Enterprise production data collaborative processing system based on multi-agent system

Abstract

This invention discloses a collaborative processing system for enterprise production data based on a multi-agent system, relating to the field of enterprise production data processing technology. It includes: a production fact piece binding module for binding source records from enterprise production data into a set of production fact pieces; a multi-agent verification module for configuring production agents to complete source verification and state aggregation; a fact chain construction module for forming a production succession fact chain according to production succession relationships; an improved SAITS collaborative reconstruction module for embedding a production succession fact reconstruction layer between two self-attention interpolation blocks and outputting reconstructed fact pieces; a succession closed-loop solidification module for writing verified reconstructed fact pieces into the collaborative shadow fact chain; and a collaborative result output module for outputting the collaborative processing results of enterprise production data. This invention can improve the consistency of multi-source production data collaboration and the traceability of reconstruction results.

Description

Technical Field

[0001] This invention relates to the field of enterprise production data processing technology, and in particular to an enterprise production data collaborative processing system based on a multi-agent system. Background Technology

[0002] In enterprise production processes, data acquisition systems, manufacturing execution systems, quality inspection systems, warehousing systems, and scheduling systems are typically connected simultaneously. Records generated by these different systems for the same production object often differ in source format, reporting time, status meaning, and continuity. Existing production data processing methods often employ centralized aggregation, rule-based validation, or single-time-series interpolation models, which can partially fill in missing data. However, when faced with situations such as duplicate reporting from multiple sources, misaligned production status times, inconsistencies between quality verification results and execution records, and conflicts between scheduling plans and on-site status, the only solutions are often to overwrite data according to fixed priorities or rely on manual verification. This approach struggles to preserve the mutual verification relationships between records from different sources and cannot achieve collaborative correction without disrupting the production continuity chain. Especially in multi-process continuous production scenarios, errors in a single record can propagate along the production object's flow path, leading to inconsistent production status views at the data management end, impacting production traceability, quality analysis, and scheduling decisions.

[0003] Therefore, how to provide a collaborative processing system for enterprise production data based on multi-agent systems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a collaborative processing system for enterprise production data based on a multi-agent system. This invention, through production fact piece binding, multi-agent verification, acceptance fact chain construction, and improved SAITS collaborative reconstruction, enables multi-source records in enterprise production data to enter a unified acceptance chain while retaining their source attribution. Compared to direct aggregation or ordinary temporal interpolation, this invention uniformly expresses source anomalies, state misalignments, acceptance breaks, and cross-agent conflicts as objects to be reconstructed in the production acceptance fact chain. Utilizing a production acceptance fact reconstruction layer embedded in the middle of the SAITS network, the initial hidden state output by the first self-attention interpolation block is rewritten for production acceptance, ensuring that the reconstruction result is simultaneously constrained by peer verification relationships, forward acceptance relationships, and reverse verification relationships. After the reconstructed fact pieces are output, this invention further distinguishes between verified and disputed fact pieces through a closed-loop solidification process, writing acceptable correction results into the collaborative shadow fact chain to avoid directly overwriting the original production fact pieces. This improves the consistency, completeness, and traceability of enterprise production data collaborative processing results, providing a more reliable data foundation for production tracking, quality analysis, and scheduling decisions.

[0005] According to an embodiment of the present invention, an enterprise production data collaborative processing system based on a multi-agent system includes: The production fact sheet binding module is used to respond to enterprise production data collaborative processing tasks, acquire enterprise production data, and bind the source records in the enterprise production data into a set of production fact sheets; The multi-agent verification module is used to configure corresponding production agents for different production data domains. Each production agent performs source verification and status aggregation on the production fact pieces belonging to its own data domain in the production fact piece set, forming a verified production fact piece set. The fact chain construction module is used to connect the production fact pieces in the verified production fact piece set according to the production acceptance relationship to form a production acceptance fact chain; An improved SAITS collaborative reconfiguration module is used to input the production acceptance fact chain into the improved SAITS network. The improved SAITS network embeds a production acceptance fact reconfiguration layer between the first self-attention interpolation block and the second self-attention interpolation block. The production acceptance fact reconfiguration layer rewrites the initial hidden state output by the first self-attention interpolation block, forming a production acceptance reconfiguration hidden state that enters the second self-attention interpolation block, and outputs a reconfiguration fact piece. The closed-loop solidification module is used to paste the reconstructed fact pieces back to the production acceptance fact chain, verify the acceptance closure relationship between the reconstructed fact pieces and the production acceptance fact chain, write the reconstructed fact pieces that have passed the verification into the collaborative shadow fact chain, and attach the reconstructed fact pieces that have not passed the verification to the corresponding production intelligent agent. The collaborative results output module is used to output the collaborative processing results of enterprise production data based on the production acceptance fact chain and the collaborative shadow fact chain.

[0006] Optionally, the production fact sheet binding module specifically includes: Perform source boundary identification on source records in enterprise production data and retain the original record fragments corresponding to different source records; Identify the production objects, status actions, and locations in the original record segments, and bind the original record segments belonging to the same production processing event into fact candidate segments; Perform source duplication and time misalignment checks on the fact candidate pieces. Merge duplicate reported original record fragments into the corresponding fact candidate pieces. Mark original record fragments with time misalignment as pending verification and then merge them into the corresponding fact candidate pieces. The verified fact candidate piece is encapsulated in the same position as the source tag, time tag, and state trust tag to form a production fact piece; The production fact pieces are arranged in the order of their succession to form a set of production fact pieces.

[0007] Optionally, the multi-agent verification module specifically comprises: Based on the origin of the production fact pieces in the production fact piece set, the production fact pieces are assigned to the production agents in the corresponding production data domain. Each production agent performs source consistency verification on the production fact pieces belonging to its own data domain, checks the reporting link, collection time and status meaning of the source record, and marks the source reliable fact pieces and source abnormal fact pieces; Perform state continuity verification on trusted fact pieces from the source, and mark production fact pieces whose state transitions are inconsistent with the state inheritance table of this data domain as abnormal fact pieces; For trusted fact fragments from the same production processing event, perform aggregation processing to merge production fact fragments with consistent state meanings and adjacent time positions into verified production fact fragments. The fact pieces with abnormal sources and abnormal states are retained as fact pieces to be collaboratively reconstructed, and together with the verified production fact pieces, they form a set of verified production fact pieces.

[0008] Optionally, the fact chain construction module specifically includes: From the verified set of production fact pieces, production fact pieces with production status succession relationships are selected as candidate fact pieces for succession. Perform acceptance verification on the order of state occurrence, object flow, and return confirmation in the candidate fact pieces, and mark the production acceptance front position, production acceptance position, and production acceptance return confirmation position; The completed production fact footage is linked together into production acceptance segments according to the production acceptance front position, production acceptance itself position, and production acceptance back verification position. Perform a check on the object continuity state and state closure result between adjacent production acceptance segments, and mark the production acceptance segments with broken acceptance as segments to be reconstructed; The production acceptance segments are connected in sequence according to their arrangement. Acceptance pointers are configured between adjacent production acceptance segments, and the original position of the segment to be reconstructed in the chain is preserved to form a production acceptance fact chain.

[0009] Optionally, the improved SAITS network includes an input embedding structure, a first self-attention interpolation block, a production fact reconstruction layer, a second self-attention interpolation block, and an interpolation output structure connected in sequence, and trainable network parameters are configured in the input embedding structure, the first self-attention interpolation block, the production fact reconstruction layer, the second self-attention interpolation block, and the interpolation output structure; the production fact reconstruction layer is embedded between the output end of the first self-attention interpolation block and the input end of the second self-attention interpolation block, and both the first self-attention interpolation block and the second self-attention interpolation block retain the self-attention interpolation block structure in the SAITS network.

[0010] Optionally, the trainable network parameters are trained as follows: Complete production fact chains, missing production fact chains, and conflicting production fact chains are compiled from historical enterprise production data, and these production fact chains are converted into training fact chain samples. Configure the chain bit markers to be reconstructed for the training fact chain samples. The chain bit markers to be reconstructed cover the randomly masked chain bits, the true missing chain bits, the conflicting fact piece chain bits, and the chain bits with low confidence sources. Configure the production fact sheet before masking, the production fact sheet after manual verification, and the original credible production fact sheet as the reconstruction supervision sheet, the correction supervision sheet, and the maintenance supervision sheet, respectively. According to the training batch, the training fact chain samples are input into the improved SAITS network, and then pass through the input embedding structure, the first self-attention interpolation block, the production receiving fact reconstruction layer, the second self-attention interpolation block and the interpolation output structure in sequence to obtain the training reconstructed fact pieces. The training reconstruction fact piece and the reconstruction supervision piece are aligned at random masking chain positions to form the interpolation reconstruction error. The training reconstructed fact pieces and the corrected supervision pieces are aligned at the true missing chain positions, conflicting fact piece chain positions, and low-confidence source chain positions to form a collaborative error correction. The training reconstructed fact piece and the holding supervision piece are aligned at non-reconstructed chain bits to form the original holding error; The interpolation reconstruction error, co-correction error, and original preservation error are combined into the total training error, and the trainable network parameters are updated in reverse based on the total training error. When the total training error corresponding to the verification fact chain sample is less than the training error threshold, or when the number of training rounds reaches the upper limit of the number of training rounds, the trainable network parameters are fixed.

[0011] Optionally, the improved SAITS collaborative reconfiguration module specifically includes: The production fact pieces in the production fact chain are arranged according to the fact piece chain position and mapped to the fact piece content embedding; The factual content embedding, chain position embedding, source credibility embedding, and reconstructed tag embedding are superimposed in the same position to form the factual hidden state sequence; The first self-attention interpolation block performs query projection, key projection, and value projection on the sequence of hidden states of the fact slice, and weights and aggregates the value hidden states according to the chain position matching relationship between the query hidden states and the key hidden states to form the initial hidden states. The production acceptance fact reconstruction layer generates a fact piece chain position index table based on the production acceptance fact chain, and performs chain position back-sorting on the initial hidden state according to the fact piece chain position index table. It then performs anchor position acceptance reconstruction around the reconstruction anchor position corresponding to the fact piece to be reconstructed, forming the production acceptance reconstruction hidden state. The second self-attention interpolation block performs a second-stage self-attention interpolation on the hidden state of production acceptance and reconstruction, forming a collaborative interpolation hidden state. The interpolation output structure restores the cooperative interpolation hidden state to the reconstructed fact piece.

[0012] Optionally, the anchor-based reconstruction around the reconstruction anchor corresponding to the fact piece to be reconstructed, forming a production-based reconstruction hidden state, specifically includes: Using the chain position corresponding to the fact piece to be reconstructed as the reconstruction anchor position, extract the corresponding mutual verification hidden state, the previous position inheritance hidden state, and the subsequent position verification hidden state from the initial hidden state after the chain position is back-rowed. The hidden state of mutual verification comes from the production fact piece that is in the same production acceptance position as the reconstruction anchor. The hidden state of the previous acceptance comes from the production fact piece that the acceptance pointer points to the reconstruction anchor. The hidden state of the subsequent verification comes from the production fact piece that the reconstruction anchor points to and has a verification relationship. After linear projection, the hidden states of mutual verification, previous inheritance, and subsequent verification maintain the same hidden state dimension as the hidden state of the reconstructed anchor position. The corresponding mutual verification hidden state and the reconstructed anchor hidden state are multiplied dimension by dimension to obtain the mutual verification consistency component, and then subtracted dimension by dimension to obtain the mutual verification deviation component. The mutual verification consistency component and the mutual verification deviation component are combined in the same position to form the corresponding mutual verification code. The previous hidden state and the reconstructed anchor hidden state are subtracted dimension by dimension, and the chain offset representation of the corresponding accepting pointer is superimposed to form the previous accepting code; The hidden state of the post-verification is subtracted from the hidden state of the reconstructed anchor position in one dimension, and the chain offset representation of the corresponding verification relationship is superimposed to form the post-verification code; The source trust representations corresponding to the same position mutual verification code, the previous position acceptance code, the subsequent position verification code, and the reconstructed anchor are concatenated and then mapped by a threshold value to form the anchor retention threshold, the previous position write threshold, and the subsequent position verification threshold. The preceding latent state is projected forward to form a forward candidate state, and the following latent state is projected backward to form a backward candidate state. The anchor position retention threshold is multiplied dimension-wise with the reconstructed anchor position hidden state. The previous write threshold is multiplied dimension-wise with the forward acceptance candidate state. The subsequent verification threshold is multiplied dimension-wise with the reverse verification candidate state. The three products are added in the same position and then written back to the reconstructed anchor position to form the production acceptance reconstructed hidden state.

[0013] Optionally, the receiving closed-loop curing module specifically includes: Establish a collaborative shadow fact chain based on the fact piece chain position and acceptance pointer in the production acceptance fact chain. The collaborative shadow fact chain retains the same chain position arrangement and acceptance pointer as the production acceptance fact chain. The reconstructed fact pieces are pasted back to the production receiving fact chain according to the corresponding chain position, and the preceding production fact pieces, following production fact pieces, and corresponding mutual verification fact pieces adjacent to the reconstructed fact pieces are located. Perform preceding succession verification on the object continuation state between the reconstructed fact piece and the preceding production fact piece; perform subsequent succession verification on the state succession result between the reconstructed fact piece and the subsequent production fact piece; and perform peer closure verification on the source mutual verification state between the reconstructed fact piece and the peer mutual verification fact piece. When the previous position acceptance verification, the subsequent position acceptance verification, and the same position closure verification all pass, the reconstructed fact piece is written into the corresponding chain position in the collaborative shadow fact chain, and the chain position pointer and source pointer of the original fact piece before reconstruction are retained in the corresponding chain position; When the verification of the previous position, the verification of the subsequent position, or the verification of the same position fails, the original fact piece before reconstruction is retained in the corresponding chain position in the collaborative shadow fact chain. The reconstructed fact piece is marked as the disputed fact piece, and the chain position, source agent, and verification type of the disputed fact piece are recorded. The disputed factual pieces are attached to the corresponding production intelligent agents, and the original factual pieces before reconstruction are retained in the production factual chain.

[0014] Optionally, the collaborative result output module specifically includes: Align the production acceptance fact chain and the collaborative shadow fact chain item by item according to the fact piece chain position, and distinguish between the original fact pieces that have not been reconstructed and the verified fact pieces written into the collaborative shadow fact chain; Configure collaborative correction tags for the verified fact pieces, and register the collaborative correction tags together with the source location, chain location, and production agent affiliation of the corresponding original fact pieces; For disputed fact pieces that are not written into the collaborative shadow fact chain, retain the original fact pieces in the production successor fact chain, and register the unverified type corresponding to the disputed fact piece as a dispute prompt mark; The enterprise production data is processed collaboratively by combining the original factual footage, verified factual footage, and dispute alerts. The collaborative processing results of enterprise production data are output to the enterprise production data management terminal, and collaborative correction markers, dispute reminder markers, and the corresponding production agent affiliation relationships are output simultaneously.

[0015] The beneficial effects of this invention are: 1. This invention first binds the source records in the enterprise's production data into production fact pieces. Then, different production intelligent agents perform source verification and status aggregation, so that data such as equipment, processes, materials, quality, warehousing, and scheduling are no longer directly mixed and summarized, but retain the source attribution, status meaning, and receiving position. By connecting and verifying the production fact pieces through the production receiving fact chain, missing, misaligned, conflicting, and abnormal states can be returned to the production receiving relationship for processing, reducing the production status breakage problem caused by simple rule coverage.

[0016] 2. This invention embeds a production acceptance fact reconstruction layer between the first and second self-attention interpolation blocks of the SAITS network. This allows the initial hidden state obtained in the first stage to undergo chain rearrangement, peer verification, forward acceptance, and reverse verification before entering the second self-attention interpolation block. This structure enables the SAITS network to no longer rely solely on ordinary time-series correlations for interpolation, but instead introduces production acceptance relationships into the intermediate hidden state rewriting process, improving the accuracy of reconstructing cross-agent conflict fact pieces and acceptance break fact pieces.

[0017] 3. This invention incorporates a closed-loop solidification process after the reconstructed fact pieces are output. Verified reconstructed fact pieces are written into the collaborative shadow fact chain, while unverified reconstructed fact pieces are linked to their corresponding production agents. By storing the original production fact chain and the collaborative shadow fact chain in parallel, it can output consistent production data processing results while preserving the source of the original and disputed fact pieces, facilitating production traceability, manual review, and auditing by the production data management end. Attached Figure Description

[0018] 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 The flowchart is a process flow of the enterprise production data collaborative processing system based on a multi-agent system proposed in this invention. Figure 2 This is a diagram of the improved SAITS network structure of the enterprise production data collaborative processing system based on a multi-agent system proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1 and Figure 2 A collaborative processing system for enterprise production data based on a multi-agent system includes: The production fact sheet binding module is used to respond to enterprise production data collaborative processing tasks, acquire enterprise production data, and bind the source records in the enterprise production data into a set of production fact sheets; The multi-agent verification module is used to configure corresponding production agents for different production data domains. Each production agent performs source verification and status aggregation on the production fact pieces belonging to its own data domain in the production fact piece set, forming a verified production fact piece set. The fact chain construction module is used to connect the production fact pieces in the verified production fact piece set according to the production acceptance relationship to form a production acceptance fact chain; An improved SAITS collaborative reconfiguration module is used to input the production acceptance fact chain into the improved SAITS network. The improved SAITS network embeds a production acceptance fact reconfiguration layer between the first self-attention interpolation block and the second self-attention interpolation block. The production acceptance fact reconfiguration layer rewrites the initial hidden state output by the first self-attention interpolation block, forming a production acceptance reconfiguration hidden state that enters the second self-attention interpolation block, and outputs a reconfiguration fact piece. The closed-loop solidification module is used to paste the reconstructed fact pieces back to the production acceptance fact chain, verify the acceptance closure relationship between the reconstructed fact pieces and the production acceptance fact chain, write the reconstructed fact pieces that have passed the verification into the collaborative shadow fact chain, and attach the reconstructed fact pieces that have not passed the verification to the corresponding production intelligent agent. The collaborative results output module is used to output the collaborative processing results of enterprise production data based on the production acceptance fact chain and the collaborative shadow fact chain.

[0021] In this embodiment, the production of the document binding module specifically includes: Perform source boundary identification on source records in enterprise production data and retain the original record fragments corresponding to different source records; Identify the production objects, status actions, and locations in the original record segments, and bind the original record segments belonging to the same production processing event into fact candidate segments; Perform source duplication and time misalignment checks on the fact candidate pieces. Merge duplicate reported original record fragments into the corresponding fact candidate pieces. Mark original record fragments with time misalignment as pending verification and then merge them into the corresponding fact candidate pieces. The verified fact candidate piece is encapsulated in the same position as the source tag, time tag, and state trust tag to form a production fact piece; The production fact pieces are arranged in the order of their succession to form a set of production fact pieces.

[0022] In the specific implementation process, after receiving the enterprise's production data, the production fact-piece binding module first identifies the source boundaries of the source records in the enterprise's production data. Source records can originate from equipment acquisition systems, manufacturing execution systems, quality inspection systems, warehousing systems, and scheduling systems. The source boundary is jointly determined by the source system identifier, acquisition node identifier, record number, message header structure, acquisition time, and reporting time. When the source system identifier of adjacent records changes, the record number becomes discontinuous, the message header structure changes, or the acquisition node switches, the data stream is split at the corresponding position to obtain multiple original record fragments. Each original record fragment retains the original fields, source system identifier, acquisition node identifier, acquisition time, reporting time, and original arrangement position to prevent the loss of source traceability information during subsequent binding.

[0023] After source boundary identification is completed, semantic localization is performed on each original record fragment. Semantic localization includes the production object, status action, and location of occurrence. The production object can be a product, semi-finished product, material batch, equipment object, or process object; the status action can be starting processing, completing processing, feeding materials, transferring to the next process, quality inspection completed, warehousing, rework, paused, or alarm; the location of occurrence can be the equipment location, process location, storage location, or quality inspection location. For structured source records, the production object, status action, and location of occurrence are determined by field names and field values; for log-type source records, the production object, status action, and location of occurrence are determined by the record template, status keyword positions, and field separators. For example, "Equipment A completes process C of batch B at 10:15" can be broken down into the production object "batch B", the status action "completed processing", and the location "equipment A / process C".

[0024] After semantic localization, original record fragments that are consistent in production object, fall within the same production processing event range, occur in the same production stage, and are collected at adjacent times are bound into fact candidate fragments. The fact candidate fragments retain the source location and original fields of each original record fragment. Records from different systems under the same production processing event are grouped together into the fact candidate fragments. For example, if the equipment acquisition system records "equipment processing completed," the manufacturing execution system records "process status completed," and the quality inspection system records "task to be inspected generated," and these three types of records correspond to the same production object and the same processing stage, they are bound into the same fact candidate fragment.

[0025] After fact candidate segments are formed, source duplication verification is performed on them. Source duplication verification uses the source system identifier, production object, status action, occurrence location, and acquisition time as the verification objects. When two original record segments have the same source system identifier, production object, status action, and occurrence location, and the difference in acquisition time falls within the continuous reporting cycle of the corresponding source system, the later-arriving original record segment is considered a duplicate report segment and merged into the existing fact candidate segments. The continuous reporting cycle is determined by the reporting frequency or data synchronization cycle of the corresponding acquisition device; for example, if the device reports once every 1 second, the continuous reporting cycle can be 1 to 2 seconds. Records of the same type exceeding this range are not directly considered duplicate reports.

[0026] When performing time misalignment checks on fact candidate segments, the sequence of data collection, reporting, and production processing events is compared simultaneously. Original record segments whose collection time matches the production processing event sequence but whose reporting time is later than the adjacent production processing event are marked with a reporting delay flag; original record segments whose collection time does not match the production processing event sequence are marked with a pending verification flag and then merged into the corresponding fact candidate segment.

[0027] After completing source duplication and time misalignment checks, the candidate fact fragments are encapsulated together with source markers, time markers, and state credibility markers to form a production fact fragment. Source markers include the source system identifier, acquisition node identifier, and reporting link identifier; time markers include the acquisition time, reporting time, and relative time position within the production processing event; state credibility markers include duplicate reporting markers, delayed reporting markers, pending verification markers, and source missing markers. The production fact fragment retains both the original record fragments and the bound fact structure, enabling it to both enter the multi-agent verification module for source verification and participate in the construction of the fact chain.

[0028] Finally, production fact pieces are arranged according to the sequence of production processing events to form a set of production fact pieces. The sequence is determined by the flow of production objects, the relationship between processes, the material delivery relationship, the quality inspection relationship, and the warehousing relationship. When multiple production fact pieces correspond to the same production processing event, the order within the pieces is determined by the collection time and the source credibility marker. The resulting set of production fact pieces retains the correspondence between source records, candidate fact pieces, and production fact pieces, serving as the input object for the multi-agent verification module.

[0029] In this embodiment, the multi-agent verification module specifically comprises: Based on the origin of the production fact pieces in the production fact piece set, the production fact pieces are assigned to the production agents in the corresponding production data domain. Each production agent performs source consistency verification on the production fact pieces belonging to its own data domain, checks the reporting link, collection time and status meaning of the source record, and marks the source reliable fact pieces and source abnormal fact pieces; Perform state continuity verification on trusted fact pieces from the source, and mark production fact pieces whose state transitions are inconsistent with the state inheritance table of this data domain as abnormal fact pieces; For trusted fact fragments from the same production processing event, perform aggregation processing to merge production fact fragments with consistent state meanings and adjacent time positions into verified production fact fragments. The fact pieces with abnormal sources and abnormal states are retained as fact pieces to be collaboratively reconstructed, and together with the verified production fact pieces, they form a set of verified production fact pieces.

[0030] In the specific implementation process, after receiving the set of production fact pieces, the multi-agent verification module first determines the source attribution according to the source markers in the production fact pieces. The source system identifier, acquisition node identifier, and reporting link identifier in the source marker jointly determine the data domain to which the production fact piece belongs; production fact pieces generated by the equipment acquisition system are assigned to the equipment production agent, production fact pieces generated by the manufacturing execution system are assigned to the process production agent, production fact pieces generated by the quality inspection system are assigned to the quality production agent, production fact pieces generated by the warehousing system are assigned to the warehousing production agent, and production fact pieces generated by the scheduling system are assigned to the scheduling production agent. When a source marker corresponds to multiple data domains, the primary data domain is determined first according to the business meaning corresponding to the state action, and other source relationships are retained as mutual verification source relationships.

[0031] After receiving a production fact piece from its assigned data domain, each production agent first performs a source consistency verification. This verification checks the source system identifier, acquisition node identifier, reporting link identifier, acquisition time, and status meaning within the production fact piece. If the record numbers within the same source link are consecutive, the acquisition node matches the data domain, the acquisition time falls within the time range of the corresponding production processing event, and the status meaning is consistent with the record type of the source system, the production fact piece is designated as a source-trusted fact piece. If the record numbers are broken, the acquisition node does not match the data domain, the reporting link is inconsistent with the source system, or the status meaning is inconsistent with the record type of the source system, the production fact piece is designated as a source-abnormal fact piece.

[0032] The source of trusted fact pieces is used for state continuity verification. Each production agent configures a state acceptance table for its corresponding data domain. The state acceptance table consists of the allowed state transition relationships for that data domain. For example, the state transition relationships in the equipment data domain can include allowed transition relationships between standby, processing, processing completed, shutdown, and alarm; the state transition relationships in the quality data domain can include allowed transition relationships between pending inspection, inspection in progress, qualified, unqualified, and rework. This data domain's state acceptance table is generated during the production agent's initialization phase. The corresponding production agent reads the process route table, equipment status dictionary, quality inspection process table, warehouse flow rule table, and historical trusted production fact pieces for this data domain, and extracts the sequential relationship of adjacent state actions from the above data. State transition relationships that have been manually confirmed, closed-loop verified, or have historically stable occurrences are registered as allowed acceptance relationships, while state combinations with reverse jumps, cross-stage jumps, and missing intermediate processing states are registered as abnormal acceptance relationships. When the production process, equipment program, quality inspection process, or warehouse flow rule is adjusted, the corresponding production agent updates the data domain's state acceptance table incrementally according to the updated configuration.

[0033] Each production agent arranges the state actions in the production fact slices according to the time of collection or the order of production processing events, and compares them item by item with the state inheritance table. When adjacent state actions have an allowed inheritance relationship in the state inheritance table, the trusted state of the source trusted fact slice is retained. When adjacent state actions lack an allowed inheritance relationship, or when there is a broken relationship such as jumping directly from the completed state back to the start state, or jumping directly from the un-feeded state to the completed processing state, the corresponding production fact slice is marked as a state abnormal fact slice.

[0034] For trusted fact pieces from the same production processing event, each production agent continues to perform state convergence. State convergence is based on the following conditions: consistent production object, consistent state meaning, consistent or adjacent occurrence location, and adjacent time location. The range of adjacent time location is determined by the sampling period, synchronization period, or business processing cycle of the corresponding data domain; for example, when equipment data is uploaded once every 1 second, the adjacent time range can be 1 to 2 seconds; when the manufacturing execution system synchronizes by process node, the adjacent time range can be the closed time period of the corresponding process node's record. Multiple production fact pieces that meet the convergence conditions are merged into one verified production fact piece. During merging, the source tag and time tag of each source fact piece are retained, and the trusted state tag is updated according to the trusted source state.

[0035] For fact pieces with abnormal sources and abnormal states, the multi-agent verification module does not delete them. Fact pieces with abnormal sources retain their original source marker, abnormal source type, and corresponding production agent affiliation; fact pieces with abnormal states retain their state actions, abnormal succession positions, and inconsistent state succession relationships. After being configured with a "to be collaboratively reconstructed" marker, these abnormal fact pieces, together with the verified production fact pieces, form a set of verified production fact pieces. This set of verified production fact pieces simultaneously retains trusted fact pieces and fact pieces to be collaboratively reconstructed, enabling it to enter the fact chain construction module to participate in the production succession relationship concatenation.

[0036] In this embodiment, the fact chain construction module is specifically as follows: From the verified set of production fact pieces, production fact pieces with production status succession relationships are selected as candidate fact pieces for succession. Perform acceptance verification on the order of state occurrence, object flow, and return confirmation in the candidate fact pieces, and mark the production acceptance front position, production acceptance position, and production acceptance return confirmation position; The completed production fact footage is linked together into production acceptance segments according to the production acceptance front position, production acceptance itself position, and production acceptance back verification position. Perform a check on the object continuity state and state closure result between adjacent production acceptance segments, and mark the production acceptance segments with broken acceptance as segments to be reconstructed; The production acceptance segments are connected in sequence according to their arrangement. Acceptance pointers are configured between adjacent production acceptance segments, and the original position of the segment to be reconstructed in the chain is preserved to form a production acceptance fact chain.

[0037] In the specific implementation process, after receiving the verified set of production fact pieces, the fact-receiving module first selects candidate fact pieces for acceptance from the production fact pieces. The selection criteria are production object markers, status actions, occurrence locations, collection times, and status credibility markers. Production fact pieces with consistent production objects, sequential status actions, and occurrence locations belonging to the same production flow path are included in the candidate acceptance range. Fact pieces with abnormal sources and abnormal states are not directly eliminated; their original chain position candidate relationships are retained, and they participate in the acceptance verification with a marker indicating pending collaborative reconstruction.

[0038] When accepting candidate fact pieces for acceptance verification, the status occurrence sequence is first organized according to the collection time, the relative time position of the event, and the production processing sequence. For the same production object, the status actions are arranged according to the status relationship such as "start processing, processing, processing completed, transfer to the next stage, and return verification confirmation"; when there is a reverse jump between two status actions, a lack of intermediate states, or an inverted occurrence time, the corresponding fact piece is marked as a suspected acceptance break. The verification result of the status occurrence sequence is used to determine the production acceptance front position and the production acceptance position.

[0039] The object flow sequence verification uses production object markers, material batch numbers, process positions, equipment positions, and storage positions as comparison objects. If the production object in the current production fact piece is consistent with the production object in the previous production fact piece, or if a corresponding relationship can be established between the material batch number, semi-finished product number, and process transfer record, an object continuity relationship is established between the two production fact pieces. If the object marker changes and transfer records, split records, or batch merging records are missing, the corresponding receiving position is marked as an object continuity anomaly. The object flow sequence verification result is used to confirm the receiving edge between the previous receiving position and the current receiving position.

[0040] The verification and confirmation sequence is based on quality inspection records, warehousing records, rework records, re-inspection records, and scheduling confirmation records. For production fact pieces that have been processed or transferred, the fact chain construction module searches for the corresponding verification fact piece in adjacent chain positions or within the same production processing event range. When the production object, status action, and occurrence location of the verification fact piece can correspond to the current production fact piece, the verification fact piece is marked as a production acceptance verification position. If a verification fact piece is missing, the verification result conflicts with the current status, or the verification time is earlier than the occurrence time of the current status, the corresponding production acceptance segment is marked as pending reconstruction.

[0041] After verifying the order of state occurrence, object flow, and verification confirmation, the fact chain construction module concatenates the preceding production acceptance, the current production acceptance, and the verification production acceptance into a production acceptance segment. A production acceptance segment can contain a preceding fact piece, a current fact piece, a verification fact piece, and corresponding source and state credibility markers. For production acceptance segments with only preceding and current fact pieces and no verification fact piece yet, a verification slot is reserved and a "to be reconstructed" marker is written. For cases with multiple verification fact pieces, the primary verification fact piece is determined based on the source credibility marker and acquisition time, while other verification fact pieces are retained as mutually verified fact pieces.

[0042] Between adjacent production segments, the object continuation status and state closure result verification continue. The object continuation status verification compares the output production object of the previous production segment with the input production object of the next production segment; the state closure result verification compares the verification result of the previous production segment with the initial state of the next production segment. When the previous segment shows processing completed and verification passed, and the next segment shows the same production object has entered the next process, a continuous connection is established between the two production segments; when the previous segment lacks a verification result, the production object in the next segment is inconsistent, or the state action cannot be inherited, the next production segment is marked as a connection break segment.

[0043] The broken segment is not deleted from the chain. The accepting fact chain construction module concatenates all production accepting segments according to their arrangement order, and configures accepting pointers between adjacent production accepting segments. The accepting pointers record the previous segment's chain position, the current segment's chain position, the backtracking segment's chain position, and the break mark, so that the segment to be reconstructed retains its original position in the production accepting fact chain. The formed production accepting fact chain contains both normal accepting segments and segments to be reconstructed, serving as input objects for the improved SAITS collaborative reconstructing module.

[0044] In this embodiment, the improved SAITS network includes an input embedding structure, a first self-attention interpolation block, a production fact reconstruction layer, a second self-attention interpolation block, and an interpolation output structure connected in sequence. Trainable network parameters are configured in the input embedding structure, the first self-attention interpolation block, the production fact reconstruction layer, the second self-attention interpolation block, and the interpolation output structure. The production fact reconstruction layer is embedded between the output end of the first self-attention interpolation block and the input end of the second self-attention interpolation block. Both the first and second self-attention interpolation blocks retain the self-attention interpolation block structure in the SAITS network.

[0045] In this embodiment, the network parameters can be trained, and the specific training process is as follows: Complete production fact chains, missing production fact chains, and conflicting production fact chains are compiled from historical enterprise production data, and these production fact chains are converted into training fact chain samples. Configure the chain bit markers to be reconstructed for the training fact chain samples. The chain bit markers to be reconstructed cover the randomly masked chain bits, the true missing chain bits, the conflicting fact piece chain bits, and the chain bits with low confidence sources. Configure the production fact sheet before masking, the production fact sheet after manual verification, and the original credible production fact sheet as the reconstruction supervision sheet, the correction supervision sheet, and the maintenance supervision sheet, respectively. According to the training batch, the training fact chain samples are input into the improved SAITS network, and then pass through the input embedding structure, the first self-attention interpolation block, the production receiving fact reconstruction layer, the second self-attention interpolation block and the interpolation output structure in sequence to obtain the training reconstructed fact pieces. The training reconstruction fact piece and the reconstruction supervision piece are aligned at random masking chain positions to form the interpolation reconstruction error. The training reconstructed fact pieces and the corrected supervision pieces are aligned at the true missing chain positions, conflicting fact piece chain positions, and low-confidence source chain positions to form a collaborative error correction. The training reconstructed fact piece and the holding supervision piece are aligned at non-reconstructed chain bits to form the original holding error; The interpolation reconstruction error, co-correction error, and original preservation error are combined into the total training error, and the trainable network parameters are updated in reverse based on the total training error. When the total training error corresponding to the verification fact chain sample is less than the training error threshold, or when the number of training rounds reaches the upper limit of the number of training rounds, the trainable network parameters are fixed.

[0046] In the specific implementation process, trainable network parameters include: fact piece content embedding parameters, chain bit embedding parameters, source credibility embedding parameters, and to-be-reconstructed tag embedding parameters in the input embedding structure; query projection parameters, key projection parameters, value projection parameters, attention output parameters, and normalization parameters in the first and second self-attention interpolation blocks; peer verification linear projection parameters, previous-position verification linear projection parameters, subsequent-position backtesting linear projection parameters, chain bit offset representation parameters, gate mapping parameters, forward verification projection parameters, and backward backtesting projection parameters in the production receiving fact reconstruction layer; and fact piece restoration parameters in the interpolation output structure. These parameters collectively participate in forward computation and error backpropagation during training.

[0047] Training samples were compiled from historical enterprise production data. First, complete production fact chains, missing production fact chains, and conflicting production fact chains were extracted from the historical data, and each production fact chain was converted into a training fact chain sample. Complete production fact chains came from production records that had been closed-loop confirmed; missing production fact chains came from production records with missing fact pieces, acquisition breakpoints, or synchronization breakpoints; conflicting production fact chains came from production records where different production agents gave inconsistent states for the same production processing event. Each training fact chain sample retained the fact piece chain position, the inheritance pointer, the source credibility state, and the chain position marker to be reconstructed.

[0048] The reconstructed link markers in the training fact chain samples cover four types of links. Randomly masked links are extracted from the complete production fact chain and used to train the model to recover masked production fact pieces; true missing links come from positions in the missing production fact chain where source records are missing; conflicting fact piece links come from positions where different source records have inconsistent states; and low-confidence source links come from positions marked with source anomalies, time misalignments, or state jump anomalies. The random masking ratio can be set to 10% to 30% of the total number of production fact pieces in the training fact chain samples, while all low-confidence source links and conflicting fact piece links retain the reconstructed link markers.

[0049] Supervision is configured according to chain segment type. Randomly masked chain segments correspond to the unmasked production fact segments, which are used as reconstruction supervision segments. Truly missing chain segments, conflicting fact segment segments, and low-confidence source chain segments correspond to manually verified production fact segments, which are used as correction supervision segments. Non-reconstruction chain segments correspond to the original credible production fact segments, which are used as maintenance supervision segments. Reconstruction supervision segments are used to train the model to restore normal production fact segments, correction supervision segments are used to train the model to correct missing, conflicting, and low-confidence source fact segments, and maintenance supervision segments are used to restrict the model from rewriting the original credible production fact segments.

[0050] During training, training fact chain samples are read in training batches, with batch size set to 16 or 32 production-receiving fact chains. Each training fact chain sample first enters the input embedding structure, which converts the production fact pieces into fact piece content embeddings and overlays chain position embeddings, source credibility embeddings, and reconstructed marker embeddings to form a fact piece hidden state sequence. The fact piece hidden state sequence enters the first self-attention imputation block, which obtains the preliminary hidden state through query projection, key projection, and value projection. The preliminary hidden state enters the production-receiving fact reconstruction layer, which performs chain position back-rowing, anchor position reconstruction, and gate value writing according to the fact piece chain position index table to form the production-receiving reconstruction hidden state. The production-receiving reconstruction hidden state enters the second self-attention imputation block, which forms a collaborative imputation hidden state. The imputation output structure restores the collaborative imputation hidden state to the training reconstructed fact piece.

[0051] Training errors are calculated separately according to chain position type. For randomly occluded chain positions, the training reconstructed fact pieces and the reconstructed supervised pieces are aligned by chain position, and the state-action, source credible state, and fact piece content embeddings are compared separately. Among them, the state-action difference is determined by whether the state category is consistent, the source credible state difference is determined by whether the credible, abnormal, unverified, and low-credible categories are consistent, and the fact piece content embedding difference is determined by the sum of the dimension-wise differences of the two embedding vectors. The three types of differences are combined into the interpolation reconstruction error.

[0052] For genuine missing chain bits, conflicting fact piece chain bits, and low-trust source chain bits, the training reconstructed fact piece and the correction supervision piece are aligned by chain bit alignment. The state actions, chain bit positions, source trust status, and fact piece content embedding after manual verification are compared respectively. When the state actions, chain bit positions, and source trust status of the training reconstructed fact piece are consistent with those of the manually verified produced fact piece, the corresponding chain bit's collaborative correction error decreases. When the training reconstructed fact piece still retains conflicting states, misaligned chain bits, or low-trust source states, the corresponding chain bit's collaborative correction error increases.

[0053] For non-reconstruction chain bits, the training reconstructed fact pieces and the maintenance supervision pieces are aligned, and the changes in the state actions, source trusted states, and fact piece content embeddings of the trusted production fact pieces before and after reconstruction are compared. The greater the change in the non-reconstruction chain bits, the greater the original maintenance error. The original maintenance error is used to limit the over-rewriting of the original trusted production fact pieces by the improved SAITS network.

[0054] The interpolation reconstruction error, collaborative correction error, and original preservation error are combined into the total training error according to their respective error weights. The error weights can be determined based on the verification fact chain samples. In one implementation, the interpolation reconstruction error weight is 1.0, the collaborative correction error weight is 1.5, and the original preservation error weight is 0.5, making the training process more focused on correcting conflicting fact pieces, true missing chain positions, and low-confidence source chain positions, while retaining the constraint of preserving the original confident chain positions. The total training error is propagated backward from the interpolation output structure, sequentially updating the trainable network parameters in the interpolation output structure, the second self-attention interpolation block, the production receiving fact reconstruction layer, the first self-attention interpolation block, and the input embedding structure. Mini-batch gradient descent or Adam optimization is used to update the parameters during training. The training error threshold is determined based on the reconstruction accuracy requirements of the verification fact chain samples; a condition for solidification can be that the total verification error is below a set threshold for several consecutive rounds. In one implementation, the training error threshold is between 0.05 and 0.10, and the upper limit of the training rounds is between 80 and 120 rounds.

[0055] In this embodiment, the SAITS collaborative reconfiguration module is improved as follows: The production fact pieces in the production fact chain are arranged according to the fact piece chain position and mapped to the fact piece content embedding; The factual content embedding, chain position embedding, source credibility embedding, and reconstructed tag embedding are superimposed in the same position to form the factual hidden state sequence; The first self-attention interpolation block performs query projection, key projection, and value projection on the sequence of hidden states of the fact slice, and weights and aggregates the value hidden states according to the chain position matching relationship between the query hidden states and the key hidden states to form the initial hidden states. The production acceptance fact reconstruction layer generates a fact piece chain position index table based on the production acceptance fact chain, and performs chain position back-sorting on the initial hidden state according to the fact piece chain position index table. It then performs anchor position acceptance reconstruction around the reconstruction anchor position corresponding to the fact piece to be reconstructed, forming the production acceptance reconstruction hidden state. The second self-attention interpolation block performs a second-stage self-attention interpolation on the hidden state of production acceptance and reconstruction, forming a collaborative interpolation hidden state. The interpolation output structure restores the cooperative interpolation hidden state to the reconstructed fact piece.

[0056] In this embodiment, anchor point reconstructing is performed around the reconstructing anchor point corresponding to the fact piece to be reconstructed, forming a production reconstructing hidden state, specifically as follows: Using the chain position corresponding to the fact piece to be reconstructed as the reconstruction anchor position, extract the corresponding mutual verification hidden state, the previous position inheritance hidden state, and the subsequent position verification hidden state from the initial hidden state after the chain position is back-rowed. The hidden state of mutual verification comes from the production fact piece that is in the same production acceptance position as the reconstruction anchor. The hidden state of the previous acceptance comes from the production fact piece that the acceptance pointer points to the reconstruction anchor. The hidden state of the subsequent verification comes from the production fact piece that the reconstruction anchor points to and has a verification relationship. After linear projection, the hidden states of mutual verification, previous inheritance, and subsequent verification maintain the same hidden state dimension as the hidden state of the reconstructed anchor position. The corresponding mutual verification hidden state and the reconstructed anchor hidden state are multiplied dimension by dimension to obtain the mutual verification consistency component, and then subtracted dimension by dimension to obtain the mutual verification deviation component. The mutual verification consistency component and the mutual verification deviation component are combined in the same position to form the corresponding mutual verification code. The previous hidden state and the reconstructed anchor hidden state are subtracted dimension by dimension, and the chain offset representation of the corresponding accepting pointer is superimposed to form the previous accepting code; The hidden state of the post-verification is subtracted from the hidden state of the reconstructed anchor position in one dimension, and the chain offset representation of the corresponding verification relationship is superimposed to form the post-verification code; The source trust representations corresponding to the same position mutual verification code, the previous position acceptance code, the subsequent position verification code, and the reconstructed anchor are concatenated and then mapped by a threshold value to form the anchor retention threshold, the previous position write threshold, and the subsequent position verification threshold. The preceding latent state is projected forward to form a forward candidate state, and the following latent state is projected backward to form a backward candidate state. The anchor position retention threshold is multiplied dimension-wise with the reconstructed anchor position hidden state. The previous write threshold is multiplied dimension-wise with the forward acceptance candidate state. The subsequent verification threshold is multiplied dimension-wise with the reverse verification candidate state. The three products are added in the same position and then written back to the reconstructed anchor position to form the production acceptance reconstructed hidden state.

[0057] In its implementation, the improved SAITS collaborative reconstruction module performs collaborative reconstruction of the fact pieces to be reconstructed in the production acceptance fact chain. The original SAITS network typically included an input embedding structure, a first self-attention interpolation block, a second self-attention interpolation block, and an interpolation output structure. The intermediate hidden states output by the first self-attention interpolation block directly entered the second self-attention interpolation block. This structure primarily relied on temporal position and variable correlation to complete interpolation, lacking explicit processing at the intermediate hidden state level for fact piece chain positions, acceptance pointers, peer verification relationships, and backtesting relationships in enterprise production data. This easily led to production records from the same time period but with weak acceptance relationships participating in the interpolation calculation. The improved SAITS network retains the first and second self-attention interpolation blocks, embedding a production acceptance fact reconstruction layer between them. This allows the initial hidden states formed in the first stage to be rewritten through production acceptance relationships before entering the second stage of self-attention interpolation.

[0058] After the production fact chain enters the improved SAITS network, the production fact pieces are first arranged according to their chain positions. Each production fact piece is mapped to a fact piece content embedding, which can be obtained by vectorizing the production object, state action, source trusted state, reconstructed marker, and original fact piece content. The fact piece content embedding, chain position embedding, source trusted embedding, and reconstructed marker embedding are superimposed at the same chain position to form a hidden state sequence of fact pieces. The chain position embedding records the position of the production fact piece in the production fact chain, the source trusted embedding records whether the production fact piece comes from a trusted source, an abnormal source, or a source to be verified, and the reconstructed marker embedding distinguishes between normal chain positions, missing chain positions, conflicting chain positions, and low-trust source chain positions.

[0059] The first self-attention interpolation block performs query projection, key projection, and value projection on the fact piece hidden state sequence. The fact piece hidden state at each chain position is mapped to a query hidden state, a key hidden state, and a value hidden state, respectively. The query hidden state and the key hidden state form an attention response according to the chain position matching relationship, and the value hidden state is weighted and aggregated according to the attention response. The weighted aggregated hidden state is then residually overlaid and normalized with the fact piece hidden state at the corresponding chain position to form the preliminary hidden state. The preliminary hidden state retains the time-related information and inter-fact piece information obtained from the first stage of self-attention interpolation, and then enters the production-reconstruction-of-facts layer.

[0060] The production acceptance fact reconstruction layer is located between the output of the first self-attention interpolation block and the input of the second self-attention interpolation block. Internally, the production acceptance fact reconstruction layer includes a chain-bit backordering structure, an anchor context extraction structure, a mutual verification difference encoding branch, an acceptance verification candidate projection branch, and a gate value writing structure. The chain-bit backordering structure generates a fact piece chain-bit index table based on the production acceptance fact chain. This table records the fact piece chain bit, acceptance pointer, verification relationship, and peer verification relationship. Following the fact piece chain-bit index table, the chain-bit backordering structure backorders the initial hidden states output by the first self-attention interpolation block chain-bit by chain, ensuring that each initial hidden state corresponds to the position of the relevant fact piece in the production acceptance fact chain.

[0061] The anchor context extraction structure uses the chain position corresponding to the fact piece to be reconstructed as the reconstruction anchor. It extracts the peer-verification hidden state, the preceding acceptance hidden state, and the subsequent verification hidden state from the initial hidden state after chain position back-sorting. The peer-verification hidden state comes from the production fact piece that is in the same production acceptance position as the reconstruction anchor; the preceding acceptance hidden state comes from the production fact piece whose acceptance pointer points to the reconstruction anchor; and the subsequent verification hidden state comes from the production fact piece pointed to by the reconstruction anchor and has a verification relationship. After extraction, the three types of hidden states are mapped to the same hidden state dimension as the reconstruction anchor hidden state through a fully connected linear projection layer. The fully connected linear projection layer weights and accumulates the components of each dimension in the hidden state vector and adds bias parameters. The projection weights and bias parameters are used as trainable network parameters for training and updating the improved SAITS network.

[0062] The mutual verification difference coding branch handles the relationship between the latent state of the same-position mutual verification and the latent state of the reconstructed anchor position. The latent state of the same-position mutual verification and the latent state of the reconstructed anchor position are multiplied dimension-wise to obtain the mutual verification consistency component; the latent state of the same-position mutual verification and the latent state of the reconstructed anchor position are subtracted dimension-wise to obtain the mutual verification deviation component. The mutual verification consistency component retains the identical responses among fact pieces from multiple sources within the same production acceptance position, while the mutual verification deviation component retains the differential responses between corresponding sources. The mutual verification consistency component and the mutual verification deviation component are combined according to their corresponding dimensional positions to form the same-position mutual verification code.

[0063] The candidate projection branch for acceptance and verification processes the preceding acceptance hidden state and the subsequent acceptance hidden state respectively. The preceding acceptance hidden state is subtracted dimension-wise from the reconstructed anchor hidden state, and the chain offset representation corresponding to the acceptance pointer is superimposed to form the preceding acceptance code. The chain offset representation is determined by the chain distance between the preceding fact piece and the reconstructed anchor, the acceptance direction, and the acceptance type, so that the preceding acceptance code carries both state differences and acceptance position differences. The preceding acceptance hidden state is then subjected to forward acceptance projection to form a forward acceptance candidate state. The subsequent acceptance hidden state is subtracted dimension-wise from the reconstructed anchor hidden state, and the chain offset representation corresponding to the verification relationship is superimposed to form the subsequent verification code. The subsequent verification hidden state then enters the reverse verification projection path. The reverse backtesting projection path first concatenates the subsequent backtesting hidden state with the subsequent backtesting encoding and the chain offset representation corresponding to the backtesting relationship, so that the input vector simultaneously contains the state content of the subsequent backtesting fact piece, the state difference between the subsequent backtesting fact piece and the reconstruction anchor, and the chain distance in the backtesting direction. Subsequently, the reverse backtesting projection path performs weighted accumulation and bias superposition on the concatenated input vector through an independent fully connected linear projection layer, mapping the subsequent backtesting hidden state from the backtesting chain to the hidden state space corresponding to the reconstruction anchor, and then normalizing it to form a reverse backtesting candidate state. This reverse backtesting candidate state is used to represent the candidate hidden state when the subsequent backtesting result is used to correct the fact piece to be reconstructed.

[0064] The threshold write structure receives the peer verification code, the preceding acceptance code, the following verification code, and the source trust representation corresponding to the reconstructed anchor. These representations are concatenated at the same chain position and then entered into the threshold mapping to obtain the anchor retention threshold, the preceding write threshold, and the following verification threshold. The anchor retention threshold is multiplied dimension-wise with the reconstructed anchor hidden state to obtain the anchor retention state; the preceding write threshold is multiplied dimension-wise with the forward acceptance candidate state to obtain the preceding write state; and the following verification threshold is multiplied dimension-wise with the reverse verification candidate state to obtain the following verification write state. The anchor retention state, the preceding write state, and the following verification write state are added along the same hidden state dimension to obtain the updated reconstructed anchor hidden state. The updated reconstructed anchor hidden state is written back to the reconstructed anchor position in the chained hidden state, while the remaining chain hidden states maintain their original chain position arrangement, forming the production acceptance reconstructed hidden state.

[0065] The production process accepts the reconstructed hidden state and enters the second self-attention interpolation block. The second self-attention interpolation block performs a second stage of self-attention interpolation based on the hidden state that has already undergone chain bit back-rowing, peer verification encoding, previous bit acceptance writing, and subsequent bit verification writing, forming a collaboratively interpolated hidden state. The interpolation output structure restores the collaboratively interpolated hidden state to a reconstructed fact piece. The reconstructed fact piece retains the corresponding chain bit, source trusted state, and the marker to be reconstructed, and can enter the acceptance closed-loop solidification module for back-attachment verification.

[0066] Through the above improvements, the hidden state transmission path between the first self-attention imputation block and the second self-attention imputation block is changed. In the original SAITS network, the intermediate hidden state obtained from the first-stage imputation directly enters the second-stage imputation; in the improved SAITS network, the initial hidden state obtained from the first-stage imputation is first rewritten in-chain by the production accepting fact reconstruction layer according to the fact piece chain position, accepting pointer, peer verification relationship, and verification relationship, and then enters the second-stage imputation. This processing ensures that the hidden state of the fact piece to be reconstructed is simultaneously affected by the peer fact piece, the preceding accepting fact piece, and the following verification fact piece, reducing the interference of adjacent data without acceptance relationships on the imputation result.

[0067] The beneficial effects of this structure are that the production acceptance fact reconstruction layer transforms source conflicts, acceptance breaks, and backtesting misalignments into peer verification codes, preceding acceptance codes, and subsequent backtesting codes, so that the model's reconstruction of fact pieces no longer relies solely on ordinary temporal adjacency. The anchor retention threshold, preceding write threshold, and subsequent backtesting threshold respectively adjust the write ratios of the initial hidden state, the forward acceptance candidate state, and the reverse backtesting candidate state, making the reconstructed fact pieces more consistent with the acceptance relationships in the production acceptance fact chain. Therefore, the improved SAITS collaborative reconstruction module can enhance the reconstruction accuracy of missing fact pieces, conflicting fact pieces, and fact pieces with low-reliability sources, and provide the acceptance closed-loop solidification module with reconstructed fact pieces that have clear chain positions and traceable sources.

[0068] In this embodiment, the closed-loop curing module is specifically as follows: Establish a collaborative shadow fact chain based on the fact piece chain position and acceptance pointer in the production acceptance fact chain. The collaborative shadow fact chain retains the same chain position arrangement and acceptance pointer as the production acceptance fact chain. The reconstructed fact pieces are pasted back to the production receiving fact chain according to the corresponding chain position, and the preceding production fact pieces, following production fact pieces, and corresponding mutual verification fact pieces adjacent to the reconstructed fact pieces are located. Perform preceding succession verification on the object continuation state between the reconstructed fact piece and the preceding production fact piece; perform subsequent succession verification on the state succession result between the reconstructed fact piece and the subsequent production fact piece; and perform peer closure verification on the source mutual verification state between the reconstructed fact piece and the peer mutual verification fact piece. When the previous position acceptance verification, the subsequent position acceptance verification, and the same position closure verification all pass, the reconstructed fact piece is written into the corresponding chain position in the collaborative shadow fact chain, and the chain position pointer and source pointer of the original fact piece before reconstruction are retained in the corresponding chain position; When the verification of the previous position, the verification of the subsequent position, or the verification of the same position fails, the original fact piece before reconstruction is retained in the corresponding chain position in the collaborative shadow fact chain. The reconstructed fact piece is marked as the disputed fact piece, and the chain position, source agent, and verification type of the disputed fact piece are recorded. The disputed factual pieces are attached to the corresponding production intelligent agents, and the original factual pieces before reconstruction are retained in the production factual chain.

[0069] In the specific implementation process, the closed-loop solidification module first establishes a collaborative shadow fact chain based on the fact piece chain position and acceptance pointer in the production acceptance fact chain. The collaborative shadow fact chain maintains the same chain position arrangement and acceptance pointer as the production acceptance fact chain. Each shadow chain position first retains the chain position pointer and source pointer of the corresponding original production fact piece. The reconstructed fact piece does not directly overwrite the original production fact piece before verification.

[0070] After the reconstructed fact piece is output, the closed-loop solidification module pastes the reconstructed fact piece back into the production receiving fact chain according to the chain position carried by the reconstructed fact piece, and locates the preceding production fact piece, following production fact piece, and peer verification fact piece adjacent to the corresponding chain position. The preceding receiving verification is used to check whether the output object, completion status and input status of the preceding production fact piece are continuous with the reconstructed fact piece; the following receiving verification is used to check whether the output status of the reconstructed fact piece can be received by the following production fact piece; the peer closure verification is used to check whether the status meaning of the reconstructed fact piece is consistent with that of other source fact pieces under the same production receiving position.

[0071] When the current position acceptance verification, the subsequent position acceptance verification, and the same position closure verification all pass, the reconstructed fact piece is written to the corresponding link in the collaborative shadow fact chain, and the link pointer and source pointer of the original production fact piece before reconstruction are retained in the corresponding link. The original production fact piece in the production acceptance fact chain remains unchanged, and the reconstructed fact piece in the collaborative shadow fact chain serves as the collaborative correction result after the verification passes.

[0072] When the verification of the current position, the verification of the subsequent position, or the verification of the same position fails, the corresponding link in the collaborative shadow fact chain continues to retain the original production fact piece before reconstruction. The acceptance closed-loop solidification module marks the reconstructed fact piece as a disputed fact piece, records the link corresponding to the disputed fact piece, the source agent, and the type of verification that failed, and attaches the disputed fact piece to the corresponding production agent.

[0073] In this embodiment, the collaborative result output module is specifically as follows: Align the production acceptance fact chain and the collaborative shadow fact chain item by item according to the fact piece chain position, and distinguish between the original fact pieces that have not been reconstructed and the verified fact pieces written into the collaborative shadow fact chain; Configure collaborative correction tags for the verified fact pieces, and register the collaborative correction tags together with the source location, chain location, and production agent affiliation of the corresponding original fact pieces; For disputed fact pieces that are not written into the collaborative shadow fact chain, retain the original fact pieces in the production successor fact chain, and register the unverified type corresponding to the disputed fact piece as a dispute prompt mark; The enterprise production data is processed collaboratively by combining the original factual footage, verified factual footage, and dispute alerts. The collaborative processing results of enterprise production data are output to the enterprise production data management terminal, and collaborative correction markers, dispute reminder markers, and the corresponding production agent affiliation relationships are output simultaneously.

[0074] In the specific implementation process, after receiving the disputed fact fragment records formed by the production acceptance fact chain, the collaborative shadow fact chain, and the acceptance closed-loop solidification module, the collaborative result output module first establishes an alignment relationship according to the fact fragment chain position. During alignment, the chain position number and acceptance pointer in the production acceptance fact chain are used as the benchmark, and the content of the same chain position in the collaborative shadow fact chain is read item by item; when the corresponding chain position in the collaborative shadow fact chain is written with a verified fact fragment, the verified fact fragment is used as the collaborative correction result; when the corresponding chain position in the collaborative shadow fact chain still retains the original fact fragment, the original fact fragment in the production acceptance fact chain is used as the retention result of the current chain position.

[0075] For verified fact pieces written into the collaborative shadow fact chain, the collaborative result output module configures a collaborative correction flag. The collaborative correction flag records at least the corrected chain position, the original fact piece source location before correction, the corrected verified fact piece source location, the corresponding production agent affiliation, and the closed-loop verification result. Fact pieces before and after correction are retained simultaneously, allowing the production data management end to view why a particular chain position was corrected, and which type of production agent and which closed-loop verification result the correction result came from.

[0076] For disputed fact pieces not written into the collaborative shadow fact chain, the collaborative result output module does not overwrite the original fact piece with the disputed fact piece. Instead, it retains the original fact piece from the production acceptance fact chain at the corresponding chain position and registers a dispute alert flag. The dispute alert flag records the chain position of the disputed fact piece, the source agent, the type of verification failure, and the corresponding preceding acceptance verification, subsequent acceptance verification, or same-position closure verification result. After receiving the dispute alert flag, the production data management terminal can directly locate the production fact piece position that requires manual review or further collaborative reconstruction.

[0077] After completing the above processing, the collaborative result output module binds the collaborative processing results of enterprise production data according to the chain position of the fact pieces. In the collaborative processing results of enterprise production data, the original fact pieces that have not been reconstructed retain their original chain positions, the verified fact pieces are written to the corresponding chain positions in the form of collaborative correction tags, and the disputed fact pieces are attached to the side of the corresponding chain positions in the form of dispute prompt tags. The resulting result retains both the original record of the production receiving fact chain and the verified correction results in the collaborative shadow fact chain.

[0078] Finally, the collaborative results output module outputs the collaborative processing results of enterprise production data to the enterprise production data management terminal, and simultaneously outputs collaborative correction markers, dispute alert markers, and the corresponding production agent affiliation relationships. The production data management terminal displays the original fact pieces, verified fact pieces, and dispute alert markers in chain position order, enabling production traceability, quality analysis, scheduling verification, and anomaly review to be completed based on the same production undertaking relationship.

[0079] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous production line for surface mount technology (SMT), soldering, inspection, and warehousing in an electronic component assembly company. This production line is connected to an equipment acquisition system, manufacturing execution system, quality inspection system, warehousing system, and scheduling system, and includes 12 SMT machines, 6 soldering machines, 4 AOI inspection machines, and 2 automated warehousing stations. The experiment used 30 consecutive days of production data, involving 428 work orders, 3160 production batches, and approximately 1.86 million source records. Problems included duplicate reporting, delayed synchronization, misaligned quality inspection records, and inconsistencies between equipment completion status and process status.

[0080] During implementation, the production fact piece binding module first identifies source boundaries and binds candidate fact pieces from different sources, resulting in approximately 427,000 production fact pieces. The multi-agent verification module configures equipment production agents, process production agents, quality production agents, warehousing production agents, and scheduling production agents to perform source consistency verification and status aggregation on the production fact pieces, resulting in approximately 314,000 verified production fact pieces, with approximately 28,000 fact pieces reserved for collaborative reconstruction. The fact chain construction module connects the production fact pieces according to production acceptance relationships, forming production acceptance fact chains corresponding to work orders and batches. The improved SAITS collaborative reconstruction module reconstructs missing, conflicting, and low-trust source chain positions in the production acceptance fact chain. The acceptance closed-loop solidification module writes the reconstructed fact pieces that pass verification into the collaborative shadow fact chain, while reconstructed fact pieces that fail verification are linked back to their corresponding production agents.

[0081] To verify the effectiveness, the system of this invention was compared with traditional rule verification methods and ordinary SAITS imputation methods. Traditional rule verification methods use fixed field matching and source priority coverage; ordinary SAITS imputation methods directly convert multi-source production records into a time series matrix and imput missing values; the system of this invention uses production fact piece binding, multi-agent verification, production-inherited fact chains, and an improved SAITS network for collaborative reconstruction. The test results are shown in the table below.

[0082] Table 1. Comparison of Collaborative Processing Methods for Production Data in Different Enterprises

[0083] As shown in Table 1, traditional rule-based verification methods can handle production records with complete fields and stable sources, but they mainly rely on fixed priority coverage for multi-source conflict records, resulting in low completeness rates for conflict fact piece identification and traceability record completion. Ordinary SAITS imputation methods can improve the completion effect of some missing data, but they directly process time series matrices without explicitly utilizing production succession relationships, source mutual verification relationships, and backtesting relationships, thus limiting the improvement in the completeness rate of conflict fact piece identification. The system of this invention, through production fact piece binding, multi-agent verification, production succession fact chain construction, and an improved SAITS network embedded in the production succession fact reconstruction layer, ensures that the fact piece to be reconstructed is simultaneously affected by the latent states of peer mutual verification, previous succession, and subsequent backtesting, and retains the verification pass results through a collaborative shadow fact chain. Therefore, the data consistency accuracy, conflict fact piece identification completeness rate, and traceability record completeness rate are all significantly improved, while the amount of manual review is reduced.

[0084] 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. A collaborative processing system for enterprise production data based on a multi-agent system, characterized in that, include: The production fact sheet binding module is used to respond to enterprise production data collaborative processing tasks, acquire enterprise production data, and bind the source records in the enterprise production data into a set of production fact sheets; The multi-agent verification module is used to configure corresponding production agents for different production data domains. Each production agent performs source verification and status aggregation on the production fact pieces belonging to its own data domain in the production fact piece set, forming a verified production fact piece set. The fact chain construction module is used to connect the production fact pieces in the verified production fact piece set according to the production acceptance relationship to form a production acceptance fact chain; An improved SAITS collaborative reconfiguration module is used to input the production acceptance fact chain into the improved SAITS network. The improved SAITS network embeds a production acceptance fact reconfiguration layer between the first self-attention interpolation block and the second self-attention interpolation block. The production acceptance fact reconfiguration layer rewrites the initial hidden state output by the first self-attention interpolation block, forming a production acceptance reconfiguration hidden state that enters the second self-attention interpolation block, and outputs a reconfiguration fact piece. The closed-loop solidification module is used to paste the reconstructed fact pieces back to the production acceptance fact chain, verify the acceptance closure relationship between the reconstructed fact pieces and the production acceptance fact chain, write the reconstructed fact pieces that have passed the verification into the collaborative shadow fact chain, and attach the reconstructed fact pieces that have not passed the verification to the corresponding production intelligent agent. The collaborative results output module is used to output the collaborative processing results of enterprise production data based on the production acceptance fact chain and the collaborative shadow fact chain.

2. The enterprise production data collaborative processing system based on a multi-agent system according to claim 1, characterized in that, The production fact-copy binding module is specifically as follows: Perform source boundary identification on source records in enterprise production data and retain the original record fragments corresponding to different source records; Identify the production objects, status actions, and locations in the original record segments, and bind the original record segments belonging to the same production processing event into fact candidate segments; Perform source duplication and time misalignment checks on the fact candidate pieces. Merge duplicate reported original record fragments into the corresponding fact candidate pieces. Mark original record fragments with time misalignment as pending verification and then merge them into the corresponding fact candidate pieces. The verified fact candidate piece is encapsulated in the same position as the source tag, time tag, and state trust tag to form a production fact piece; The production fact pieces are arranged in the order of their succession to form a set of production fact pieces.

3. The enterprise production data collaborative processing system based on a multi-agent system according to claim 1, characterized in that, The multi-agent verification module is specifically as follows: Based on the origin of the production fact pieces in the production fact piece set, the production fact pieces are assigned to the production agents in the corresponding production data domain. Each production agent performs source consistency verification on the production fact pieces belonging to its own data domain, checks the reporting link, collection time and status meaning of the source record, and marks the source reliable fact pieces and source abnormal fact pieces; Perform state continuity verification on trusted fact pieces from the source, and mark production fact pieces whose state transitions are inconsistent with the state inheritance table of this data domain as abnormal fact pieces; For trusted fact fragments from the same production processing event, perform aggregation processing to merge production fact fragments with consistent state meanings and adjacent time positions into verified production fact fragments. The fact pieces with abnormal sources and abnormal states are retained as fact pieces to be collaboratively reconstructed, and together with the verified production fact pieces, they form a set of verified production fact pieces.

4. The enterprise production data collaborative processing system based on a multi-agent system according to claim 1, characterized in that, The aforementioned fact chain construction module specifically includes: From the verified set of production fact pieces, production fact pieces with production status succession relationships are selected as candidate fact pieces for succession. Perform acceptance verification on the order of state occurrence, object flow, and return confirmation in the candidate fact pieces, and mark the production acceptance front position, production acceptance position, and production acceptance return confirmation position; The completed production fact footage is linked together into production acceptance segments according to the production acceptance front position, production acceptance itself position, and production acceptance back verification position. Perform a check on the object continuity state and state closure result between adjacent production acceptance segments, and mark the production acceptance segments with broken acceptance as segments to be reconstructed; The production acceptance segments are connected in sequence according to their arrangement. Acceptance pointers are configured between adjacent production acceptance segments, and the original position of the segment to be reconstructed in the chain is preserved to form a production acceptance fact chain.

5. The enterprise production data collaborative processing system based on a multi-agent system according to claim 1, characterized in that, The improved SAITS network includes an input embedding structure, a first self-attention interpolation block, a production fact reconstruction layer, a second self-attention interpolation block, and an interpolation output structure connected in sequence. Trainable network parameters are configured in the input embedding structure, the first self-attention interpolation block, the production fact reconstruction layer, the second self-attention interpolation block, and the interpolation output structure. The production fact reconstruction layer is embedded between the output of the first self-attention interpolation block and the input of the second self-attention interpolation block. Both the first and second self-attention interpolation blocks retain the self-attention interpolation block structure in the SAITS network.

6. The enterprise production data collaborative processing system based on a multi-agent system according to claim 5, characterized in that, The specific training process for the trainable network parameters is as follows: Complete production fact chains, missing production fact chains, and conflicting production fact chains are compiled from historical enterprise production data, and these production fact chains are converted into training fact chain samples. Configure the chain bit markers to be reconstructed for the training fact chain samples. The chain bit markers to be reconstructed cover the randomly masked chain bits, the true missing chain bits, the conflicting fact piece chain bits, and the chain bits with low confidence sources. Configure the production fact sheet before masking, the production fact sheet after manual verification, and the original credible production fact sheet as the reconstruction supervision sheet, the correction supervision sheet, and the maintenance supervision sheet, respectively. According to the training batch, the training fact chain samples are input into the improved SAITS network, and then pass through the input embedding structure, the first self-attention interpolation block, the production receiving fact reconstruction layer, the second self-attention interpolation block and the interpolation output structure in sequence to obtain the training reconstructed fact pieces. The training reconstruction fact piece and the reconstruction supervision piece are aligned at random masking chain positions to form the interpolation reconstruction error. The training reconstructed fact pieces and the corrected supervision pieces are aligned at the true missing chain positions, conflicting fact piece chain positions, and low-confidence source chain positions to form a collaborative error correction. The training reconstructed fact piece and the holding supervision piece are aligned at non-reconstructed chain bits to form the original holding error; The interpolation reconstruction error, co-correction error, and original preservation error are combined into the total training error, and the trainable network parameters are updated in reverse based on the total training error. When the total training error corresponding to the verification fact chain sample is less than the training error threshold, or when the number of training rounds reaches the upper limit of the number of training rounds, the trainable network parameters are fixed.

7. The enterprise production data collaborative processing system based on a multi-agent system according to claim 1, characterized in that, The improved SAITS collaborative reconfiguration module is specifically as follows: The production fact pieces in the production fact chain are arranged according to the fact piece chain position and mapped to the fact piece content embedding; The factual content embedding, chain position embedding, source credibility embedding, and reconstructed tag embedding are superimposed in the same position to form the factual hidden state sequence; The first self-attention interpolation block performs query projection, key projection, and value projection on the sequence of hidden states of the fact slice, and weights and aggregates the value hidden states according to the chain position matching relationship between the query hidden states and the key hidden states to form the initial hidden states. The production acceptance fact reconstruction layer generates a fact piece chain position index table based on the production acceptance fact chain, and performs chain position back-sorting on the initial hidden state according to the fact piece chain position index table. It then performs anchor position acceptance reconstruction around the reconstruction anchor position corresponding to the fact piece to be reconstructed, forming the production acceptance reconstruction hidden state. The second self-attention interpolation block performs a second-stage self-attention interpolation on the hidden state of production acceptance and reconstruction, forming a collaborative interpolation hidden state. The interpolation output structure restores the cooperative interpolation hidden state to the reconstructed fact piece.

8. The enterprise production data collaborative processing system based on a multi-agent system according to claim 7, characterized in that, The process of anchoring and reconstructing around the anchor points corresponding to the fact pieces to be reconstructed, forming a hidden state of production-accepted reconstruction, is as follows: Using the chain position corresponding to the fact piece to be reconstructed as the reconstruction anchor position, extract the corresponding mutual verification hidden state, the previous position inheritance hidden state, and the subsequent position verification hidden state from the initial hidden state after the chain position is back-rowed. The hidden state of mutual verification comes from the production fact piece that is in the same production acceptance position as the reconstruction anchor. The hidden state of the previous acceptance comes from the production fact piece that the acceptance pointer points to the reconstruction anchor. The hidden state of the subsequent verification comes from the production fact piece that the reconstruction anchor points to and has a verification relationship. After linear projection, the hidden states of mutual verification, previous inheritance, and subsequent verification maintain the same hidden state dimension as the hidden state of the reconstructed anchor position. The corresponding mutual verification hidden state and the reconstructed anchor hidden state are multiplied dimension by dimension to obtain the mutual verification consistency component, and then subtracted dimension by dimension to obtain the mutual verification deviation component. The mutual verification consistency component and the mutual verification deviation component are combined in the same position to form the corresponding mutual verification code. The previous hidden state and the reconstructed anchor hidden state are subtracted dimension by dimension, and the chain offset representation of the corresponding accepting pointer is superimposed to form the previous accepting code; The hidden state of the post-verification is subtracted from the hidden state of the reconstructed anchor position in one dimension, and the chain offset representation of the corresponding verification relationship is superimposed to form the post-verification code; The source trust representations corresponding to the same position mutual verification code, the previous position acceptance code, the subsequent position verification code, and the reconstructed anchor are concatenated and then mapped by a threshold value to form the anchor retention threshold, the previous position write threshold, and the subsequent position verification threshold. The preceding latent state is projected forward to form a forward candidate state, and the following latent state is projected backward to form a backward candidate state. The anchor position retention threshold is multiplied dimension-wise with the reconstructed anchor position hidden state. The previous write threshold is multiplied dimension-wise with the forward acceptance candidate state. The subsequent verification threshold is multiplied dimension-wise with the reverse verification candidate state. The three products are added in the same position and then written back to the reconstructed anchor position to form the production acceptance reconstructed hidden state.

9. A collaborative processing system for enterprise production data based on a multi-agent system according to claim 1, characterized in that, The aforementioned closed-loop curing module specifically includes: Establish a collaborative shadow fact chain based on the fact piece chain position and acceptance pointer in the production acceptance fact chain. The collaborative shadow fact chain retains the same chain position arrangement and acceptance pointer as the production acceptance fact chain. The reconstructed fact pieces are pasted back to the production receiving fact chain according to the corresponding chain position, and the preceding production fact pieces, following production fact pieces, and corresponding mutual verification fact pieces adjacent to the reconstructed fact pieces are located. Perform preceding succession verification on the object continuation state between the reconstructed fact piece and the preceding production fact piece; perform subsequent succession verification on the state succession result between the reconstructed fact piece and the subsequent production fact piece; and perform peer closure verification on the source mutual verification state between the reconstructed fact piece and the peer mutual verification fact piece. When the previous position acceptance verification, the subsequent position acceptance verification, and the same position closure verification all pass, the reconstructed fact piece is written into the corresponding chain position in the collaborative shadow fact chain, and the chain position pointer and source pointer of the original fact piece before reconstruction are retained in the corresponding chain position; When the verification of the previous position, the verification of the subsequent position, or the verification of the same position fails, the original fact piece before reconstruction is retained in the corresponding chain position in the collaborative shadow fact chain. The reconstructed fact piece is marked as the disputed fact piece, and the chain position, source agent, and verification type of the disputed fact piece are recorded. The disputed factual pieces are attached to the corresponding production intelligent agents, and the original factual pieces before reconstruction are retained in the production factual chain.

10. A collaborative processing system for enterprise production data based on a multi-agent system according to claim 1, characterized in that, The collaborative result output module is specifically as follows: Align the production acceptance fact chain and the collaborative shadow fact chain item by item according to the fact piece chain position, and distinguish between the original fact pieces that have not been reconstructed and the verified fact pieces written into the collaborative shadow fact chain; Configure collaborative correction tags for the verified fact pieces, and register the collaborative correction tags together with the source location, chain location, and production agent affiliation of the corresponding original fact pieces; For disputed fact pieces that are not written into the collaborative shadow fact chain, retain the original fact pieces in the production successor fact chain, and register the unverified type corresponding to the disputed fact piece as a dispute prompt mark; The enterprise production data is processed collaboratively by combining the original factual footage, verified factual footage, and dispute alerts. The collaborative processing results of enterprise production data are output to the enterprise production data management terminal, and collaborative correction markers, dispute reminder markers, and the corresponding production agent affiliation relationships are output simultaneously.

Images