Intelligent material management method and system based on big data processing

Abstract

The application discloses an intelligent material management method and system based on big data processing, and relates to the technical field of material management.The method comprises the following steps: S1, acquiring multi-source data and performing preprocessing; S2, performing sliding window segmentation, calculating time lag response values, and screening effective management segments; S3, mapping the effective management segments into associated nodes, establishing a directed connection relationship according to the time lag response values, performing edge weight configuration, and constructing a material management associated graph; S4, grouping the neighborhood of the associated nodes, respectively performing convolution aggregation on the neighborhood characteristics under different connection edge types, and performing multi-round fusion updating; S5, performing K-optimal bidirectional path search, screening candidate processing chains, and performing constraint projection and sorting screening to generate a management result sequence; and S6, acquiring feedback data, and performing incremental updating on the material management associated graph.The application can realize automatic identification of material change trends and abnormal management links, and improve the accuracy of material allocation.

Description

Technical Field

[0001] This invention relates to the field of materials management technology, and in particular to an intelligent materials management method and system based on big data processing. Background Technology

[0002] As businesses expand and the number of material circulation stages increases, material management is no longer limited to simple inbound, outbound, and inventory records. In actual management, basic material information, procurement plans, requisition records, warehouse changes, execution feedback, and on-site data are often distributed across different systems. Existing material management systems typically record this data through ERP systems, warehouse management systems, or procurement management systems, and combine this with information such as inventory quantities, historical consumption, and planned demand to complete management tasks such as material inquiries, early warnings, replenishment, and allocation.

[0003] While existing technologies enable information-based management of material data, most systems still rely primarily on manual rules and fixed thresholds for judgment. For example, a replenishment reminder is triggered when inventory falls below a set threshold, or a backlog warning is issued when a certain type of material has not been circulated for an extended period. While these methods are simple to implement, they are poorly adaptable to changes in material demand, usage patterns, and cross-stage influences, making it difficult to reflect continuous changes in the actual management status in a timely manner.

[0004] Meanwhile, material data from different sources differ in field format, recording time, and update frequency. Existing systems typically only perform simple data summaries or sort by time, making it difficult to detect delays and correlations between different records. When an anomaly occurs in a management process, its impact often manifests gradually in subsequent processes. Relying solely on a single record or indicator for judgment can easily lead to untimely warnings, inaccurate handling plans, or unreasonable resource allocation.

[0005] Furthermore, existing materials management systems typically save the execution results as ordinary records after generating processing plans, rarely using feedback to continuously revise subsequent judgment rules. When the pace of materials flow, demand structure, or execution environment changes, the system struggles to adjust its analysis criteria in a timely manner, causing subsequent management results to still rely on old rules, thus affecting the accuracy and adaptability of materials management.

[0006] Therefore, how to provide an intelligent material management method and system based on big data processing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent materials management method and system based on big data processing. This invention fully utilizes big data preprocessing, time-delay correlation analysis, graph structure modeling, relational graph convolution, and feedback incremental update technology. It describes in detail the intelligent materials management process from unified organization of multi-source materials data, identification of management change relationships, screening of candidate processing schemes to execution feedback correction. It has the advantages of high data processing consistency, timely response to materials changes, high matching degree of processing schemes, and strong continuous adaptability.

[0008] An intelligent materials management method based on big data processing according to an embodiment of the present invention includes the following steps: S1. Acquire multi-source data in the material management scenario and preprocess it to form a standard management sequence; S2. The standard management sequence is divided into multiple management segments by a sliding window. The time delay response values ​​between management segments are calculated, and the effective management segments are screened out using mutual information metric. S3. Map effective management segments to associated nodes, establish directed connections based on the time delay response values ​​between management segments, classify connection edge types according to the segment association determination results, configure edge weights for directed connections, and construct a material management association graph. S4. In the material management association graph, the neighborhood of the associated node is grouped according to the type of connection edge. A relational graph convolutional network is used to perform convolutional aggregation on the neighborhood features under different connection edge types. The aggregation results of different neighborhoods are then fused and updated in multiple rounds through gated state recursion to obtain the node processing score. S5. Calculate the path cost based on the node processing score and the weight of the connecting edge, perform K-optimal bidirectional path search in the material management association graph, screen out candidate processing chains, and perform constraint projection and sorting screening on the candidate processing chains based on multi-objective constraints to generate a management result sequence. S6. Obtain the feedback data corresponding to the management result sequence, and use the recursive least squares method to incrementally update the material management association diagram.

[0009] Optionally, the preprocessing specifically includes: Field mapping is performed on multi-source data, and a unified index is built according to object identifier, action identifier, and time identifier, where: The multi-source data consists of basic material records, material circulation records, planned demand records, execution feedback records, and collection log records. The object identifier is used to mark the same material management object, the action identifier is used to mark the management action corresponding to the data record, and the time identifier is used to mark the order in which the data records occur. Data records from different sources are linked and merged based on a unified index, and the merged data records are sorted by time to obtain an initial management sequence; Missing fields in the initial management sequence are filled with adjacent windows, duplicate records are deduplicated by hashing, and records with abnormal values ​​are truncated by digit division. The processed data records are time-aligned according to the preset time granularity, and the data records within adjacent time granularities are aggregated and encoded to obtain a standard management sequence.

[0010] Optionally, S2 specifically includes: S21. According to the preset window length and sliding step, the sequence items in each segmented window are cut out in the order of the standard management sequence. The sequence items in each segmented window are arranged according to the time identifier, and the field values ​​of each sequence item are spliced ​​together in a fixed field order to obtain multiple management segments. S22. Using any two management segments as the test segment pair, select one management segment as the reference segment, and perform forward and backward translations on the other management segment according to a preset lag step size. Determine the overlap interval under each lag step size, where: The preset lag step size uses the time granularity of the standard management sequence as the step size unit, including negative lag, zero lag and positive lag. The maximum lag step size is less than the length of the management segment, and the lag step size that does not form an overlapping interval does not participate in the response calculation. S23. Within each overlapping interval, the segment vectors of the test segment pairs are standardized, the difference distance between the segment vectors is calculated, and the response matching value is determined based on the difference distance. The lag step size corresponding to the maximum response matching value is taken as the target lag step size, and the maximum response matching value is taken as the time delay response value. S24. Align the test segments according to the target lag step length, classify the aligned segment vectors into discrete value intervals, calculate the mutual information metric based on the frequency of occurrence of the discrete value intervals and the frequency of synchronous occurrence, and determine the management segments that meet the screening conditions for both the time delay response value and the mutual information metric as valid management segments.

[0011] Optionally, S23 specifically includes: S231. For the overlapping interval under each lag step, extract the segment vector at the corresponding position in the segment pair to be tested, and form a reference vector sequence and a matching vector sequence according to the field order. S232. Calculate the mean and variance of each field in the baseline vector sequence and the vector sequence to be matched, and standardize the corresponding field values ​​based on the mean and variance to obtain the standardized baseline vector sequence and the standardized vector sequence to be matched. S233. Calculate the difference distance between the standardized reference vector sequence and the standardized matching vector sequence according to the corresponding positions, and determine the response matching value under the current lag step size based on the inverse relationship of the difference distance from small to large. S234. Traverse all response matching values ​​corresponding to the lag step length, determine the maximum response matching value as the time-delay response value, and determine the lag step length corresponding to the maximum response matching value as the target lag step length.

[0012] Optionally, S3 specifically includes: S31. Perform deduplication on the valid management segments, take the segment vector of each valid management segment as the node attribute, and generate the corresponding associated nodes according to the window position of the valid management segment in the standard management sequence. S32. For each pair of segments to be tested in the segment association determination result, determine the connection direction between the corresponding associated nodes based on the target lag step size: When the target lag step is positive, the associated node corresponding to the reference segment points to the associated node corresponding to the segment to be matched. When the target lag step is negative, the associated node corresponding to the segment to be matched points to the associated node corresponding to the reference segment. When the target lag step is zero, a bidirectional directed connection is established between the associated node corresponding to the reference segment and the associated node corresponding to the segment to be matched. S33. Classify the connection edge types of directed connections based on the value of the target lag step: Mark directed connections with positive target lag steps as positive time-delay connections; Directed connections with negative target lag steps are marked as reverse time-delay connections; Mark directed connections with a target lag step of zero as synchronous connection edges; S34. Based on the time delay response value, mutual information metric value and target lag step size, the directed connection relationship is coupled and weighted to obtain the edge weight value. Based on the associated node, directed connection relationship, connection edge type and edge weight value, a material management association graph is constructed.

[0013] Optionally, S4 specifically includes: S41. Take each associated node as the current processing node, map the node attributes of the current processing node to the initial node features, and extract the adjacent associated nodes corresponding to the current processing node according to the connection edge type to form the type neighborhood of the corresponding connection edge type. S42. For each type of neighborhood, generate edge weight gating coefficients based on the edge weights of the directed connection relationship, adjust the node features of adjacent associated nodes using the edge weight gating coefficients, and use the relational convolution kernel corresponding to the connection edge type to perform convolution propagation on the adjusted node features to obtain the type convolution result. S43. Perform residual calculation on the convolution results corresponding to different connection edge types and the previous round node features of the current processing node. Generate a fusion gate value based on the residual calculation results, and recursively correct the previous round node features according to the fusion gate value to obtain the current round node features. S44. Use the current round node features as the input features for the next round of recursion, repeat the convolutional propagation and recursion correction until the preset recursion round is reached, and map the last round node features to the node processing score.

[0014] Optionally, S42 specifically includes: S421. For each type of neighborhood, read the edge weights of the directed connections within the neighborhood and the node features of the adjacent nodes, linearly combine the edge weights with the gating parameters corresponding to the type of the connection edge, and obtain the edge weight gating value through S-shaped function mapping. S422. Based on the edge weight gating value, channel gating is performed on the node features of adjacent associated nodes. Feature channels that match the current connection edge type are retained, and feature channels that deviate from the current connection edge type are suppressed to obtain gated neighborhood features. S423. Take the node features of the current processing node as the zero-order convolution term, take the result of the gated neighborhood features after passing along the directed connection relationship as the first-order convolution term, and construct the second-order convolution term based on the result of the second passing of the first-order convolution term. S424. Using the relational convolution kernel corresponding to the current connection edge type, perform convolution mapping on the zero-order convolution term, the first-order convolution term, and the second-order convolution term respectively, and perform gated fusion on the convolution mapping results according to the convolution order to obtain the type convolution result corresponding to the current connection edge type.

[0015] Optionally, S5 specifically includes: S51. Generate node cost potential based on node processing score, generate edge transfer cost based on edge weight, and write node cost potential, edge transfer cost and processing chain length penalty into path cost recursion formula, and gradually update the path cost of each processing chain as the associated nodes expand. S52. Perform bidirectional multi-label search in the material management association graph. The forward label is recursively pushed along the directed connection relationship, and the reverse label is recursively pushed along the reverse of the directed connection relationship. Each search label records the end associated node, the cumulative path cost, the sequence of traversed nodes, and the constraint residual vector. S53. Filter the search labels that reach the same end associated node, compare the cumulative path cost and constraint residual value of each search label, retain the search labels with smaller cumulative path cost and smaller constraint residual value, and generate candidate processing chains based on the retained forward and reverse labels. S54. Convert the multi-objective constraints into Lagrange penalty terms, calculate the link evaluation value based on the cumulative path cost of the candidate processing chain and the Lagrange penalty term, and select K candidate processing chains according to the link evaluation values ​​to generate a management result sequence.

[0016] Optionally, S6 specifically includes: S61. Obtain the feedback data corresponding to the management result sequence. According to the candidate processing chain in the management result sequence, extract the associated nodes, connection edge weights, and node processing scores that the feedback data passes through, and encode the extracted results into a feedback observation vector, where: The feedback data consists of: result confirmation records, status change records, abnormal return records, and delayed response records, which are used to characterize the completion status, segment changes, execution deviations, and response lags after the execution of the management result sequence. S62. Calculate the actual feedback response based on the feedback observation vector, and calculate the predicted feedback response based on the current connection edge weights and gate state recursive parameters. Determine the deviation between the actual feedback response and the predicted feedback response as the execution residual. S63. Based on the feedback observation vector and execution residual, perform recursive least squares estimation, calculate the covariance recursive amount and parameter update gain, and generate the connection edge weight correction amount and the gated state recursive parameter correction amount based on the parameter update gain. S64. Based on the adjustment amount of the connection edge weight, incrementally adjust the connection edge weight in the material management association diagram, and update the gate status recursive parameter in the node processing scoring process based on the adjustment amount of the gate status recursive parameter.

[0017] An intelligent materials management system based on big data processing according to an embodiment of the present invention includes: The preprocessing module is used to acquire multi-source data in the material management scenario and preprocess it to form a standard management sequence; The segment selection module is used to perform sliding window segmentation on the standard management sequence to obtain multiple management segments, calculate the time delay response values ​​between management segments, and use mutual information metric to select effective management segments. The graph construction module is used to map effective management segments to associated nodes, establish directed connections based on the time delay response values ​​between management segments, classify the connection edge types according to the segment association determination results, configure the edge weights of the directed connections, and construct the material management association graph. The graph recursion module is used to group the neighborhoods of associated nodes according to the type of connection edge in the material management association graph. It uses a relational graph convolutional network to perform convolutional aggregation on the neighborhood features under different connection edge types, and performs multiple rounds of fusion and update of different neighborhood aggregation results through gating state recursion to obtain the node processing score. The path filtering module is used to calculate the path cost based on the node processing score and the weight of the connecting edge, perform K-optimal bidirectional path search in the material management association graph, screen candidate processing chains, and perform constraint projection and sorting filtering on the candidate processing chains based on multi-objective constraints to generate a management result sequence. The feedback update module is used to obtain feedback data corresponding to the management result sequence and incrementally update the material management association diagram using the recursive least squares method.

[0018] The beneficial effects of this invention are: First, this invention performs unified preprocessing on multi-source material management data to form a continuous standard management sequence, reducing the differences in field format, time granularity and recording order of data from different systems, and improving the consistency and usability of data processing.

[0019] Secondly, this invention identifies the delay effects and correlation changes between different management links through time delay response calculation, mutual information screening, and correlation graph analysis, and generates a management result sequence that matches the current management status, thereby improving the accuracy and timeliness of material handling plans.

[0020] Finally, the present invention incrementally updates the correlation and recursive parameters based on feedback data, enabling the system to continuously correct the analysis basis as material demand, circulation rhythm and execution results change, thereby improving the dynamic adaptability and continuous operation stability of material management. Attached Figure Description

[0021] 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: Figure 1 This is an overall flowchart of an intelligent materials management method based on big data processing proposed in this invention; Figure 2 This invention presents a flowchart of the relationship graph convolution and gating state recursion of an intelligent material management method based on big data processing. Figure 3 This is a module structure diagram of an intelligent materials management system based on big data processing proposed in this invention. Detailed Implementation

[0022] 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.

[0023] refer to Figures 1-2 An intelligent materials management method based on big data processing includes the following steps: S1. Acquire multi-source data in the material management scenario and preprocess it to form a standard management sequence; S2. The standard management sequence is divided into multiple management segments by a sliding window. The time delay response values ​​between management segments are calculated, and the effective management segments are screened out using mutual information metric. S3. Map effective management segments to associated nodes, establish directed connections based on the time delay response values ​​between management segments, classify connection edge types according to the segment association determination results, configure edge weights for directed connections, and construct a material management association graph. S4. In the material management association graph, the neighborhood of the associated node is grouped according to the type of connection edge. A relational graph convolutional network is used to perform convolutional aggregation on the neighborhood features under different connection edge types. The aggregation results of different neighborhoods are then fused and updated in multiple rounds through gated state recursion to obtain the node processing score. S5. Calculate the path cost based on the node processing score and the weight of the connecting edge, perform K-optimal bidirectional path search in the material management association graph, screen out candidate processing chains, and perform constraint projection and sorting screening on the candidate processing chains based on multi-objective constraints to generate a management result sequence. S6. Obtain the feedback data corresponding to the management result sequence, and use the recursive least squares method to incrementally update the connection edge weights and gate status recursive parameters of the material management association graph.

[0024] In this embodiment, the preprocessing specifically includes: Field mapping is performed on multi-source data, and a unified index is built according to object identifier, action identifier, and time identifier, where: Multi-source data consists of basic material records, material circulation records, planned demand records, execution feedback records, and collection log records. Object identifiers are used to mark the same material management object, action identifiers are used to mark the management action corresponding to the data record, and time identifiers are used to mark the order in which the data record occurs. Data records from different sources are linked and merged based on a unified index, and the merged data records are sorted by time to obtain an initial management sequence; Missing fields in the initial management sequence are filled with adjacent windows, duplicate records are deduplicated by hashing, and records with abnormal values ​​are truncated by digit division. The processed data records are time-aligned according to the preset time granularity, and the data records within adjacent time granularities are aggregated and encoded to obtain a standard management sequence. In this embodiment, the preset time granularity is set to 1 hour; When processing missing fields in the initial management sequence, the data records corresponding to the same object identifier within three time granularities before and after the missing field are selected as the center for neighboring window filling. If no usable field value is available within three time granularities before and after, the median value of the 10 most recent valid data records under the same object identifier will be used to fill the gap. When processing duplicate records, the object identifier, action identifier, and time identifier are combined into a ternary hash key, and only one data record with the same ternary hash key is retained; When processing records with abnormal values, the 1st quantile and the 99th quantile are used as cutoff boundaries; After completing the above processing, the data records within the same time granularity are aggregated and encoded according to a fixed field order to form a 32-dimensional management status vector, and the continuously arranged management status vectors are used as the standard management sequence.

[0025] In this embodiment, S2 specifically includes: S21. According to the preset window length and sliding step, the sequence items in each segmented window are cut out in the order of the standard management sequence. The sequence items in each segmented window are arranged according to the time identifier, and the field values ​​of each sequence item are spliced ​​together in a fixed field order to obtain multiple management segments. S22. Using any two management segments as the test segment pair, select one management segment as the reference segment, and perform forward and backward translations on the other management segment according to a preset lag step size. Determine the overlap interval under each lag step size, where: The preset lag step size uses the time granularity of the standard management sequence as the step size unit, including negative lag, zero lag, and positive lag. The maximum lag step size is less than the length of the management segment, and the lag step size that does not form an overlapping interval is not included in the response calculation. S23. Within each overlapping interval, the segment vectors of the test segment pairs are standardized, the difference distance between the segment vectors is calculated, and the response matching value is determined based on the difference distance. The lag step size corresponding to the maximum response matching value is taken as the target lag step size, and the maximum response matching value is taken as the time delay response value. S24. Align the test segments according to the target lag step length, classify the aligned segment vectors into discrete value intervals, calculate the mutual information metric based on the frequency of occurrence of the discrete value intervals and the frequency of synchronous occurrence, and determine the management segments that meet the screening conditions for both the time delay response value and the mutual information metric as valid management segments.

[0026] In this embodiment, S23 specifically includes: S231. For the overlapping interval under each lag step, extract the segment vector at the corresponding position in the segment pair to be tested, and form a reference vector sequence and a matching vector sequence according to the field order. S232. Calculate the mean and variance of each field in the baseline vector sequence and the vector sequence to be matched, and standardize the corresponding field values ​​based on the mean and variance to obtain the standardized baseline vector sequence and the standardized vector sequence to be matched. S233. Calculate the difference distance between the standardized reference vector sequence and the standardized matching vector sequence according to the corresponding positions, and determine the response matching value under the current lag step size based on the inverse relationship of the difference distance from small to large. S234. Traverse all response matching values ​​corresponding to the lag step length, determine the maximum response matching value as the time-delay response value, and determine the lag step length corresponding to the maximum response matching value as the target lag step length.

[0027] In this embodiment, S3 specifically includes: S31. Perform deduplication on the valid management segments, take the segment vector of each valid management segment as the node attribute, and generate the corresponding associated nodes according to the window position of the valid management segment in the standard management sequence. S32. For each pair of segments to be tested in the segment association determination result, determine the connection direction between the corresponding associated nodes based on the target lag step size: When the target lag step is positive, the associated node corresponding to the reference segment points to the associated node corresponding to the segment to be matched. When the target lag step is negative, the associated node corresponding to the segment to be matched points to the associated node corresponding to the reference segment. When the target lag step is zero, a bidirectional directed connection is established between the associated node corresponding to the reference segment and the associated node corresponding to the segment to be matched. S33. Classify the connection edge types of directed connections based on the value of the target lag step: Mark directed connections with positive target lag steps as positive time-delay connections; Directed connections with negative target lag steps are marked as reverse time-delay connections; Mark directed connections with a target lag step of zero as synchronous connection edges; S34. Based on the time delay response value, mutual information metric value and target lag step size, the directed connection relationship is coupled and weighted to obtain the edge weight value. Based on the associated node, directed connection relationship, connection edge type and edge weight value, a material management association graph is constructed.

[0028] In this embodiment, S4 specifically includes: S41. Take each associated node as the current processing node, map the node attributes of the current processing node to the initial node features, and extract the adjacent associated nodes corresponding to the current processing node according to the connection edge type to form the type neighborhood of the corresponding connection edge type. S42. For each type of neighborhood, generate edge weight gating coefficients based on the edge weights of the directed connection relationship, adjust the node features of adjacent associated nodes using the edge weight gating coefficients, and use the relational convolution kernel corresponding to the connection edge type to perform convolution propagation on the adjusted node features to obtain the type convolution result. S43. Perform residual calculation on the convolution results corresponding to different connection edge types and the previous round node features of the current processing node. Generate a fusion gate value based on the residual calculation results, and recursively correct the previous round node features according to the fusion gate value to obtain the current round node features. S44. Use the current round node features as the input features for the next round of recursion, repeat the convolutional propagation and recursion correction until the preset recursion round is reached, and map the last round node features to the node processing score.

[0029] In this embodiment, S42 specifically includes: S421. For each type of neighborhood, read the edge weights of the directed connections within the neighborhood and the node features of the adjacent nodes, linearly combine the edge weights with the gating parameters corresponding to the type of the connection edge, and obtain the edge weight gating value through S-shaped function mapping. S422. Based on the edge weight gating value, channel gating is performed on the node features of adjacent associated nodes. Feature channels that match the current connection edge type are retained, and feature channels that deviate from the current connection edge type are suppressed to obtain gated neighborhood features. S423. Take the node features of the current processing node as the zero-order convolution term, take the result of the gated neighborhood features after passing along the directed connection relationship as the first-order convolution term, and construct the second-order convolution term based on the result of the second passing of the first-order convolution term. S424. Using the relational convolution kernel corresponding to the current connection edge type, perform convolution mapping on the zero-order convolution term, the first-order convolution term, and the second-order convolution term respectively, and perform gated fusion on the convolution mapping results according to the convolution order to obtain the type convolution result corresponding to the current connection edge type. In this embodiment, the node features of each associated node are 128-dimensional. For forward time-delay connection edges, reverse time-delay connection edges, and synchronous connection edges, a set of gating parameters and a set of relational convolution kernels are configured respectively. After linear combination and sigmoid function mapping, the edge weights are obtained as edge weight gate values ​​ranging from 0 to 1; The edge weight gate value is expanded into a 128-dimensional channel gate vector. The original feature value is retained for feature channels with a channel gate value of not less than 0.5, and the feature channel with a channel gate value of less than 0.5 is attenuated by a factor of 0.2 to obtain the gated neighborhood features. When constructing the convolution term, the node features of the current processing node are used as the zero-order convolution term. The gated neighborhood features are passed along the directed connection relationship once to obtain the first-order convolution term. The first-order convolution term is then passed along the directed connection relationship once to obtain the second-order convolution term. The zero-order, first-order, and second-order convolutional terms are mapped to 128-dimensional convolutional features by relational convolutional kernels of the corresponding connection edge types, and then fused according to the convolutional order gating value. The convolution order gate value is normalized to between 0 and 1 after being calculated by the sigmoid function, and after fusion, the type convolution result corresponding to the current connection edge type is obtained.

[0030] In this embodiment, S5 specifically includes: S51. Generate node cost potential based on node processing score, generate edge transfer cost based on edge weight, and write node cost potential, edge transfer cost and processing chain length penalty into path cost recursion formula, and gradually update the path cost of each processing chain as the associated nodes expand. S52. Perform bidirectional multi-label search in the material management association graph. The forward label is recursively pushed along the directed connection relationship, and the reverse label is recursively pushed along the reverse of the directed connection relationship. Each search label records the end associated node, the cumulative path cost, the sequence of traversed nodes, and the constraint residual vector. S53. Filter the search labels that reach the same end associated node, compare the cumulative path cost and constraint residual value of each search label, retain the search labels with smaller cumulative path cost and smaller constraint residual value, and generate candidate processing chains based on the retained forward and reverse labels. S54. Convert the multi-objective constraints into Lagrange penalty terms, calculate the link evaluation value based on the cumulative path cost of the candidate processing chains and the Lagrange penalty terms, and select K candidate processing chains according to the link evaluation values ​​to generate a management result sequence, specifically including: The multi-objective constraint is broken down into multiple constraint terms, and a constraint direction and constraint threshold are set for each constraint term; For each candidate processing chain, the actual values ​​of each constraint term are calculated based on the node processing score, connection edge weight, and link length of the associated nodes within the chain. The deviation between the actual value of each constraint term and the corresponding constraint threshold is calculated: For upper limit constraints, only the portion exceeding the constraint threshold is considered as the constraint deviation value; For lower bound constraints, only the portion below the constraint threshold is considered as the constraint deviation value; When the constraint threshold is met, the corresponding constraint deviation value is recorded as zero; Configure corresponding Lagrange coefficients for each constraint term, and convert each constraint deviation value into a penalty value based on the Lagrange coefficients to obtain the Lagrange penalty term corresponding to the candidate processing chain; Substituting the cumulative path cost and Lagrange penalty term of the candidate processing chain into the link evaluation formula, we obtain the link evaluation value; Candidate processing chains are sorted from smallest to largest according to their link evaluation values, the top K candidate processing chains are selected, and the selected candidate processing chains are converted into a management result sequence.

[0031] In this embodiment, S6 specifically includes: S61. Obtain the feedback data corresponding to the management result sequence. According to the candidate processing chain in the management result sequence, extract the associated nodes, connection edge weights, and node processing scores that the feedback data passes through, and encode the extracted results into a feedback observation vector, where: The feedback data consists of: result confirmation records, status change records, exception return records, and delayed response records, which are used to characterize the completion status, segment changes, execution deviations, and response lags after the execution of the management result sequence. S62. Calculate the actual feedback response based on the feedback observation vector, and calculate the predicted feedback response based on the current connection edge weights and gate state recursive parameters. Determine the deviation between the actual feedback response and the predicted feedback response as the execution residual. S63. Based on the feedback observation vector and execution residual, perform recursive least squares estimation, calculate the covariance recursive amount and parameter update gain, and generate the connection edge weight correction amount and the gated state recursive parameter correction amount based on the parameter update gain. S64. Based on the adjustment amount of the connection edge weight, incrementally adjust the connection edge weight in the material management association diagram, and update the gate status recursive parameter in the node processing scoring process based on the adjustment amount of the gate status recursive parameter.

[0032] refer to Figure 3 An intelligent materials management system based on big data processing includes: The preprocessing module is used to acquire multi-source data in the material management scenario and preprocess it to form a standard management sequence; The segment selection module is used to perform sliding window segmentation on the standard management sequence to obtain multiple management segments, calculate the time delay response values ​​between management segments, and use mutual information metric to select effective management segments. The graph construction module is used to map effective management segments to associated nodes, establish directed connections based on the time delay response values ​​between management segments, classify the connection edge types according to the segment association determination results, configure the edge weights of the directed connections, and construct the material management association graph. The graph recursion module is used to group the neighborhoods of associated nodes according to the type of connection edge in the material management association graph. It uses a relational graph convolutional network to perform convolutional aggregation on the neighborhood features under different connection edge types, and performs multiple rounds of fusion and update of different neighborhood aggregation results through gating state recursion to obtain the node processing score. The path filtering module is used to calculate the path cost based on the node processing score and the weight of the connecting edge, perform K-optimal bidirectional path search in the material management association graph, screen candidate processing chains, and perform constraint projection and sorting filtering on the candidate processing chains based on multi-objective constraints to generate a management result sequence. The feedback update module is used to obtain the feedback data corresponding to the management result sequence, and to incrementally update the connection edge weights and gate status recursive parameters of the material management association graph using the recursive least squares method.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a production support material management scenario. This scenario involves managing a variety of materials, including both routine consumables and critical materials with low turnover rates but significant impacts on production continuity.

[0034] The original management method mainly relied on inventory ledgers, requisition records, and manual threshold reminders. The data update times of different systems were inconsistent, and the system records for some materials were delayed after they were actually requisitioned, causing managers to often only replenish the materials after shortages were discovered. At the same time, although the book quantity of some materials was sufficient, difficulties in temporary allocation still arose due to continuous use, abnormal returns, or concentrated release of subsequent demand.

[0035] In this scenario, the basic material records, circulation records, planned demand records, execution feedback records, and data collection logs within a continuous operating cycle are selected as the processing objects, involving a total of 1240 material codes and approximately 128,000 original data records. Using this invention, data from different sources are first organized according to object identifiers, action identifiers, and time identifiers, and a standard management sequence is formed with a 1-hour time granularity. Then, 12 time granularities are used as sliding windows to calculate the time lag response between management segments, with the lag step size set to -6 to +6. When the time lag response value is not less than 0.55 and the mutual information metric value is not less than 0.15, the corresponding segment is included in subsequent correlation analysis.

[0036] In practical applications, this invention does not trigger alerts solely based on a single inventory quantity, but instead incorporates the delayed changes between different management segments into the calculation. For a certain type of frequently turnoverable material, the system identifies that changes in planned demand typically affect requisition records after 3 to 4 time granularities, further influencing subsequent replenishment processing. For a certain type of material with a high number of abnormal returns, the system identifies a strong correlation between abnormal return records and subsequent secondary allocation. Based on the above analysis, the system constructs a material management relationship graph and obtains node processing scores through graph convolution and gating state recursion.

[0037] The node feature dimension is set to 128 dimensions, the recursion round is set to 3 rounds, the number of candidate processing chains K is set to 10, and the processing chain length is limited to 2 to 8 associated nodes. After filtering by path cost and constraints, the system forms a management result sequence to assist in generating supplementary, allocation, review, and anomaly handling suggestions.

[0038] To verify the effectiveness, the original manual rule-based approach was compared with that of this invention. The results show that the data field consistency rate was 82.6% with the original method, which increased to 97.8% after preprocessing with this invention. The original method took an average of 4.8 hours to identify abnormal workflows, which was reduced to 1.3 hours with this invention. The matching rate between the processing suggestions generated by the original method and the subsequent actual execution results was 74.9%, which increased to 91.2% with this invention. Regarding feedback correction, the original method required an average of 136 manual rule adjustments per running cycle, which was reduced to 52 with this invention. For delay-related issues, the original method achieved an accuracy rate of 61.5%, which increased to 87.6% with this invention.

[0039] Table 1 Comparison of the Implementation Effects of Materials Management

[0040] As can be seen from the data in Table 1, the present invention can more accurately identify material change trends and abnormal management processes in scenarios where there are inconsistencies in multi-source data, lags in management processes, and continuous changes in feedback results, and generate management results that are more consistent with the actual execution status.

[0041] The above description is only a preferred embodiment 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. An intelligent materials management method based on big data processing, characterized in that, Includes the following steps: S1. Acquire multi-source data in the material management scenario and preprocess it to form a standard management sequence; S2. The standard management sequence is divided into multiple management segments by a sliding window. The time delay response values ​​between management segments are calculated, and the effective management segments are screened out using mutual information metric. S3. Map effective management segments to associated nodes, establish directed connections based on the time delay response values ​​between management segments, classify connection edge types according to the segment association determination results, configure edge weights for directed connections, and construct a material management association graph. S4. In the material management association graph, the neighborhood of the associated node is grouped according to the type of connection edge. The relationship graph convolutional network is used to perform convolutional aggregation on the neighborhood features under different connection edge types. The aggregation results of different neighborhoods are then fused and updated in multiple rounds through gated state recursion to obtain the node processing score. S5. Calculate the path cost based on the node processing score and the weight of the connecting edge, perform K-optimal bidirectional path search in the material management association graph, screen out candidate processing chains, and perform constraint projection and sorting screening on the candidate processing chains based on multi-objective constraints to generate a management result sequence. S6. Obtain the feedback data corresponding to the management result sequence, and use the recursive least squares method to incrementally update the material management association diagram.

2. The intelligent materials management method based on big data processing according to claim 1, characterized in that, The preprocessing specifically includes: Field mapping is performed on multi-source data, and a unified index is built according to object identifier, action identifier, and time identifier, where: The multi-source data consists of basic material records, material circulation records, planned demand records, execution feedback records, and collection log records. The object identifier is used to mark the same material management object, the action identifier is used to mark the management action corresponding to the data record, and the time identifier is used to mark the order in which the data records occur. Data records from different sources are linked and merged based on a unified index, and the merged data records are sorted by time to obtain an initial management sequence; Missing fields in the initial management sequence are filled with adjacent windows, duplicate records are deduplicated by hashing, and records with abnormal values ​​are truncated by digit division. The processed data records are time-aligned according to the preset time granularity, and the data records within adjacent time granularities are aggregated and encoded to obtain a standard management sequence.

3. The intelligent materials management method based on big data processing according to claim 1, characterized in that, S2 specifically includes: S21. According to the preset window length and sliding step, the sequence items in each segmented window are cut out in the order of the standard management sequence. The sequence items in each segmented window are arranged according to the time identifier, and the field values ​​of each sequence item are spliced ​​together in a fixed field order to obtain multiple management segments. S22. Using any two management segments as a pair of segments to be tested, select one management segment as the reference segment, and perform forward and backward translations on the other management segment according to a preset lag step size. Determine the overlapping interval under each lag step size, where: The preset lag step size uses the time granularity of the standard management sequence as the step size unit, including negative lag, zero lag and positive lag. The maximum lag step size is less than the length of the management segment, and the lag step size that does not form an overlapping interval does not participate in the response calculation. S23. Within each overlapping interval, the segment vectors of the test segment pairs are standardized, the difference distance between the segment vectors is calculated, and the response matching value is determined based on the difference distance. The lag step size corresponding to the maximum response matching value is taken as the target lag step size, and the maximum response matching value is taken as the time delay response value. S24. Align the test segments according to the target lag step length, classify the aligned segment vectors into discrete value intervals, calculate the mutual information metric based on the frequency of occurrence of the discrete value intervals and the frequency of occurrence of synchronization, and determine the management segments that meet the screening conditions for both the time delay response value and the mutual information metric as valid management segments.

4. The intelligent materials management method based on big data processing according to claim 3, characterized in that, S23 specifically includes: S231. For the overlapping interval under each lag step, extract the segment vector at the corresponding position in the segment pair to be tested, and form a reference vector sequence and a matching vector sequence according to the field order. S232. Calculate the mean and variance of each field in the baseline vector sequence and the vector sequence to be matched, and standardize the corresponding field values ​​based on the mean and variance to obtain the standardized baseline vector sequence and the standardized vector sequence to be matched. S233. Calculate the difference distance between the standardized reference vector sequence and the standardized vector sequence to be matched according to the corresponding positions, and determine the response matching value under the current lag step size based on the inverse relationship of the difference distance from small to large. S234. Traverse all response matching values ​​corresponding to the lag step length, determine the maximum response matching value as the time-delay response value, and determine the lag step length corresponding to the maximum response matching value as the target lag step length.

5. The intelligent materials management method based on big data processing according to claim 1, characterized in that, S3 specifically includes: S31. Perform deduplication on the valid management segments, take the segment vector of each valid management segment as the node attribute, and generate the corresponding associated nodes according to the window position of the valid management segment in the standard management sequence. S32. For each pair of segments to be tested in the segment association determination result, determine the connection direction between the corresponding associated nodes based on the target lag step size: When the target lag step is positive, the associated node corresponding to the reference segment points to the associated node corresponding to the segment to be matched. When the target lag step is negative, the associated node corresponding to the segment to be matched points to the associated node corresponding to the reference segment. When the target lag step is zero, a bidirectional directed connection is established between the associated node corresponding to the reference segment and the associated node corresponding to the segment to be matched. S33. Classify the connection edge types of directed connections based on the value of the target lag step: Mark directed connections with positive target lag steps as positive time-delay connections; Directed connections with negative target lag steps are marked as reverse time-delay connections; Mark directed connections with a target lag step of zero as synchronous connection edges; S34. Based on the time delay response value, mutual information metric value and target lag step size, the directed connection relationship is coupled and weighted to obtain the edge weight value. Based on the associated node, directed connection relationship, connection edge type and edge weight value, a material management association graph is constructed.

6. The intelligent materials management method based on big data processing according to claim 1, characterized in that, S4 specifically includes: S41. Take each associated node as the current processing node, map the node attributes of the current processing node to the initial node features, and extract the adjacent associated nodes corresponding to the current processing node according to the connection edge type to form the type neighborhood of the corresponding connection edge type. S42. For each type of neighborhood, generate edge weight gating coefficients based on the edge weights of the directed connection relationship, adjust the node features of adjacent associated nodes using the edge weight gating coefficients, and use the relational convolution kernel corresponding to the connection edge type to perform convolution propagation on the adjusted node features to obtain the type convolution result. S43. Perform residual calculation on the convolution results corresponding to different connection edge types and the previous round node features of the current processing node. Generate a fusion gate value based on the residual calculation results, and recursively correct the previous round node features according to the fusion gate value to obtain the current round node features. S44. Use the current round node features as the input features for the next round of recursion, repeat the convolutional propagation and recursion correction until the preset recursion round is reached, and map the last round node features to the node processing score.

7. The intelligent materials management method based on big data processing according to claim 6, characterized in that, S42 specifically includes: S421. For each type of neighborhood, read the edge weights of the directed connections within the neighborhood and the node features of the adjacent nodes, linearly combine the edge weights with the gating parameters corresponding to the type of the connection edge, and obtain the edge weight gating value through S-shaped function mapping. S422. Based on the edge weight gating value, channel gating is performed on the node features of adjacent associated nodes. Feature channels that match the current connection edge type are retained, and feature channels that deviate from the current connection edge type are suppressed to obtain gated neighborhood features. S423. Take the node features of the current processing node as the zero-order convolution term, take the result of the gated neighborhood features after passing along the directed connection relationship as the first-order convolution term, and construct the second-order convolution term based on the result of the second passing of the first-order convolution term. S424. Using the relational convolution kernel corresponding to the current connection edge type, perform convolution mapping on the zero-order convolution term, the first-order convolution term, and the second-order convolution term respectively, and perform gated fusion on the convolution mapping results according to the convolution order to obtain the type convolution result corresponding to the current connection edge type.

8. The intelligent materials management method based on big data processing according to claim 1, characterized in that, S5 specifically includes: S51. Generate node cost potential based on node processing score, generate edge transfer cost based on edge weight, and write node cost potential, edge transfer cost and processing chain length penalty into path cost recursion formula, and gradually update the path cost of each processing chain as the associated nodes expand. S52. Perform bidirectional multi-label search in the material management association graph. The forward label is recursively pushed along the directed connection relationship, and the reverse label is recursively pushed along the reverse of the directed connection relationship. Each search label records the end associated node, the cumulative path cost, the sequence of traversed nodes, and the constraint residual vector. S53. Filter the search labels that reach the same end associated node, compare the cumulative path cost and constraint residual value of each search label, retain the search labels with smaller cumulative path cost and smaller constraint residual value, and generate candidate processing chains based on the retained forward and reverse labels. S54. Convert the multi-objective constraints into Lagrange penalty terms, calculate the link evaluation value based on the cumulative path cost of the candidate processing chain and the Lagrange penalty term, and select K candidate processing chains according to the link evaluation values ​​to generate a management result sequence.

9. The intelligent materials management method based on big data processing according to claim 1, characterized in that, S6 specifically includes: S61. Obtain the feedback data corresponding to the management result sequence. According to the candidate processing chain in the management result sequence, extract the associated nodes, connection edge weights, and node processing scores that the feedback data passes through, and encode the extracted results into a feedback observation vector, where: The feedback data consists of: result confirmation records, status change records, abnormal return records, and delayed response records, which are used to characterize the completion status, segment changes, execution deviations, and response lags after the execution of the management result sequence. S62. Calculate the actual feedback response based on the feedback observation vector, and calculate the predicted feedback response based on the current connection edge weights and gate state recursive parameters. Determine the deviation between the actual feedback response and the predicted feedback response as the execution residual. S63. Based on the feedback observation vector and execution residual, perform recursive least squares estimation, calculate the covariance recursive amount and parameter update gain, and generate the connection edge weight correction amount and the gated state recursive parameter correction amount based on the parameter update gain. S64. Based on the adjustment amount of the connection edge weight, incrementally adjust the connection edge weight in the material management association diagram, and update the gate status recursive parameter in the node processing scoring process based on the adjustment amount of the gate status recursive parameter.

10. An intelligent materials management system based on big data processing, comprising executing the intelligent materials management method based on big data processing as described in any one of claims 1 to 9, characterized in that, include: The preprocessing module is used to acquire multi-source data in the material management scenario and preprocess it to form a standard management sequence; The segment selection module is used to perform sliding window segmentation on the standard management sequence to obtain multiple management segments, calculate the time delay response values ​​between management segments, and use mutual information metric to select effective management segments. The graph construction module is used to map effective management segments to associated nodes, establish directed connections based on the time delay response values ​​between management segments, classify the connection edge types according to the segment association determination results, configure the edge weights of the directed connections, and construct the material management association graph. The graph recursion module is used to group the neighborhoods of associated nodes according to the type of connection edge in the material management association graph. It uses a relational graph convolutional network to perform convolutional aggregation on the neighborhood features under different connection edge types, and performs multiple rounds of fusion and update of different neighborhood aggregation results through gating state recursion to obtain the node processing score. The path filtering module is used to calculate the path cost based on the node processing score and the weight of the connecting edge, perform K-optimal bidirectional path search in the material management association graph, screen candidate processing chains, and perform constraint projection and sorting filtering on the candidate processing chains based on multi-objective constraints to generate a management result sequence. The feedback update module is used to obtain feedback data corresponding to the management result sequence and incrementally update the material management association diagram using the recursive least squares method.

Images