresponding propagation path.

[0150] The dominant disturbance source node, secondary disturbance source nodes and their corresponding propagation paths are organized according to a unified time index and associated with the corresponding source driving values ​​to generate disturbance inversion results.

[0151] In this embodiment, the generation of multidimensional interpretive prediction results specifically includes:

[0152] Read the material demand forecast, inventory forecast and loss forecast values ​​corresponding to each time index in the material forecast results, and read the dominant disturbance source node, secondary disturbance source node and their corresponding propagation path corresponding to each time index in the disturbance inversion results, and extract the source driving value corresponding to each disturbance source node.

[0153] For each time index position in the deviation-driven time index set, the source driving value and corresponding propagation path of the corresponding dominant disturbance source node are extracted, and the source driving value is associated with the material demand forecast, inventory forecast and loss forecast corresponding to the time index position to obtain the dominant explanatory value.

[0154] When generating the dominant explanatory value, each time index value in the deviation-driven time index set is read one by one, and the dominant disturbance source node corresponding to the time index value is extracted from the disturbance inversion result. At the same time, the source driving value and the corresponding propagation path corresponding to the dominant disturbance source node are read to obtain the dominant source driving value and dominant propagation path corresponding to the time index position. In the material forecasting result, the material demand forecast value, inventory forecast value, and loss forecast value corresponding to the time index position with the same time index value are read. The dominant source driving value is multiplied with the material demand forecast value, inventory forecast value, and loss forecast value respectively to obtain the demand explanatory value, inventory explanatory value, and loss explanatory value respectively. The demand explanatory value, inventory explanatory value, and loss explanatory value are combined to obtain the dominant explanatory vector corresponding to the time index position. The dominant explanatory vector is associated with the corresponding dominant disturbance source node and dominant propagation path to obtain the dominant explanatory value of the dominant disturbance source node corresponding to the time index position.

[0155] For each time index position in the deviation-driven time index set, the source driving value and corresponding propagation path corresponding to each secondary disturbance source node are read, and the source driving value corresponding to each secondary disturbance source node is associated with the material demand forecast value, inventory forecast value and loss forecast value corresponding to the time index position to obtain the secondary explanation value.

[0156] When generating secondary explanatory values, each time index value in the deviation-driven time index set is read one by one, and the secondary disturbance source node corresponding to the time index value is extracted from the disturbance inversion result. At the same time, the source driving value and the corresponding propagation path corresponding to each secondary disturbance source node are read to obtain the set of secondary disturbance source nodes, the corresponding source driving value sequence, and the corresponding propagation path set corresponding to the time index position. In the material forecasting result, the material demand forecast value, inventory forecast value, and loss forecast value corresponding to the time index position with the same time index value are read. For each secondary disturbance source node in the set of secondary disturbance source nodes, its corresponding source driving value is multiplied with the material demand forecast value, inventory forecast value, and loss forecast value to obtain the corresponding demand explanatory value, inventory explanatory value, and loss explanatory value, respectively. The demand explanatory value, inventory explanatory value, and loss explanatory value corresponding to each secondary disturbance source node are combined to obtain the corresponding secondary explanatory vector. Each secondary explanatory vector is associated with the corresponding secondary disturbance source node and its corresponding propagation path to obtain the secondary explanatory value of each secondary disturbance source node corresponding to the time index position.

[0157] The propagation paths corresponding to the dominant and secondary disturbance source nodes are associated and labeled, and the path activation values ​​corresponding to each propagation path are extracted to generate path interpretation results.

[0158] For each time index position in the deviation-driven time index set, the corresponding material demand forecast, inventory forecast, and loss forecast are combined with the corresponding dominant explanatory value, each secondary explanatory value, and path explanation result to obtain the explanatory forecast result.

[0159] For time index positions that do not belong to the deviation-driven time index set, the corresponding material demand forecast, inventory forecast, and loss forecast are read and associated with the corresponding time index to obtain the forecast results. The explanatory forecast results are then summarized with the forecast results to generate multidimensional explanatory forecast results.

[0160] Example 1: To verify the feasibility of this invention in practice, it was applied to the supply chain management system of a manufacturing enterprise. This enterprise involves multiple production workshops and warehousing centers, facing complex situations in daily operations such as large fluctuations in order volume, significant differences in inventory levels, and logistics transportation being affected by the external environment. Traditional material forecasting methods can only make single predictions based on historical order and inventory data, making it difficult to cope with material supply risks caused by changes in production rhythm, unforeseen events, and transportation delays. This leads to frequent occurrences of excessive inventory or shortages, affecting production efficiency and cost control. This invention addresses this problem by using multimodal data fusion and perturbation causal propagation modeling to provide enterprises with more accurate and interpretable predictions of material requirements, inventory, and losses.

[0161] In this scenario, the invention first collects internal enterprise data on order generation, inventory status, production processes, and transportation, while simultaneously integrating external disturbance event data, including climate change and sudden logistics events. During data preprocessing, missing values ​​are imputed, outliers are removed, and all data is aligned using a unified time index. Subsequently, an event-driven decomposition method is used to identify abnormal fluctuations in orders, inventory, or production processes, and these disturbance trigger signals are aggregated to form a disturbance event sequence. In the causal analysis stage, the invention correlates disturbance events with changes in material state, generating candidate disturbance events and their corresponding causal relationships, constructing an initial disturbance causal relationship graph, and further filtering through causal propagation paths to form a multimodal disturbance causal propagation network to quantify the disturbance intensity and propagation impact.

[0162] In practical applications, enterprises can fuse disturbance propagation characteristics with standardized material feature matrices to form a causal evidence fusion feature representation, which is then input into a time-series prediction network for material forecasting. Combining the forecast results with the actual material status, the disturbance inversion mechanism can identify dominant and secondary disturbance sources, tracing their impact on material demand, inventory, and losses along the propagation path. Multidimensional interpretable forecast results enable managers to clearly understand which factors cause material fluctuations and make reasonable adjustments in daily scheduling and replenishment decisions. Through the implementation of this invention, enterprises can more accurately predict material demand and inventory levels when facing order fluctuations, changes in production line load, and logistics delays. Simultaneously, they possess the ability to interpret the sources of forecast results, improving the flexibility and reliability of supply chain management and achieving dynamic matching between material supply and production plans.

[0163] To verify the performance of the present invention in practice, it was compared with traditional prediction methods, and the results are shown in Table 1.

[0164] Table 1. Performance Comparison of Multimodal Fusion Prediction Method and Traditional Prediction Method

[0165] method Forecast accuracy (material requirements) Forecast accuracy (inventory level) Inventory turnover volatility Loss rate Anomaly response time (hours) Traditional single-mode prediction 78% 74% 15% 4.2% 12 This invention 91% 88% 7% 2.1% 4

[0166] As can be seen from the comparison in Table 1, the multimodal material fusion prediction method proposed in this invention is superior to the traditional single-modal prediction method in all key indicators.

[0167] In terms of material demand forecasting accuracy, this invention improves accuracy by 13 percentage points, enabling more precise prediction of the impact of order fluctuations on production and warehousing, and reducing the risk of stockouts or overstocking due to forecasting errors. Regarding inventory level forecasting accuracy, it also improves from 74% to 88%, enhancing the precision of inventory control and allowing companies to reduce unnecessary inventory buildup while ensuring supply.

[0168] Inventory turnover volatility decreased from 15% with traditional methods to 7% with this invention, indicating that the multimodal fusion method can stabilize inventory change trends and reduce inventory instability caused by abnormal orders or transportation delays. Simultaneously, the loss rate was halved, from 4.2% to 2.1%, demonstrating that perturbation causal propagation modeling and abnormal perturbation tracking can effectively identify potential risk nodes, enabling timely control measures and reducing material loss. Abnormal response time was also shortened from 12 hours to 4 hours. This invention can quickly capture perturbation events and perform inverse analysis, allowing managers to obtain decision-making references in the shortest possible time, achieving dynamic intervention and optimized scheduling.

[0169] The performance improvement is primarily due to the invention's use of multimodal data fusion and a disturbance event-driven causal propagation network. This approach not only integrates information on business transactions, inventory status, production processes, logistics, and external disturbances, but also quantifies disturbance intensity, propagation paths, and cumulative effects. It combines the predicted results with the actual material state to provide interpretable decision-making support. This method overcomes the limitations of traditional single data sources and static prediction models, resulting in more accurate predictions, effective control of inventory and loss fluctuations, and significantly improved response speed. It fully demonstrates the value of multimodal fusion and causal analysis in supply chain management.

[0170] 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 multimodal material fusion prediction method, characterized in that, Includes the following steps: Collect multimodal data related to the target material and construct a multi-source heterogeneous material data set; The dataset of multi-source heterogeneous materials is preprocessed and then fused and stitched together according to a unified time index to generate a standardized multimodal material feature matrix. Based on the standardized multimodal material feature matrix, event-driven decomposition is performed on each modal feature to extract the corresponding disturbance trigger signal and construct a disturbance event sequence; A correlation analysis is performed on the sequence of disturbance events and the sequence of material state changes to construct a set of candidate causal relationships between disturbance events and changes in material demand, inventory and loss, and an initial disturbance causal relationship diagram is generated. The initial perturbation causal relationship graph is subjected to causal structure learning and path filtering to identify perturbation source nodes, propagation paths and terminal nodes, and a multimodal perturbation causal propagation network is constructed. Based on a multimodal perturbation causal propagation network, perturbation intensity quantification is performed on each perturbation source node, and the impact attenuation and superposition effect of each perturbation during the propagation process are calculated to generate a perturbation propagation feature set. The perturbation propagation feature set is fused with the standardized multimodal material feature matrix to construct a causal evidence fusion feature representation and generate material prediction results; Based on the material prediction results and the actual material state, perturbation inversion is performed on the multimodal perturbation causal propagation network to generate perturbation inversion results; By combining material prediction results with perturbation inversion results, multidimensional interpretive prediction results are generated.

2. The multimodal material fusion prediction method according to claim 1, characterized in that, The multimodal data includes business transaction data, inventory status data, production process data, logistics and transportation data, and external disturbance event data.

3. The multimodal material fusion prediction method according to claim 1, characterized in that, The preprocessing includes missing value imputation, outlier removal, time alignment, uniform encoding, and numerical normalization.

4. The multimodal material fusion prediction method according to claim 1, characterized in that, The construction of the perturbation event sequence specifically includes: Based on the standardized multimodal material feature matrix, business transaction features, inventory status features, production process features, logistics and transportation features, and external disturbance event features are grouped and extracted according to feature modes to form time series feature subsets of the corresponding modes; For each modal time series feature subset, sliding segmentation is performed, and the change amplitude, change rate and fluctuation intensity of feature values ​​within each time window are calculated to generate a feature change sequence; A mutation detection is performed on the feature change sequence to identify the time point that meets the disturbance triggering condition, and the multi-dimensional vector composed of the corresponding feature change amplitude, change rate and fluctuation intensity is extracted as the disturbance triggering signal; The disturbance trigger signals identified in each mode are aggregated in chronological order, and the comprehensive change intensity value and feature type value of the disturbance event segment are determined. Based on the start time, end time, duration, comprehensive change intensity, and characteristic type of each disturbance event segment, disturbance event candidates are generated, and the disturbance event candidates are arranged in chronological order to form a disturbance event sequence.

5. The multimodal material fusion prediction method according to claim 1, characterized in that, The generation of the initial perturbation causal relationship graph includes: Based on the disturbance event sequence, the start time value, end time value, duration value, comprehensive change intensity value and feature type value corresponding to each disturbance event candidate are extracted, and sorted according to a unified time index to form an ordered disturbance event sequence; Obtain the material state change sequence within the time range corresponding to the disturbance event sequence, and perform time alignment on the material state change sequence. The material state change sequence includes the material demand change sequence, inventory change sequence, and loss change sequence. For each disturbance event candidate, the corresponding candidate influence time window is determined according to the time interval relationship between its end time value and the material state change sequence. The material state change value within the candidate influence time window is extracted to generate event-state aligned data pairs. For each event-state aligned data pair, perform change correlation analysis and calculate the correlation strength between the comprehensive change intensity value of the disturbance event and the changes in material demand, inventory, and loss. The relationship between disturbance event candidates and material state changes is filtered based on the correlation strength value, and a set of causal candidate relationships between disturbance events and material state changes is constructed. The candidate causal relationships are categorized and organized according to time order and feature type values, and initial perturbation causal relationship edges are constructed with perturbation event candidates as the starting point and material state changes as the ending point. All initial disturbance causal relationship edges are summarized, and connections are established according to the chronological order of disturbance event candidates to form an initial disturbance causal relationship graph containing disturbance event nodes and material status nodes.

6. The multimodal material fusion prediction method according to claim 1, characterized in that, The generation of the multimodal perturbation causal propagation network includes: Read each initial disturbance causal relationship edge in the initial disturbance causal relationship graph, and extract the disturbance event candidate, material state change, event end time index value, feature type value and correlation strength value corresponding to each initial disturbance causal relationship edge to form an initial causal relationship edge set; Based on the event end time index value corresponding to each initial perturbation causal relationship edge, the event is arranged in chronological order to determine the temporal relationship between each perturbation event candidate and to generate the temporal connection relationship between the perturbation event candidate; Connect the perturbation event candidates with temporal connections to construct a set of candidate propagation paths; For each candidate propagation path in the candidate propagation path set, the edge weights are accumulated according to the correlation strength values ​​of each initial perturbation causal relationship edge in the path to obtain the path strength value of the corresponding candidate propagation path, and the path length value is calculated according to the number of relation edges contained in the path. Based on the path strength and path length values ​​of each candidate propagation path, a set of effective propagation paths is formed; For each candidate propagation path in the effective propagation path set, the starting point is identified, the disturbance event candidates with zero incoming edges are extracted as disturbance source nodes, the material status nodes corresponding to the end of the path are extracted as action terminal nodes, and the disturbance event candidates in the middle of the path are extracted as propagation intermediate nodes. By summarizing the source nodes of disturbances, intermediate nodes of propagation, terminal nodes of effects, and the set of effective propagation paths, a multimodal disturbance causal propagation network is constructed, which includes node connection relationships, edge weights, and path order.

7. The multimodal material fusion prediction method according to claim 1, characterized in that, The generation of the disturbance propagation feature set specifically includes: Read the source nodes, intermediate nodes, terminal nodes, and edge weights of each causal connection in the multimodal perturbation causal propagation network. For each disturbance source node, the comprehensive change intensity value of the disturbance event candidates corresponding to the disturbance source node is used as the disturbance intensity value of the disturbance source node, and as the initial propagation intensity value of the first hop of the corresponding propagation path; Along each propagation path in the multimodal perturbation causal propagation network, the edge weights and event pair propagation strength values ​​in each path are extracted and arranged in the order of the paths to form a sequence of path propagation parameters. For each propagation path, the propagation strength values ​​of each edge weight and event pair in the path propagation parameter sequence are read sequentially to generate single-hop propagation strength values ​​and form a path propagation strength sequence. The single-hop propagation intensity values ​​in the same propagation path are accumulated to obtain the cumulative path impact value of the corresponding propagation path, and the terminal single-hop propagation intensity value in the path propagation intensity sequence is extracted to obtain the terminal path impact value. For multiple propagation paths pointing to the same terminal node, the cumulative impact value and the terminal impact value of each propagation path are read respectively, and the cumulative impact value and the terminal impact value of each path are added together to obtain the superimposed impact value of the terminal node. The disturbance intensity value corresponding to each disturbance source node, the cumulative path impact value corresponding to each propagation path, and the superimposed impact value corresponding to each terminal node are summarized to generate a disturbance propagation feature set.

8. The multimodal material fusion prediction method according to claim 1, characterized in that, The generation of the material prediction results specifically includes: Read the disturbance propagation feature set, and align the disturbance intensity value, path cumulative impact value, and superimposed impact value with time according to a unified time index to obtain the disturbance propagation feature vector; Multimodal material feature vectors corresponding to each time index are extracted from the standardized multimodal material feature matrix. The perturbation propagation feature vectors are time-aligned with the multimodal material feature vectors, and then normalized and spliced ​​to generate a fused feature vector, which serves as the fused feature representation of causal evidence. The causal evidence fusion feature representations of multiple consecutive time steps are sequentially combined to construct the input feature sequence. The input feature sequence is then input into the time-series fusion Transformer prediction network to obtain the material demand forecast, inventory forecast, and loss forecast under the corresponding time index. The material demand forecast, inventory forecast, and loss forecast values ​​under each time index are arranged and summarized in chronological order to generate material forecast results.

9. The multimodal material fusion prediction method according to claim 1, characterized in that, The generation of the perturbation inversion results specifically includes: Read the material demand forecast, inventory forecast, and loss forecast values ​​corresponding to each time index in the material forecast results, and obtain the actual material demand, inventory, and loss values ​​under the corresponding time index to obtain the material demand deviation, inventory deviation, and loss deviation values. Based on a unified time index, material demand deviation, inventory deviation, and loss deviation are combined to construct a material status deviation vector and form a deviation-driven time index set. For each time index position in the deviation-driven time index set, the corresponding terminal node is located in the multimodal disturbance causal propagation network, and the superimposed influence value is obtained. The material state deviation vector and the superimposed influence value are combined to obtain the terminal driving value. Path activation is performed on each propagation path in the multimodal perturbation causal propagation network to obtain the path activation value corresponding to each propagation path, and the propagation paths are filtered to form a set of activated propagation paths. For each propagation path in the set of active propagation paths, read the disturbance source node corresponding to the starting node of the path and its corresponding disturbance intensity value, and combine the disturbance intensity value with the path activation value of the corresponding propagation path to obtain the source driving value corresponding to each disturbance source node; The source driving values ​​corresponding to each disturbance source node are sorted, and the source driving values ​​of multiple disturbance source nodes are compared under the same time index. The disturbance source node with the largest source driving value is selected as the dominant disturbance source node, and its corresponding propagation path is determined as the dominant propagation path. For disturbance source nodes that are not selected as the dominant disturbance source node, secondary disturbance source nodes are obtained by filtering based on the difference between their source driving value and the source driving value of the dominant disturbance source node, and the corresponding propagation paths are extracted. The dominant disturbance source node, secondary disturbance source nodes and their corresponding propagation paths are organized according to a unified time index and associated with the corresponding source driving values ​​to generate disturbance inversion results.

10. The multimodal material fusion prediction method according to claim 1, characterized in that, The generation of the multidimensional interpretive prediction results specifically includes: Read the material demand forecast, inventory forecast and loss forecast values ​​corresponding to each time index in the material forecast results, and read the dominant disturbance source node, secondary disturbance source node and their corresponding propagation path corresponding to each time index in the disturbance inversion results, and extract the source driving value corresponding to each disturbance source node. For each time index position in the deviation-driven time index set, the source driving value and corresponding propagation path of the corresponding dominant disturbance source node are extracted, and the source driving value is associated with the material demand forecast, inventory forecast and loss forecast corresponding to the time index position to obtain the dominant explanatory value. For each time index position in the deviation-driven time index set, the source driving value and corresponding propagation path corresponding to each secondary disturbance source node are read, and the source driving value corresponding to each secondary disturbance source node is associated with the material demand forecast value, inventory forecast value and loss forecast value corresponding to the time index position to obtain the secondary explanation value. The propagation paths corresponding to the dominant and secondary disturbance source nodes are associated and labeled, and the path activation values ​​corresponding to each propagation path are extracted to generate path interpretation results. For each time index position in the deviation-driven time index set, the corresponding material demand forecast, inventory forecast, and loss forecast are combined with the corresponding dominant explanatory value, each secondary explanatory value, and path explanation result to obtain the explanatory forecast result. For time index positions that do not belong to the deviation-driven time index set, the corresponding material demand forecast, inventory forecast, and loss forecast are read and associated with the corresponding time index to obtain the forecast results. The explanatory forecast results are then summarized with the forecast results to generate multidimensional explanatory forecast results.

Images