A Deep Learning-Based Collaborative Approach to Digital Audit Management

Abstract

This invention discloses a deep learning-based collaborative method for digital audit management, comprising the following steps: S1, extracting and indenting audit flow records to form a set of audit clue fragments; S2, threading along the corresponding points of the items and inserting them into the item anchor positions to form item caliber bundles; S3, peeling off the caliber fibers and snapping together the boundary pieces between slots to form a caliber slot matrix; S4, inputting the improved EdgeNeXt network model, forming a caliber slot embedded state through the caliber slot embedding layer; S5, splitting the data through the slot snap attention layer and the slot snap untangling latching mechanism to form three residual paths; S6, re-attaching, correcting, and calibrating the three residual paths through the untangling correction layer to form a caliber untangling state; S7, pulling the item anchor positions through the closed output layer to form the audit collaborative processing result. This invention uses audit caliber fiber untangling and an improved EdgeNeXt network model to achieve collaborative processing in digital audit management.

Description

Technical Field

[0001] This invention relates to the field of intelligent audit management technology, and in particular to a collaborative digital audit management method based on deep learning. Background Technology

[0002] Existing digital audit management methods typically collect records of projects, funds, contracts, procurement, assets, acceptance, rectification, and responsibilities, and then identify audit issues through rule comparison, anomaly screening, or process node verification. While these methods can improve data processing efficiency, most processing remains at the level of field matching, material integrity verification, and process status judgment, lacking in-depth analysis and structured connection of the relationships between the same audit matter under different management perspectives.

[0003] With the application of deep learning technology to audit data analysis, some methods have begun to use neural networks to identify audit risks or to use graph structures to link multi-source audit data. However, existing models usually encode multi-caliber audit data uniformly and then make overall judgments, which can easily mix misconnected relationships into ordinary feature associations, making it difficult to identify hidden problems such as caliber misappropriation, caliber drift, attribution gaps, and collaborative closure distortion.

[0004] In addition, existing systems mostly use model identification results as risk warnings, and then manually assign verification, rectification and review tasks. The models lack a structure for untangling and diverting residuals for audit matters, which affects the accuracy and traceability of audit collaborative processing results.

[0005] Therefore, how to provide a collaborative digital audit management method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based collaborative digital audit management method. This invention employs a combination of audit caliber fiber untangling and an improved EdgeNeXt network model to extract, indent, and thread audit flow records, constructing a caliber bundle capable of supporting multiple management caliber relationships. This bundle is further separated into caliber fibers and caliber slot matrices. Based on this, a caliber slot embedding state is formed through a caliber slot embedding layer. A slot buckle untangling latching mechanism is introduced in the slot buckle attention layer, transforming the original unified transposition attention path into a slot buckle structure with parallel slot retention branches, slot buckle engagement branches, and reverse buckle stripping branches. This allows for the diversion of normal acceptance, misaligned acceptance, and missing acceptance, forming main residuals, isolation residuals, and fault residuals. Finally, an untangling correction layer corrects the attribution of multiple caliber relationships and calibrates faults, and a closed output layer generates the audit collaborative processing results. This invention can identify issues such as misdeduction, shell company operations, drift, and gaps in the management of the same audit matter under different management standards, and avoids the simple mixing and judgment of audit data from multiple standards. It has the advantages of clear correlation between standards, accurate collaborative handling, traceable responsibility nodes, and reliable audit closure results.

[0007] A deep learning-based collaborative method for digital audit management according to an embodiment of the present invention includes the following steps: S1. By meticulously analyzing and verifying the execution of audit flow records, the traces and collaborative actions of the audited entities are linked together to form audit clue fragments, thus creating a set of audit clue fragments. S2. Audit clue fragments are strung together along the subject matter to the execution position. The audit clue fragments with different scopes are superimposed to the subject anchor position. The gap fragments are pressed into the boundary of the subject anchor position to form a subject scope bundle. S3. The diameter bundle of the item is stripped into diameter fibers, and the notched piece is retained at the boundary of the item anchor position. Adjacent diameter fibers are snapped together along the flow direction to form a groove boundary piece, forming a diameter groove matrix. S4. The diameter slot matrix is ​​fed into the improved EdgeNeXt network model. The diameter slot embedding layer presses the diameter fiber into slot segments and presses the inter-slot boundary sheet into the inter-slot receiving position to form the diameter slot embedding state. S5. Pay attention to the groove insertion state of the groove bearing interface and embed the groove buckle untangling latching mechanism. The groove retains the support to lock the groove segment. The groove inter-clamping support will clamp the groove bearing position that is in the same direction as the anchor position into the main residual. The reverse clamping peeling support will transfer the groove bearing position that is not in the same direction to the isolation residual. The missing groove segment will be transferred to the fault residual, forming three residuals. S6. The unwrapping correction layer receives the three residuals. The main residual is back-attached to the event anchor position. The isolation residual is corrected for attribution. The fault residual and the attribution gap are merged and calibrated to form the caliber unwrapping state. S7. Close the output layer interface untangling status, guide the corresponding anchor position along the untangling status, lock the collaborative verification object, responsibility confirmation node and rectification chain position, backfill the review mark, and form the audit collaborative handling result.

[0008] Optionally, S1 specifically includes: S11. Perform field regularization and empty space masking on the audit flow records, retain the record segments that point to the audit object, management scope and collaborative actions, and remove redundant record segments that are irrelevant to the attribution of the matter to form a record segment to be extracted. S12. The record segment to be extracted is extracted along the continuous items of the audit object. Adjacent record segments are continuously verified by the chain of attribution. When the source relationship and the direction of flow both belong to the same audit object, they are connected as the same item thread. When the chain of attribution is broken, it is marked as a broken thread piece. S13. The item wire is connected to the management scope and the indentation is calibrated for the scope name, scope source and scope change traces to obtain the item wire with scope indentation. The broken wire piece is pressed into the boundary of the corresponding item wire. S14. The item wire with caliber indentation is encapsulated in the same position as the cooperating action execution, and the cooperating action is attached to the corresponding item wire's transfer position to form an audit clue piece; S15. Audit clue fragments are arranged in order of their subject matter, and broken thread fragments retain boundary markers along with their corresponding audit clue fragments, and are collected into an audit clue fragment set.

[0009] Optionally, S2 specifically includes: S21. Audit clue fragments are set up according to the matter, with the audit object trace falling into the object position, the caliber indentation falling into the caliber position, and the coordinated action falling into the action position, thus creating a clue fragment to be threaded. S22. The thread pieces to be threaded are threaded along the object position in the execution sequence. The thread pieces to be threaded that are continuous in the object position and connected by the action position are strung together to form a matter line. The thread pieces to be threaded that have inconsistent names of the caliber positions retain the original caliber indentation and are arranged with the matter line. S23. The item line is pressed into the item anchor position, the continuous thread piece falls into the main layer of the anchor position, and the dispersed diameter thread piece is attached to the side layer of the anchor position along the original diameter indentation. The main layer of the anchor position and the side layer of the anchor position are embedded to form the anchor position layer. S24. The thread piece to be threaded is peeled off into a belonging gap piece when the object position is disconnected, the action position is reversed, and the caliber position is missing. The belonging gap piece is pressed into the boundary of the anchor position and snapped to the edge of the anchor layer. S25. The main anchor layer, side anchor layer, and belonging gap piece within the anchor position are bound together in the order of threading to form the anchor diameter bundle.

[0010] Optionally, S3 specifically includes: S31. The bundle of matters is spread out along the anchor position of the matter as the surface to be stripped. The audit clue pieces within the bundle fall into the corresponding bundle path according to the indentation of the matters, and the belonging gap pieces are docked at the boundary of the anchor position of the matter. S32. Audit clue pieces in the same bundle within the fiber bundle to be stripped are attached end to end according to the order of transfer, with a notch embedded at the break point to occupy the space, and stripped into the diameter of the fiber carrying the anchor position and the order of transfer. S33. Fibers of the same diameter are pressed into the diameter groove according to the diameter indentation. Fibers of the same diameter in the same groove are arranged in sequence along the flow order. Fibers of different diameters across grooves are aligned with the anchor position along the direction to form a diameter groove row with notches. S34. Adjacent diameter slots extract end-side fibers along the flow direction. When the end-side fibers are connected, they are pressed together to form a slot boundary piece. When the end-side fibers are disconnected, the notch piece pulls the notch to occupy the corresponding slot edge. S35, the diameter slot array, the inter-slot boundary plates and the slot edge notch occupants are assembled into a diameter slot matrix according to the anchor positions.

[0011] Optionally, S4 specifically includes: S41. The improved EdgeNeXt network model includes a slot embedding layer, a slot attention layer, an unwrapping correction layer, and a closed output layer. S42. The diameter groove matrix enters the diameter groove embedding layer. The diameter groove embedding layer spreads the matrix groove surface along the anchor position. The diameter fiber falls into the groove crimping position. The groove boundary piece falls into the groove receiving position candidate area. The notch occupant is left at the groove edge notch position along with the assigned notch piece. S43. The diameter fibers in the crimping position in the groove are arranged adjacently along the flow direction. The diameter fibers of the same anchor position are crimped end to end to form a segment in the groove. The diameter fibers of the anchor position that are broken stop at the notch position on the side of the groove and fit with the notch piece. S44. The inter-slot boundary piece in the inter-slot receiving position candidate area is projected to the end side of the adjacent inter-slot segment. The inter-slot boundary piece with the end side anchor position is pressed into the inter-slot receiving position. The inter-slot boundary piece with the end side anchor position is transferred to the slot edge notch position. S45. The segments inside the slot, the receiving positions between slots, and the notches on the edge of the slot are encapsulated in the order of the diameter slots. The segments inside the slot carry the position of the diameter inside the slot, the receiving positions between slots carry the connection position of the adjacent slots, and the notches on the edge of the slot carry the position of the notch to which they belong, thus forming a diameter slot embedded state.

[0012] Optionally, S5 specifically includes: S51. After the groove fastening layer bearing interface diameter groove is embedded, the unified transfer attention channel is disassembled. The segment inside the groove remains in the groove along the diameter groove and is retained in the groove. The groove bearing position is transferred into the groove fastening support along the adjacent groove edge. The groove edge notch position is transferred into the reverse fastening peeling support along the anchor boundary to form a parallel groove fastening attention surface. S52. The slot retainer performs slot transposition on the slot segment. The slot segment only has contact crimping with adjacent segments in the same diameter slot. The original diameter of the segment and the anchor position are locked in the slot retainer, thus forming a slot retainer piece. S53. The inter-slot fastening support performs double-end alignment on the inter-slot receiving position. The two ends of the inter-slot receiving position pull the retaining piece in the adjacent diameter slot. The two-end segments of the anchor position and the flow direction receiving are pressed into the same fastening position to form the inter-slot fastening piece. S54. The reverse stripping support performs reverse verification on the slot connection position that has not entered the fastening position. The segments with opposite anchor positions, reverse flow direction, and belonging gap are removed from the normal fastening path and enter the isolation channel. The slot retaining piece without adjacent end segment is pressed into the fault channel. S55, the slot buckle unwinding latching mechanism performs latching and merging on the slot retainer, slot interlocking clip, isolation channel and fault channel within the slot buckle attention layer. The slot interlocking clip is incorporated into the main propagation path, and the isolation channel and fault channel are isolated from the main propagation path. S56. The main propagation path carries the inter-slot interlocking piece back to the residual main slot. The segment that is separated from the normal interlocking path is pressed into the residual isolation slot through the isolation channel. The slot retaining piece that lacks the adjacent end segment is pressed into the residual fault slot through the fault channel. The residual main slot, residual isolation slot and residual fault slot converge into three residual paths.

[0013] Optionally, S55 specifically includes: S551, the slot buckle unwinding latch mechanism is set at the three junctions of the slot buckle attention layer. A latch junction surface is opened along the diameter attention surface. The latch junction surface receives the slot buckle piece, the segment in the isolation channel and the segment in the fault channel, and introduces the buckle latch groove, the reverse buckle isolation groove and the fault retention groove respectively. The retaining piece in the groove is attached to the edge of the corresponding diameter groove as the end side verification piece. S552, the latching slot first clamps the two end segments of the inter-slot latching piece, and then pulls the end side verification piece in the two diameter slots. When the end segment and the end side verification piece are aligned and connected in the same position on the anchor position and in the flow direction, the inter-slot latching piece is embedded with the end side verification piece to form a latching closed piece, and the latching closed piece is connected to the main propagation path. S553. Fragments that have not been made into snap-fit ​​closures do not enter the main propagation path. They are diverted along the cause of failure by the snap-fit ​​latch groove. Fragments with incorrect anchor positions and reverse connection in the flow direction are cut into the reverse snap isolation groove. Fragments with missing end-side verification pieces are transferred into the tomographic retention groove. Only snap-fit ​​closures are retained in the main propagation path. S554. After the cut-in segment is received by the reverse-clamping isolation groove, the cut-in segment is clamped in the opposite direction along the original groove receiving position, the reconnection relationship between the cut-in segment and the normal fastening edge is stripped off, and then the misalignment cause mark is pressed into the cut-in segment to make a reverse-clamping isolation plate that is separated from the main propagation path. S555. After verifying the missing segment of the fault retention groove receiving end side, press the missing end segment along the groove edge notch. The missing end segment and the assigned notch segment are stacked on the same side. The disconnected groove receiving position is locked as the fault boundary, and a fault retention segment that does not participate in the groove fastening is made. S556, the residual outflow channel sequentially receives the snap-fit ​​closing plate, the reverse snap-fit ​​isolation plate and the fault retention plate. The snap-fit ​​closing plate is reinjected into the residual main channel along the main propagation path and pressed into the main residual. The reverse snap-fit ​​isolation plate is gathered into the residual isolation channel and pressed into the isolation residual. The fault retention plate is gathered into the residual fault channel and pressed into the fault residual. The residual isolation channel and the residual fault channel cut off the reinjection connection to the residual main channel.

[0014] Optionally, S6 specifically includes: S61. After the unwrapping correction layer receives the three residuals, a residual correction surface is opened. The main residual is attached to the main edge of the anchor position, the isolation residual falls into the side edge of the anchor position, and the fault residual is embedded in the edge of the anchor position gap. The three types of residuals are aligned under the same anchor position. S62. The main residual is corrected by backing along the main edge of the anchor position. The normally fastened slot connection position is backed to the adjacent diameter fiber end section. The slot segment returns to its original position with the original diameter slot to form the main closed piece. S63. The isolation residual is spread out along the side of the anchor position to form an isolation segment. The isolation segment is first stripped from the original groove connection position, and then transferred to the corresponding anchor position boundary along the cause of the misalignment of the diameter to make a correction piece. S64. The fault residual is attached to the edge of the anchorage gap and the fault calibration piece is pressed into the same fault boundary to make the fault calibration piece. S65, the main closure piece, the attribution correction piece, and the fault calibration piece are merged along the anchor position of the matter, and the results of normal snapping, attribution correction, and fault retention are merged into the caliber unwrapping state.

[0015] Optionally, S7 specifically includes: S71. After the unwrapping state of the closed output layer bearing interface, unfold the closed treatment surface along the anchor position. The main closed piece that has been reattached is incorporated into the closed chain position, and the correction piece is assigned to the traction verification slot. The fault calibration piece is assigned to the traction verification slot, and the treatment surface is divided into slots. S72. The closed chain position is approved along the anchor position of the matter. The slot receiving position after the backing is completed is locked with the corresponding slot segment to form a continuous diameter chain. The continuous diameter chain is pressed into the bottom slot of the collaborative treatment. S73. Verify the slot receiving the correction sheet, lock the collaborative verification object along the deviation position in the correction sheet, and attach the collaborative verification object to the corresponding item anchor position. S74. Verify the fault calibration plate of the trench, locate the missing acceptance position along the fault boundary, assign the missing plate to the responsibility confirmation node, and use the missing acceptance position to guide the rectification chain position. Verify the backfilling of the mark to the boundary of the anchor position. S75. Collaboratively handle the processing pieces within the bottom trench closing chain position, verification trench position, and review trench position, and encapsulate them as audit collaborative processing results according to the item anchor position.

[0016] The beneficial effects of this invention are: This invention uses a process of item extraction, caliber creasing, and co-location threading to connect scattered audit object traces, management calibers, and collaborative actions in audit workflow records into audit clues, which are further structured into item caliber bundles and caliber slot matrices. Through this method, scattered records of the same audit item under different management calibers can be compressed into a unified item anchor point, avoiding the unclear caliber correlation problems caused by relying solely on field matching, process status, or material completeness judgments in traditional digital auditing.

[0017] This invention introduces an improved EdgeNeXt network model and sets up a slot untangling latch mechanism in the slot attention layer. This mechanism locks, latches, and strips the intra-slot segments, inter-slot connection points, and missing connection relationships within the slot embedding state, forming primary residuals, isolation residuals, and discontinuity residuals. As a result, normal connection relationships, misaligned connection relationships, and missing connection relationships are no longer mixed into the same propagation path, enabling more accurate differentiation of misaligned, misleading, drifting, and discontinuity states in audit matters, thus improving the accuracy and traceability of multi-caliber audit data processing.

[0018] This invention further uses an unwrapping correction layer and a closed output layer to perform attribution correction, fault calibration, and collaborative processing and merging of the three residuals, forming audit collaborative verification objects, responsibility confirmation nodes, rectification chain positions, and review markers. This invention can directly transform model recognition results into collaborative processing results in digital audit management, reducing the cost of repeated manual verification and task assignment, and improving the accuracy of audit collaborative closure, the clarity of responsibility positioning, and the reliability of rectification review. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a deep learning-based collaborative digital audit management method proposed in this invention. Figure 2 This is a schematic diagram of the improved EdgeNeXt network model in a deep learning-based collaborative digital audit management method proposed in this invention. Figure 3 This is a schematic diagram of the slotted latch untangling mechanism of a deep learning-based digital audit management collaborative method proposed in this invention. Detailed Implementation

[0020] 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.

[0021] refer to Figures 1-3 A deep learning-based collaborative approach to digital audit management includes the following steps: S1. By meticulously analyzing and verifying the execution of audit flow records, the traces and collaborative actions of the audited entities are linked together to form audit clue fragments, thus creating a set of audit clue fragments. S2. Audit clue fragments are strung together along the subject matter to the execution position. The audit clue fragments with different scopes are superimposed to the subject anchor position. The gap fragments are pressed into the boundary of the subject anchor position to form a subject scope bundle. S3. The diameter bundle of the item is stripped into diameter fibers, and the notched piece is retained at the boundary of the item anchor position. Adjacent diameter fibers are snapped together along the flow direction to form a groove boundary piece, forming a diameter groove matrix. S4. The diameter slot matrix is ​​fed into the improved EdgeNeXt network model. The diameter slot embedding layer presses the diameter fiber into slot segments and presses the inter-slot boundary sheet into the inter-slot receiving position to form the diameter slot embedding state. S5. Pay attention to the groove insertion state of the groove bearing interface and embed the groove buckle untangling latching mechanism. The groove retains the support to lock the groove segment. The groove inter-clamping support will clamp the groove bearing position that is in the same direction as the anchor position into the main residual. The reverse clamping peeling support will transfer the groove bearing position that is not in the same direction to the isolation residual. The missing groove segment will be transferred to the fault residual, forming three residuals. S6. The unwrapping correction layer receives the three residuals. The main residual is back-attached to the event anchor position. The isolation residual is corrected for attribution. The fault residual and the attribution gap are merged and calibrated to form the caliber unwrapping state. S7. Close the output layer interface untangling status, guide the corresponding anchor position along the untangling status, lock the collaborative verification object, responsibility confirmation node and rectification chain position, backfill the review mark, and form the audit collaborative handling result.

[0022] In this embodiment, S1 specifically includes: The audit workflow is first broken down into record segments. Each record segment is the smallest processed fragment retained after field standardization. Each record segment corresponds to an audit object, a management scope, and a collaborative action location. Field standardization is performed on the audit object number, object name, document source, management scope, collaborative action, responsible entity, processing time, attachment number, and version number. The audit object number retains 6 to 32 valid characters, the contract number, purchase number, and asset number retain 4 to 40 valid characters, and the attachment number retains 4 to 64 valid characters. Numbers exceeding the range are truncated to the main number segment; if the truncated segment still does not meet the range, the time stamp is left blank. The processing time is uniformly formatted as year, month, day, hour, minute, and second. If a time stamp is missing, it is first supplemented using the version number, which adopts a two-segment structure with the main version number first and the revision number second. If the main version numbers are the same, the record segment with the smaller revision number is placed first. If the version number is still missing, it is supplemented according to the system's receiving order, which is an incrementing sequence generated when the platform receives the record segment, starting from the first record in the same processing batch and incrementing by 1.

[0023] Empty space masking targets fields containing empty strings, hyphens, "none," zero-value placeholders, and fields containing only spaces. Name fields are cleaned by removing spaces, full-width characters, parenthetical comments, system-added words, and consecutive duplicate characters; fields with fewer than two Chinese characters after cleaning are marked as empty. System-added words are version hints, import batch hints, and temporary status hints appended to the end of the name by the platform. Records where the audit object number, document source, and collaborative action are all empty are stripped into redundant record segments. When the number of empty fields exceeds 60% of the total number of record fields, the record is moved to a manual entry pool. After completion, it re-participates in field regularization; record segments that have not completed regularization are not included in this round of item decomposition. Record segments that have undergone field regularization and empty space masking constitute the record segments to be decomposition.

[0024] The record segments to be extracted are processed in ascending order of processing time. The continuity of audit objects is verified through object keys, source relationships, and flow direction. The object key is constructed from valid fields in the audit object number, contract number, purchase number, asset number, and attachment number. Valid fields are non-empty, have character lengths within the corresponding range, and have not been truncated by the manual entry pool. Audit objects are considered continuous when object keys are completely identical. When an object key is missing, object name similarity is calculated. Object name similarity is the number of identical characters in the cleaned object names divided by the number of characters in the longer object name, ranging from 0 to 1. A similarity of 0.90 indicates continuity, while a similarity below 0.75 indicates a break. Records between 0.75 and 0.90 enter the review pool. If the review passes, the audit object is added to the audit thread; if the review fails, it is marked as a broken thread. Source relationships are determined by the source chain number, with project initiation, budget, contract, procurement, acceptance, assets, rectification, and review recorded as 1 to 8 respectively. Source acceptance is established when the source chain number of the subsequent record segment is the same as or increases by 1 compared to the previous record segment. Flow direction is determined by the collaborative action number, with submission, processing, confirmation, and review recorded as 1 to 4 respectively. Flow acceptance is established when the action number of the subsequent record segment is not less than the action number of the previous record segment. When object key acceptance, source acceptance, and flow acceptance are all valid simultaneously, adjacent record segments are linked as the same thread. When a disconnection occurs in audit object disconnection, source disconnection, or flow reversal, it is marked as a broken thread, and the disconnection field, disconnection location, and disconnection reason are recorded.

[0025] After the item thread receives the management caliber position, it undergoes indentation calibration. The management caliber position is the column position of the management caliber field, the source module position, and the version position. The caliber change trace consists of the caliber name before the change, the caliber name after the change, the caliber source, the version number, and the change action. The change action is limited to addition, replacement, withdrawal, and revision. When the caliber names are the same, the source chain sequence number has not rolled back, and the revision number difference is 1, continuous indentation is formed; when the object key remains unchanged and the caliber name similarity is less than 0.80, change indentation is formed; when the caliber source changes but the object key and the major version number remain consistent, cross indentation is formed. The broken thread piece is pressed into the boundary of the corresponding item thread, and the reason for the breakage is simultaneously attached to the boundary mark, creating an item thread with caliber indentation. The item thread with caliber indentation is encapsulated in the same position as the collaborative action execution. The collaborative action, responsible party, and processing time are attached to the same flow position in the item thread. The flow position is determined by the record segment arrangement sequence number, which starts from 1 and increments. After encapsulation, it constitutes an audit clue piece. Audit clue fragments are arranged in the order of the events they point to. The order of events is determined by the processing time, version number, and system reception order, and they are eventually compiled into a set of audit clue fragments.

[0026] In this embodiment, S2 specifically includes: When a set of audit clue fragments enters the same-position threading process, threading positions are first created according to the order of the matters they point to. These threading positions include object positions, scope positions, and action positions. The object position receives the audit object's traces, the scope position receives the scope imprints, and the action position receives the coordinated actions. All three positions share the same fragment number, starting from 1 and increasing sequentially according to the matter's direction. Once the audit object's traces, scope imprints, and coordinated actions fall into the same threading position, a thread fragment to be threaded is created.

[0027] The thread pieces to be threaded are sequentially threaded along the object positions. Continuity of object positions is determined by object key and name similarity; if object keys are identical, they are considered continuous. If an object key is missing, a name similarity of 0.90 or higher is considered continuous, below 0.75 is considered broken, and those between 0.75 and 0.90 enter the review pool. Action position continuation is determined by the collaborative action sequence number: Submission, Processing, Confirmation, and Review are numbered 1 to 4 respectively. An action position continuation is determined when the action sequence number of the subsequent thread piece to be threaded is not less than that of the previous thread piece; otherwise, it is considered a reverse continuation. Thread pieces to be threaded with continuous object positions and continuating action positions are linked together to form an event line. When the names of the caliber positions are inconsistent but the object positions are continuous, the event line is not cut, and the original caliber indentation remains on the event line along with the thread piece to be threaded. When a caliber position is missing, the corresponding thread piece to be threaded is not merged into the event line but becomes a missing piece. Missing caliber positions include caliber name being empty, caliber source being empty, and caliber version number being empty.

[0028] After the item line is pressed into the item anchor position, continuous item line pieces fall into the main anchor position layer, and dispersed diameter item line pieces are attached to the side anchor position layer along the original diameter indentation. The main anchor position layer receives continuous flow segments of the same item, while the side anchor position layer receives the diameter name, diameter source, and diameter version. When the main anchor position layer and the side anchor position layer are fitted together, the position with the same segment number is used as the fitting point, and the side anchor position layer is attached to the outside of the corresponding segment of the main anchor position layer to form the anchor position layer. Item line pieces with broken object position, reversed action position, or missing diameter position are peeled off as belonging gap pieces. The belonging gap pieces retain the item source, reason for breakage, and segment number, and are pressed into the item anchor position boundary. The item anchor position boundary records the gap type, gap occurrence number, and item line position; the gap type includes broken object, reversed action, and missing diameter; the gap occurrence number increases from 1 according to the order in which the belonging gap pieces enter the item anchor position boundary; and the item line position is the segment number of the previous item line piece to be fitted adjacent to the belonging gap piece.

[0029] The anchorage main layer, anchorage side layer, and assigned gap piece are bound together in the order of threading. The threading order first compares the item line segment number, then the caliber version number, and finally the gap occurrence number. The segment with the smaller segment number is placed in the front position, the side layer with the smaller caliber version number is attached to the front side, and the assigned gap piece with the smaller gap occurrence number is left at the front boundary. After binding, the item caliber bundle is formed, which carries the continuous item line, the dispersed caliber indentation, and the assigned gap boundary.

[0030] In this embodiment, S3 specifically includes: Before the fiber stripping process, the fiber bundle is spread out along the anchor points to form the fiber bundle surface to be stripped. An anchor point is a fixed position that receives the same audit matter. The anchor point number is generated by sequentially concatenating the audit object number, document source number, and matter line segment number. The audit object number retains 6 to 32 characters, and the document source number retains 4 to 40 characters. If there is no document source number, the first 20 valid characters of the cleaned document source name are combined with the system's receiving order to generate the anchor point number. When comparing anchor point numbers, the audit object number is compared first, then the document source number, and finally the matter line segment number. If the first few characters are the same, the shorter number is ranked first, and the smaller numerical sequence number is ranked first. The fiber bundle surface to be stripped receives the anchor points horizontally and the diameter indentation vertically. The horizontal single block capacity is 256 anchor points. If there are more than 256 anchor points, it is cut into multiple bundle blocks in ascending order of anchor point number. Adjacent bundle blocks retain 8 anchor points as overlapping boundaries. The longitudinal diameter slot capacity is 8 slots, which are respectively for funds, projects, contracts, procurement, assets, acceptance, rectification and responsibility. When the corresponding diameter is missing, an empty slot is reserved. The empty slot occupies the slot index but does not participate in the inter-slot boundary sheet pressing.

[0031] When an audit clue falls into a bundle channel, the caliber name, caliber source, major version number, and revision number are extracted from the caliber indentation. Audit clues with the same caliber name, caliber source, and major version number fall into the same bundle channel; those with missing caliber names or missing caliber sources enter the gap bundle channel. Bundle channel numbers start from 1 and increase sequentially according to the vertical caliber slots, with gap bundle channel numbers recorded as 0. Attribution gap clues are anchored at the corresponding item anchor boundary. The attribution gap clue number is generated by concatenating the item anchor number, gap type, and gap occurrence sequence number, which increases sequentially from 1.

[0032] Audit clue fragments within the same channel are assigned a flow order code. The flow order code consists of a time ranking, a collaborative action sequence number, a caliber version ranking, and a fragment sequence number. The calculation method is: time ranking multiplied by 10000, plus a collaborative action sequence number multiplied by 1000, plus a caliber version ranking multiplied by 100, and then the fragment sequence number. The time ranking is the ascending order of processing time within the same channel, with the earliest processing time ranked 1. If a processing time is missing, the caliber version ranking is used to fill the gap; if both processing time and caliber version are missing, the system receiving order is used to fill the gap. The caliber version ranking first compares the major version number, then the revision number; the smaller major version number ranks first, and if the major version numbers are the same, the smaller revision number ranks first. The collaborative action sequence numbers are: Submit 1, Process 2, Confirm 3, Review 4. Audit clue fragments with smaller flow order codes are ranked first, and the flow direction is determined by the ascending order of the flow order code.

[0033] Adjacent audit clue segments within the same bundle are judged for beginning and end connection. Adjacent audit clue segments are considered continuous segments when the anchor numbers are consistent, the sequence number of the subsequent collaborative action is not less than the sequence number of the preceding collaborative action, and the difference between the sequence number of the subsequent segment and the sequence number of the preceding segment equals 1. Disconnections are identified when the anchor numbers are inconsistent, the collaborative action sequence number is reversed, the difference in segment sequence numbers is greater than 1, or a disconnection occurs during gap bundle intervention. A gap placeholder is embedded at the disconnection point. The gap placeholder number is generated by concatenating the anchor number, bundle number, and gap occurrence sequence number, and records the segment sequence number before and after the disconnection, the disconnection distance, the gap type, and the assigned gap segment number; the disconnection distance is the segment sequence number after the disconnection minus the segment sequence number before the disconnection. Continuous segments are stripped into diameter fibers. The diameter fiber number is generated by concatenating the anchor number, bundle number, and the sequence number of the first segment, and carries the anchor number, bundle number, flow sequence code, and gap placeholder number.

[0034] Fibers of different diameters are pressed into the diameter slots according to their bundle number. Fibers with the same bundle number enter the same diameter slot, while those with notches do not enter normal diameter slots. Each diameter slot retains a maximum of 512 fibers. If more than 512 fibers are retained, the slot is divided into multiple segments according to the ascending order of the flow sequence code. Adjacent segments retain 16 fibers as a boundary. Fibers within a diameter slot are arranged in ascending order of the flow sequence code; when flow sequence codes are the same, the fiber with the smaller item anchor number is placed first. Fibers crossing slots are aligned along the item anchor number. Fibers with the same item anchor number are pressed into the same anchor alignment row; when item anchor numbers are different, the fibers remain in their respective anchor alignment rows. Fibers with notches retain their notches in their original positions; these notches do not participate in normal alignment but remain at the slot edge along with the corresponding fiber. After completing the in-slot arrangement and cross-slot alignment, the fibers within a single diameter slot form a column of notched diameter slots.

[0035] Adjacent diameter slots extract end-side fibers along the flow direction, with adjacent diameter slots determined by the longitudinal diameter slot sequence. End-side fibers are those closest to the boundary of adjacent slots, with left and right end-side fibers using corresponding diameter fiber numbers. When the anchor positions of the left and right end-side fibers are identical, and there are no gaps between the two end-side fibers, they are pressed together to form an inter-slot boundary piece. The inter-slot boundary piece number is generated by concatenating the left end-side fiber number, the right end-side fiber number, and the anchor position number; the boundary sequence is calculated by adding the left end-side fiber flow sequence code and the right end-side fiber flow sequence code, then dividing by 2, with the result rounded to two decimal places. When the anchor positions are inconsistent or there is a gap between the end-side fibers, the assigned gap piece is placed at the slot edge where the gap is located. The caliber slots, inter-slot boundary pieces, and slot edge notches are assembled into a caliber slot matrix. The matrix rows are ordered in ascending order of the anchor position number, and the matrix columns are ordered in the order of the eight longitudinal caliber slots. A single matrix unit supports a set of caliber fibers for one anchor position within a caliber slot. The inter-slot boundary pieces are embedded between adjacent matrix columns, and the slot edge notches are reserved at the corresponding matrix unit slot edge position.

[0036] In this embodiment, S4 specifically includes: The improved EdgeNeXt network model is based on the lightweight visual feature encoding structure of the original EdgeNeXt network model. The original EdgeNeXt network model encodes visual features through convolutional local aggregation, channel splitting, depthwise convolution, and transposed attention. The split channel groups enter a unified transposed attention path and are then reconnected via ordinary residual connections. In the collaborative scenario of digital audit management, the processing objects are aperture fibers, slot boundary pieces, and gap occupants. Directly using the original unified channel attention structure would mix normal connection, misaligned connection, and missing connection into the same path. The improved EdgeNeXt network model sets up an aperture slot embedding layer, a slot buckle attention layer, an unwrapping correction layer, and a closed output layer. These four layers sequentially handle the aperture slot matrix, the aperture slot embedding state, the three residuals, and the aperture unwrapping state.

[0037] The caliber slot embedding layer consists of a matrix slot surface unfolding unit, an in-slot pressing unit, an inter-slot projection unit, a notch retention unit, and a slot sealing unit. The matrix slot surface unfolding unit carries the caliber slot matrix, with matrix rows corresponding to item anchor positions and matrix columns corresponding to management caliber slots. Each matrix slot surface carries 256 item anchor position rows and 8 caliber slot columns. The 8 caliber slot columns sequentially carry funding, project, contract, procurement, asset, acceptance, rectification, and responsibility calibers. If there are fewer than 256 rows, empty anchor positions are used to fill the gap. Empty anchor positions only retain row numbers and placeholder marks and do not participate in pressing, projection, or notch retention.

[0038] After the diameter slot matrix enters the diameter slot embedding layer, the diameter fibers fall into the slot crimping position, the inter-slot boundary pieces fall into the inter-slot receiving position candidate area, and the notch occupant is left at the notch position on the slot edge along with the assigned notch piece. The in-slot crimping units arrange the diameter fibers in ascending order of the flow sequence code. When the item anchor row number is the same and the sequence number of the next segment minus the sequence number of the previous segment equals 1, the diameter fibers are crimped end to end as segments in the slot; when the item anchor row number is different, the difference in segment sequence number is greater than 1, or the disconnection state occurs during the notch occupant intervention, the diameter fiber stops at the notch position on the slot edge and fits with the assigned notch piece. The inter-slot projection unit attaches the fiber number on the left side of the inter-slot boundary piece to the end of the slot with the front diameter, and the fiber number on the right side to the end of the slot with the rear diameter. When the anchor row numbers on both sides are consistent and neither end position carries a gap, the inter-slot boundary piece is pressed into the inter-slot receiving position. When the anchor row numbers on both sides are inconsistent or a disconnection occurs during gap placement, the inter-slot boundary piece is transferred to the gap position on the slot edge. The boundary sequence is obtained by adding the fiber flow sequence codes on the left and right sides and dividing by 2, retaining two decimal places.

[0039] The gap placement unit arranges the gap positions on the slot edge according to the item anchor row number and the gap occurrence sequence number. In the gap status value, normal (no gap) is recorded as 0, object disconnection as 1, action reversal as 2, diameter missing as 3, and cross-slot disconnection as 4. The slot encapsulation unit completes the encapsulation according to the item anchor row number, diameter slot column number, and flow sequence code. Each encapsulation position is made into a 128-dimensional slot embedding vector. The slot index code occupies 16 dimensions and is obtained by dividing the diameter slot column number by 8; the item anchor row number code occupies 16 dimensions and is obtained by dividing the current item anchor row number by 256; the flow sequence code occupies 32 dimensions and is obtained by dividing the current flow sequence code by the maximum flow sequence code in the same slot; the acceptance mark code occupies 16 dimensions, with 1 indicating successful inter-slot acceptance and 0 indicating unsuccessful acceptance; the gap status code occupies 16 dimensions and is obtained by dividing the gap status value by 4; the position completion code occupies 32 dimensions and stores the fragment sequence number and boundary sequence normalization result. After encapsulation, the segments inside the slot, the receiving positions between slots, and the notches on the edge of the slot together form the in-slot embedded state.

[0040] Compared to the original EdgeNeXt network model, the improvement lies in transforming the original image channel-oriented embedding preprocessing into a structured embedding layer oriented towards audit caliber slots. The original model's channel splitting primarily preserved the local and global feature relationships between image channels, lacking anchor points, slot connection structures, and gap placement structures. The improved caliber slot embedding layer, before entering the slot attention layer, has already crimped the caliber fibers into intra-slot segments, crimped the inter-slot boundary pieces into inter-slot connection positions, and left unconnected segments at slot edge gap positions. The slot attention layer receives the caliber slot embedding state with intra-slot position, inter-slot connection, and gap status, which helps distinguish between normal connection, misaligned connection, and missing connection at the slot attention layer, preventing different management calibers from being mixed prematurely in the unified attention path.

[0041] In this embodiment, S5 specifically includes: After the slot buckle is embedded in the bearing interface, the bearing attention surface is first opened. Each encapsulation position in the bearing slot embedding state is a 128-dimensional vector, which carries the slot index code, the item anchor row number code, the flow sequence code, the acceptance mark code, the gap status code, and the position completion code. The bearing attention surface arranges the slot segments within the same bearing slot on the inside of the bearing slot, arranges the slot acceptance positions between adjacent bearing slots on the bearing slot boundary, and arranges the slot edge gap positions into the anchor edge. The slotted attention layer consists of attention surface splitting unit, slot retention branch, slot interlocking branch, reverse stripping branch, latching confluence entrance and residual outlet slot; attention surface splitting unit is responsible for splitting the unified transposed attention channel, slot retention branch maintains the continuity relationship inside a single diameter, slot interlocking branch handles the normal connection between adjacent diameters, reverse stripping branch intercepts misconnection and missing connection segments, latching confluence entrance performs pre-convergence diversion on the results of the three branches, and residual outlet slot outputs the main residual, isolation residual and fault residual.

[0042] The splitting unit first reads the caliber slot column number, slot index code, item anchor row number, flow sequence code, and gap status value from the encapsulation position. The caliber slot column number is the arrangement number of the caliber slot in the 8 managed caliber slots, with a value from 1 to 8; the slot index code is obtained by dividing the caliber slot column number by 8, with a value range from 0.125 to 1. Segment segments with the same caliber slot column number and a gap status value of 0 remain in the slot retention branch; slot connection positions with a difference of 1 between caliber slot column numbers are transferred to the slot engagement branch; slot edge gap positions with a gap status value greater than 0 enter the reverse stripping branch. The split slot segment, slot connection position, and slot edge gap position continue to retain the 128-dimensional vector, item anchor row number, and flow sequence code, so that the three branch paths can be re-aligned under the same item anchor position.

[0043] The in-slot retention branch performs in-slot transpose attention on in-slot segments. Each in-slot segment's 128-dimensional vector is split into four attention heads, each with 32 dimensions. Query slices, key slices, and value slices are obtained by multiplying the vector of the same in-slot segment by its corresponding parameter matrix. The parameter matrix is ​​128x128 in size, initialized with a normal distribution with a mean of 0 and a standard deviation of 0.02, and updated via backpropagation during training. Query slices and key slices are multiplied dimension-wise, summed, and then divided by the square root of 32 to obtain the in-slot attention intensity. The attention intensity within the same caliber slot is first subtracted from the maximum value within the same slot, then exponentialized and normalized, with the normalized value ranging from 0 to 1. The normalized value is multiplied by the corresponding value slice and summed to create the in-slot retention slice. The slot-reserved segment only absorbs segments with the same anchor row number for items within the same slot and adjacent segments after being arranged in ascending order of the flow sequence code; segments spanning slots do not participate in the slot-reserved branch calculation, and their original slot affiliation and item anchor position are locked within the slot-reserved branch.

[0044] The inter-slot fastening support performs double-end fastening at the inter-slot receiving position. The two ends of the inter-slot receiving position pull the retaining pieces in the left and right diameter slots. The left diameter slot is the adjacent diameter slot with the smaller slot column number, and the right diameter slot is the adjacent diameter slot with the larger slot column number. When the anchor row numbers of the two ends are the same, the anchor fastening item is scored as 0.40; when the right-end transfer sequence code is not less than the left-end transfer sequence code, the forward receiving item is scored as 0.30; when both the left-end and right-end gap status values ​​are 0, the complete receiving item is scored as 0.20; when the difference in slot column numbers is 1, the adjacent slot item is scored as 0.10. The above weights are fixed according to the importance order of item attribution, transfer receiving, gap integrity, and slot adjacency, totaling 1. If the corresponding conditions are not met, the score is 0. The four items are added together to obtain the inter-slot fastening score, with a score range of 0 to 1. When the inter-slot fastening score reaches 0.80, the retaining pieces in both slots and the inter-slot receiving position are pressed into the same fastening position to form an inter-slot fastening piece; when the score is lower than 0.80, the inter-slot receiving position is transferred to the reverse fastening stripping support for re-inspection.

[0045] The reverse-deposit stripping support performs stripping judgment on the inter-slot bearing positions and slot edge gap positions that have not entered the engagement position. When the anchor row number is inconsistent, the anchor misconnection item is scored as 0.40; when the right-end flow sequence code is less than the left-end flow sequence code, the flow direction reverse connection item is scored as 0.30; when a gap condition occurs where both the left-end and right-end gap status values ​​are greater than 0, the gap entanglement item is scored as 0.20; when the difference between the diameter and slot column number is greater than 1, the cross-slot misconnection item is scored as 0.10. If the corresponding condition is not met, the score is 0. The four items are added together to obtain the reverse-deposit score, which ranges from 0 to 1. When the reverse deduction score reaches 0.60, the corresponding segment leaves the normal deduction path and enters the isolation channel. When the reverse deduction score is below 0.60 and both the left and right end slot retainer segments exist simultaneously, the un-deducted segment is pushed into the low-confidence boundary position. The low-confidence boundary position is the location for storing segments that have not reached the deduction threshold and have not triggered the fault condition. The low-confidence boundary position is not connected to the main propagation path and enters the residual isolation slot with the isolation channel. When the left end slot retainer segment exists and the right end slot retainer segment is missing, the left end slot retainer segment corresponding to the missing right end position enters the fault channel; when the right end slot retainer segment exists and the left end slot retainer segment is missing, the right end slot retainer segment corresponding to the missing left end position enters the fault channel. After the slot edge gap position enters the reverse deduction stripping branch, the gap status value directly participates in the gap entanglement term judgment. After the judgment is completed, the gap status value enters the isolation channel with the segment whose reverse deduction score reaches 0.60, enters the residual isolation slot with the segment in the low-confidence boundary position, and enters the fault channel with the missing end segment position.

[0046] The slot buckle untangling latching mechanism performs pre-merging splitting within the slot buckle attention layer. The inter-slot buckling pieces are connected to the main propagation path, while the fragments that are stripped by reverse buckling and fragments within the low-confidence boundary position are connected to the isolation channel. The missing pieces on the end side are connected to the fault channel. The latching merge entry only switches the propagation direction and does not rewrite the fragment vector content. The residual main slot, residual isolation slot, and residual fault slot respectively output the main residual, isolation residual, and fault residual, and retain the item anchor row number, caliber slot column number, flow sequence code, and gap status value.

[0047] Compared to the original EdgeNeXt network model's split-depth transposed attention structure, which uniformly performs transposed attention after splitting channels and then reconnects via ordinary residual connections, the slot-tightening attention layer decomposes the unified attention channel into slot-retaining branches, slot-inter-tightening branches, and reverse-tightening stripping branches. This ensures that continuous relationships within a single caliber, normal connection relationships between adjacent calibers, and abnormal connection relationships are separated before entering the residual path. Slot-interaction processing no longer relies solely on vector similarity but incorporates the anchor row number, flow sequence code, gap status value, and caliber slot column number difference into the connection score and reverse-tightening score calculation. This prevents highly similar segments with anchor misconnections, reverse flow connections, or gap entanglements from entering the main propagation path. After ordinary residual reconnection is replaced by residual main slots, residual isolation slots, and residual fault slots, normal connection, misaligned connection, and missing connection are output along different residual slots. The unwrapping correction layer can directly receive the already split caliber relationships, reducing false closures caused by premature mixing of multiple managed calibers in the unified attention path.

[0048] In this embodiment, S55 specifically includes: The slotted snap-locking mechanism is located at the confluence of the three branches of the slotted snap-locking layer, directly receiving the retained segments within the slot, the snap-locking segments between slots, the segments within the isolation channel, and the segments within the fault channel. The slotted snap-locking mechanism includes a snap-locking confluence surface, a snap-locking slot, a reverse snap-locking isolation slot, a fault retention slot, and a residual output slot. The snap-locking confluence surface processes segments between adjacent diameter slots under the same anchor position as a single unit; the snap-locking slot confirms normal snap-locking; the reverse snap-locking isolation slot handles the retention of misconnected, reverse-connected, and diameter-misaligned segments; the fault retention slot handles the locking of end-side defects and gap boundaries; and the residual output slot handles the segmented output of the main residual, isolation residual, and fault residual.

[0049] Before receiving segments at the latching confluence surface, the source of segments entering the mechanism is first confirmed by boundary verification. Inter-slot latching pieces are adjacent diameter receiving pieces with an inter-slot latching score of 0.80; segments within the isolation channel are misaligned segments with a reverse latching score of 0.60; segments within the fault channel are slot-retained pieces formed after the end-side segment is missing. Inter-slot latching pieces are introduced into the latching latch slot, segments within the isolation channel are introduced into the reverse latching isolation slot, and segments within the fault channel are introduced into the fault retention slot. Slot-retained pieces are attached to the corresponding diameter slot edge as end-side verification pieces. The end-side verification pieces retain a 128-dimensional segment vector, item anchor row number, diameter slot column number, flow sequence code, segment sequence number, and gap status value. The segment sequence number is the original segment arrangement number within the diameter fiber, continuously increasing from 1; a gap status value of 0 represents no gap, while values ​​of 1, 2, 3, and 4 represent object disconnection, action reversal, diameter loss, and cross-slot disconnection, respectively.

[0050] The latching slot processes one inter-slot latching piece at a time, clamping its left and right ends. The left end comes from the side with the smaller diameter slot number, and the right end comes from the side with the larger diameter slot number; the difference in diameter slot numbers between the left and right ends is 1. After the latching slot pulls the left and right end-side verification pieces, it performs a latch consistency check, which includes anchor alignment, end-side fit, smooth transition, and notch closure. When the anchor row numbers of the left end segment, right end segment, left end side check piece, and right end side check piece are all consistent, the anchor position matching item is 0.35; when the segment sequence number of the left end segment is consistent with that of the left end side check piece and the segment sequence number of the right end segment is consistent with that of the right end side check piece, the end side fitting item is 0.25; when the right end flow sequence code is not less than the left end flow sequence code, the flow sequence item is 0.25; when the gap status values ​​of the left end side check piece and the right end side check piece are both 0, the gap closure item is 0.15. Unsatisfied check items are 0. The four items are added together to obtain the latch consistency degree, with a value range of 0 to 1. The above weights are configured according to the importance order of item attribution, end side fitting, flow sequence, and gap closure, totaling 1. When the latch consistency degree reaches 0.85, the slot interlocking piece and the end side check piece are embedded to form a latching closed piece, and the latching closed piece is connected to the main propagation path.

[0051] When the latch consistency is below 0.85, the latching slot activation failure is branched off. The failure branch first checks the missing end-side calibration piece status. If the left end segment exists but the left end-side calibration piece is empty, the missing left end segment is transferred to the fault retention slot; if the right end segment exists but the right end-side calibration piece is empty, the missing right end segment is transferred to the fault retention slot. An empty slot is a slot edge position with no segment vector, no anchor row number, and a gap status value greater than 0. If the end-side calibration piece is not missing but the anchor position co-location item is 0, the corresponding segment is cut into the reverse latching isolation slot and marked as misconnected anchor; if the flow sequence item is 0, the corresponding segment is cut into the reverse latching isolation slot and marked as reverse flow; if the gap closure item is 0, the corresponding segment carrying the gap status value is cut into the reverse latching isolation slot. When multiple failure causes occur simultaneously, the following steps are performed: first, end-side missing judgment is performed; then, anchor misconnection judgment is performed; then, flow direction reverse connection judgment is performed; finally, gap entanglement judgment is performed. Segments that have already entered the fault retention trench will no longer enter the reverse isolation trench.

[0052] After the intercepting isolation slot receives the intercepted segment, it clamps the segment in the opposite direction along the original slot connection point and cuts off the return edge between the intercepted segment and the main propagation path. The return edge is the propagation connection from the latching slot to the main residual slot. After being cut off, the intercepted segment no longer enters the main residual. The intercepting isolation slot is used to press the intercepted segment into the inlet diameter misalignment cause mark. The cause mark uses a 3-bit binary code: anchor misalignment is recorded as 100, flow reversal is recorded as 010, and gap entanglement is recorded as 001. When multiple causes occur simultaneously, the corresponding bits are set to 1. The cause mark and the 128-dimensional vector of the intercepted segment are concatenated to form a 131-dimensional intermediate vector, which is then compressed back to 128 dimensions by a 131x128 linear compression matrix to form the intercepting isolation piece. The initial value of the linear compression matrix is ​​assigned using a normal distribution with a mean of 0 and a standard deviation of 0.02, and is updated through backpropagation during the training phase.

[0053] After verifying the missing segment of the fault-retaining groove end-side check piece, the missing segment is pressed down along the groove edge gap position, and the missing segment is superimposed on the same side as the assigned gap piece. Superimposed on the same side means that the missing segment and the assigned gap piece are located at the same diameter groove column boundary, and the item anchor row number is consistent. The assigned gap piece provides the gap type, gap occurrence sequence number, and item line position; the gap type adopts four states: object disconnection, action reversal, diameter missing, and cross-groove disconnection, coded as 1, 2, 3, and 4 respectively. The fault-retaining groove locks the disconnected inter-groove connection position as the fault boundary. The fault boundary is encoded as a 4-dimensional vector, consisting of the missing side code, the missing type normalization value, the missing occurrence sequence number normalization value, and the event anchor row number normalization value; the left missing side code is 1, and the right missing side code is 2; the missing type normalization value is obtained by dividing the missing type code by 4; the missing occurrence sequence number normalization value is obtained by dividing the missing occurrence sequence number by the total number of missing sides within the same event anchor, and the total number of missing sides within the same event anchor starts from 1; the event anchor row number normalization value is obtained by dividing the event anchor row number by the maximum event anchor row number within the current processing batch.

[0054] The attribution gap piece encoding is a 4-dimensional vector, composed of a gap type normalization value, a gap occurrence sequence normalization value, an event line position normalization value, and a caliber slot column number normalization value. The event line position normalization value is obtained by dividing the event line position by the largest segment sequence number within the same event anchor position, and the caliber slot column number normalization value is obtained by dividing the caliber slot column number by 8. The 128-dimensional segment vector of the gap end segment, the 4-dimensional fault boundary encoding, and the 4-dimensional attribution gap piece encoding are concatenated into a 136-dimensional intermediate vector, which is then compressed back to 128 dimensions using a 136x128 linear compression matrix to create the fault retention piece. The fault retention piece does not participate in inter-slot closure but only enters the residual fault slot; the 136x128 linear compression matrix is ​​initialized with a normal distribution with a mean of 0 and a standard deviation of 0.02, and is updated through backpropagation during the training phase.

[0055] The residual outflow groove receives the snap-fit ​​closure piece, the reverse snap-fit ​​isolation piece, and the fault retention piece. The snap-fit ​​closure piece is fed back into the residual main groove along the main propagation path and compressed into the main residual, which retains the normal caliber acceptance relationship; the reverse snap-fit ​​isolation piece is converged into the residual isolation groove and compressed into the isolation residual, which retains the causes of misconnection, reverse connection, and gap entanglement; the fault retention piece is converged into the residual fault groove and compressed into the fault residual, which retains the end-side missing, fault boundary, and attribution gap location. Neither the residual isolation groove nor the residual fault groove is fed back into the residual main groove. The three types of residuals carry the item anchor row number, caliber groove column number, flow sequence code, gap status value, and cause code, respectively, for the unwrapping correction layer to identify normal acceptance, misaligned acceptance, and missing acceptance.

[0056] Compared to the original EdgeNeXt network model where the split-depth transposed attention structure only performs unified attention aggregation on the feature channels after splitting and then re-adds them through the ordinary residual path, the improvement of the slot-lock untangling latch mechanism lies in directly modifying the attention convergence port and the residual return port: a latching latch slot, a reverse latching isolation slot, and a fault retention slot are added in the attention convergence port, so that the latching pieces between slots no longer directly enter the main propagation path based on similarity, but must undergo latch consistency verification through anchor position alignment, end-side fitting, flow connection, and gap closure; a residual main slot, a residual isolation slot, and a residual fault slot are added in the residual return port, so that normal receiving pieces, misaligned receiving pieces, and missing receiving pieces are pressed into different residual slots. Through the above structural modifications, misconnected segments with highly similar but inconsistent anchor positions, drifting segments with reversed flow directions, and segments that use other event boundaries to complete surface closure cannot be reinjected into the main residual along the ordinary residual path. Instead, they are intercepted by the reverse deduction isolation slot and enter the isolation residual carrying the cause code. Segments with missing ends are locked as fault boundaries by the fault retention slot and enter the fault residual. The above improvements change the original model's propagation logic of strong attention (aggregation) and direct residual addition, enabling normal deduction, abnormal deduction, and missing faults between multiple management standards in digital auditing to be hard-split within the model. This provides the untangling correction layer with isolated standard relationships, preventing misaligned standards from being treated as normal associations and continuing to spread.

[0057] In this embodiment, S6 specifically includes: The unwrapping correction layer is positioned after the slot buckle attention layer, receiving the main residual, isolation residual, and fault residual. The unwrapping correction layer includes a residual correction surface, a main residual reattachment unit, an isolation residual correction unit, a fault residual calibration unit, and an unwrapping state merging unit. The residual correction surface is configured with the main anchor edge, side anchor edge, and gap anchor edge according to the anchor row. The main residual is attached to the main anchor edge, the isolation residual is placed on the side anchor edge, and the fault residual is embedded in the gap anchor edge. Each residual fragment carries a 128-dimensional fragment vector, the anchor row number, the caliber slot column number, the flow sequence code, the version number, the gap status value, and the cause code. The alignment order first compares the anchor row number, then the caliber slot column number, and finally the flow sequence code. When the caliber slot column number is missing, the residual fragment remains within the correction boundary corresponding to its entry point and does not participate in cross-edge coverage.

[0058] The main residual back-paste unit performs back-paste correction along the main edge of the item anchor position. When the main residual matches the item anchor position row number of the main edge of the item anchor position, the anchor position back-paste item is set to 0.40; when the difference between the diameter slot column numbers of the left and right end segments is 1, the receiving back-paste item is set to 0.25; when the right end segment flow sequence code is not less than the left end segment flow sequence code, the sequence back-paste item is set to 0.20; when the gap status values ​​of the left and right end segments are both 0, the complete back-paste item is set to 0.15. Unmet items are set to 0. The four items are added together to obtain the main back-paste score, which ranges from 0 to 1. When it reaches 0.85, the slot receiving position back-pastes the adjacent diameter fiber end segment, and the segment in the slot returns to the original diameter slot and is made into a main closed piece; when it is lower than 0.85, the main residual is transferred to the anchor position side and marked as a closed low-confidence piece.

[0059] The isolation residual correction unit expands the isolation residual along the anchor side. After expansion, the isolation fragment first cuts off the loop edge between itself and the original main propagation path, and then retains the left segment number, right segment number, original item anchor row number, and original diameter slot column number. The candidate boundary is selected from the anchor side under the same item anchor row number; when the anchor misconnection cause exists, the candidate boundary is extended to the anchor side corresponding to the left and right segment item anchor row numbers. The preceding fragment is the fragment within the candidate boundary whose flow sequence code is less than the isolation fragment and whose value is closest; when there is no preceding fragment, the sequence continuation item and version continuity item are set to 0. The correction matching score is obtained by adding the anchor consistency item 0.40, the slot adjacent item 0.25, the sequence continuation item 0.20, and the version continuity item 0.15, with a value range of 0 to 1. When the score reaches 0.75, the isolated fragment is transferred to the candidate boundary and made into an attribution correction fragment; when the score is below 0.75, the isolated fragment is kept on the side of the anchor position and marked as a correction fragment to be verified.

[0060] The fault residual calibration unit processes the fault residual along the edge of the anchorage gap. The fault residual is aligned with the attribution gap piece. In the missing segment coding, the missing left segment is recorded as 1, the missing right segment as 2, and the missing both sides as 3. In the source chain sequence number, project initiation, budget, contract, procurement, acceptance, asset, rectification, and review are recorded as 1 to 8 respectively. In the break direction, forward break is recorded as 1, backward break as 2, and bidirectional break as 3. The fault residual, attribution gap piece, missing segment, gap source, and break direction are pressed into the same fault boundary to form a fault calibration piece. The fault boundary coding consists of the missing segment code, the gap source normalization value, the break direction code, and the item anchor row number normalization value. The gap source normalization value is obtained by dividing the source module code by 8, and the item anchor row number normalization value is obtained by dividing the item anchor row number by the maximum item anchor row number entering the unwrapping correction layer.

[0061] The untangling status merging unit gathers the main closed segment, attribution correction segment, verification correction segment, and fault calibration segment along the anchor row number of the event. The merging order first compares the anchor row number of the event, then compares the caliber slot column number, and finally compares the flow sequence code. The main closed segment carries a normal engagement status code 0, the attribution correction segment carries a misalignment correction status code 1, the verification correction segment carries a verification status code 2, and the fault calibration segment carries a fault status code 3. When multiple types of segments exist simultaneously within a single anchor row of an event, the main closed segment is fixed to the main edge of the anchor row, the attribution correction segment and the verification correction segment are fixed to the side edge of the anchor row, and the fault calibration segment is fixed to the edge of the anchor row gap. The boundaries of the three types of corrections do not overlap. After merging, a caliber unwrapping status is formed. The caliber unwrapping status includes the anchor row number, caliber slot column number, status code, cause code, main post score, correction matching score, fault boundary code, and original residual source. In the original residual source, the main residual is recorded as 1, the isolation residual is recorded as 2, and the fault residual is recorded as 3.

[0062] Compared to the original EdgeNeXt network model where residual output continues to propagate along a unified feature flow, the improvement of the unwrapping correction layer lies in transforming the post-propagation processing position of the residual into a three-sided correction structure. The main residual first verifies the normal continuity relationship at the main edge of the anchor position using the main return score; the isolation residual peels away the original slot continuity position at the anchor side and uses the correction matching score to determine the destination of the correction; the fault residual merges with the destination gap piece at the anchor gap edge to form the fault boundary. This improvement fixes normal closure, misalignment correction, and fault calibration within different correction boundaries, ensuring that the closure output layer receives an unwrapped caliber state with status code, cause code, correction score, and fault boundary, preventing caliber misalignment and continuity loss from being masked during ordinary residual propagation.

[0063] In this embodiment, S7 specifically includes: After the caliber unwrapping status of the closed output layer is reached, a closed processing surface is opened. The caliber unwrapping status includes the item anchor row number, caliber slot column number, status code, cause code, main feedback score, correction matching score, fault boundary code, and original residual source. The closed output layer includes a processing surface unfolding unit, a closed chain position verification unit, a verification object locking unit, a review chain position traction unit, and a processing bottom slot encapsulation unit. The closed processing surface uses the item anchor row as the basic unit. Each item anchor row is equipped with a closed chain position, a verification slot, and a review slot. The main closed piece with status code 0 enters the closed chain position, the assigned correction piece and the correction piece to be verified with status codes 1 and 2 enter the verification slot, and the fault calibration piece with status code 3 enters the review slot. The processing surface slotting order first compares the item anchor row number, then compares the status code, and finally compares the flow sequence code.

[0064] The closed-loop approval unit performs closure approval on the main closed segment. When the anchor row number of the inter-slot receiving position matches the item anchor row number of the segment within the slot, the anchor closure item is set to 0.35; when the difference between the slot column numbers of the left and right ends is 1, the slot continuity item is set to 0.25; when the right-end flow sequence code is not less than the left-end flow sequence code, the flow closure item is set to 0.20; when the main reply score reaches 0.85, the reply credibility item is set to 0.20. Unmet items are set to 0. The four items are added together to obtain the closure approval score, ranging from 0 to 1. When the score reaches 0.85, the inter-slot receiving position and the corresponding segment within the slot are locked together to form a continuous caliber chain; when it is below 0.85, the main closed segment is transferred to the verification slot and marked as a closed verification segment. The continuous caliber chain consists of the item anchor row number, the starting caliber slot column number, the ending caliber slot column number, the starting flow sequence code, and the ending flow sequence code, and is then pressed into the collaborative processing bottom slot.

[0065] The verification object locking unit receives the attribution correction piece, the correction piece to be verified, and the closed verification piece. The deviation distance is obtained by subtracting the original diameter slot number from the corrected diameter slot number and taking the absolute value, with a value range of 0 to 7. The closed verification piece does not have a corrected diameter slot number, so the deviation distance is recorded as 0, and the closed verification score is used as the source of confidence. The cause coding adopts a 3-bit binary form, with anchor misconnection as 100, flow direction reversal as 010, and gap involvement as 001; the number of cause setting 1s is the number of bits with a value of 1 in the code, with a value range of 0 to 3. The attribution correction piece and the correction piece to be verified use the reverse value of the correction matching score, and the closed verification piece uses the reverse value of the closed verification score. The reverse value is 1 minus the corresponding score. The verification priority value is obtained by dividing the deviation distance by 7 and multiplying it by 0.40, dividing the number of cause setting 1s by 3 and multiplying it by 0.30, and multiplying the confidence reverse value by 0.30, and then adding them together, with a value range of 0 to 1. When the verification priority value reaches 0.60, the collaborative verification object is attached to the corresponding item anchor position and enters the collaborative processing slot; when it is below 0.60, the collaborative verification object remains in the verification slot and is marked with a review mark. When the cause code is 100, the item anchor position is locked; when the cause code is 010, the flow sequence relationship is locked; when the cause code is 001, the gap data relationship is locked; when multiple cause codes are set to 1 simultaneously, the collaborative verification object retains the corresponding verification mark.

[0066] The review chain position traction unit receives the fault calibration piece and locates the missing receiving position from the fault boundary code. When the missing end segment code is 1, the left end segment is located; when it is 2, the right end segment is located; and when it is 3, both the left and right end segments are located. In the gap type code, object disconnection is recorded as 1, action reversal as 2, caliber missing as 3, and cross-gap disconnection as 4; the gap source code takes the value according to the source chain sequence number, and is recorded as 1 to 8 from project initiation to review. The attribution gap piece traction responsibility confirmation node is composed of the item anchor line number, gap source code, gap type code, and responsible entity code. When the responsible entity field is a personnel number, the personnel number is directly taken; when it is a department name, the department code within the platform is taken; when the responsible entity field is empty, the effective responsible entity whose intra-item flow sequence code is less than the current fault calibration piece and is the closest is taken. The effective responsible entity is the responsible entity field that can be matched with a unique number in the personnel table or department table; when it is still empty, the responsible entity code is recorded as 0 and enters the manual confirmation slot. The manual confirmation process records the anchor row number, fault calibration piece number, and missing field name for the item. The fault calibration piece number is formed by concatenating the item anchor row number, caliber slot column number, flow sequence code, and fault status code. The fault status code is fixed at 3. After manual confirmation, the responsible entity code is backfilled. The responsibility confirmation node number is formed by concatenating the item anchor row number, gap source code, gap type code, and responsible entity code, and is used to link with the rectification chain position and review mark.

[0067] The rectification chain is generated based on the missing information's location and includes the item anchor line number, missing segment code, break direction code, rectification priority, and review deadline marker. The rectification priority is obtained by dividing the gap type code by 4 and multiplying by 0.30, dividing the missing segment code by 3 and multiplying by 0.30, multiplying the normalized gap occurrence number by 0.20, and dividing the break direction code by 3 and multiplying by 0.20, then adding the results. The value ranges from 0 to 1. The normalized gap occurrence number is obtained by dividing the gap occurrence number by the largest gap occurrence number within the same item anchor line. When the rectification priority reaches 0.80, the review deadline is marked as urgent review, with a deadline of 3 calendar days; when it reaches 0.50 but is below 0.80, it is regular review, with a deadline of 7 calendar days; and when it is below 0.50, it is follow-up review, with a deadline of 15 calendar days. The rectification chain number is formed by concatenating the item anchor line number, missing segment code, break direction code, and review deadline marker. The review marker consists of the responsibility confirmation node number, rectification chain position number, review status code, and backfill sequence number. The initial value of the review status code is 0, which is changed to 1 after the review is completed, and changed to 2 when the review is returned. The backfill sequence number increments by 1 starting from the first review marker in the same anchor row. When any of the following missing states occur: missing responsibility confirmation node number, missing rectification chain position number, or missing anchor row number, the review marker will not backfill the anchor boundary of the item and will be transferred to the exception handling slot.

[0068] The collaborative processing involves closing the chain of processing segments within the verification and review segments. Each processing segment is a unit formed by linking a continuous caliber chain, a collaborative verification object, a responsibility confirmation node, a rectification chain position, and a review mark. The processing segment number is formed by concatenating the item anchor line number, processing type code, and flow sequence code; in the processing type code, continuous caliber chain is denoted as 0, collaborative verification object as 1, responsibility confirmation node as 2, rectification chain position as 3, and review mark as 4. The encapsulation order first compares the item anchor line number, then the processing type code, and finally the flow sequence code; processing segments with smaller values ​​are listed first. After encapsulation, the audit collaborative processing result is formed, including the item anchor line number, continuous caliber chain, collaborative verification object, responsibility confirmation node, rectification chain position, review mark, status code, and cause code. The handling segment with missing anchor row number is not included in the result encapsulation and is transferred to the exception handling slot; the exception handling slot records the missing field, the handling segment number and the original residual source. In the original residual source, the main residual is recorded as 1, the isolation residual is recorded as 2 and the fault residual is recorded as 3.

[0069] Compared to the original EdgeNeXt network model where the output layer only provides classification results, feature responses, or regression results, the improvement of the closed output layer lies in transforming the model's output into a closed structure oriented towards collaborative digital auditing. The closed output layer doesn't simply read the untangled state of the criteria and output risk labels; instead, it internally establishes closed chain positions, verification slots, review slots, and collaborative processing bottom slots. This locks normal closed relationships into continuous caliber chains, locks misalignment correction relationships into collaborative verification objects, and guides gap / missing relationships into responsibility confirmation nodes, rectification chain positions, and review markers. This improvement transforms the deep learning model's final output from a single identification result into an executable collaborative auditing structure. Auditors receive not just anomaly judgments, but closed results with verification objects, responsibility nodes, rectification chain positions, and review boundaries. This reduces the gap between model results and manual processing, improving the executability and traceability of digital auditing collaboration.

[0070] Example 1: To verify the feasibility of this invention in practice, it was applied to a digital audit management scenario for a collaborative agricultural research and industrialization project. In this scenario, the audit objects cover research funding expenditures, commissioned service contracts, equipment procurement, asset registration, results acceptance, problem rectification, and responsibility transfer records. Different business systems use different management standards for the same matters. The original audit system mainly relied on field matching, process node verification, and manual review to determine the completeness of data. This easily led to situations where the funding standards corresponded to the contract standards, but the procurement content, asset recording, acceptance materials, and rectification responsibilities could not be continuously closed. This resulted in misclassification, misrepresentation of matters, drifting standards, and gaps in responsibility hidden within a seemingly complete material chain.

[0071] In application, the system first extracts relevant records from the digital audit platform, including project initiation, budget execution, contract fulfillment, procurement acceptance, asset registration, rectification review, and responsibility transfer. It then meticulously analyzes and traces the execution items within these audit flow records, linking audit object traces with collaborative actions to form audit clues. These clues are then strung together to form a bundle of audit criteria. Subsequently, the system separates the audit criteria bundle into criteria fibers and presses them into a criteria slot matrix. After the criteria slot matrix enters the improved EdgeNeXt network model, the criteria slot embedding layer presses the criteria fibers into intra-slot segments and presses inter-slot boundary pieces into inter-slot receiving positions. The slot-clamping attention layer uses a slot-clamping untangling latching mechanism to process intra-slot segments, inter-slot receiving positions, and missing receiving segments, forming primary residuals, isolation residuals, and fault residuals. The untangling correction layer then re-attaches, corrects, and merges the three residuals, and finally, the closed output layer generates the audit collaborative processing results.

[0072] To ensure the comparability of the implementation data, a sample of audit items that had been archived and manually verified was selected. The sample included 820 audit items related to research project funding expenditures, procurement contract performance, asset registration and acceptance, and rectification closure, involving approximately 12,600 audit records. In the manual verification results, 176 items were confirmed to have multi-caliber coordination anomalies, while the remaining items were considered normal closure items. During the comparison process, the original rule comparison method, the ordinary deep learning audit identification method, and the method of this invention used the same original records, the same verification standards, and the same handling criteria. Indicators such as multi-caliber anomaly identification, fault location, shell company identification, drift identification, responsibility node location, and coordination efficiency were statistically analyzed. The statistical results are shown in Table 1 below.

[0073] Table 1 Comparison of the Collaborative Handling Effects of Digital Auditing of Agricultural Scientific Research Achievement Transformation Projects

[0074] As can be seen from the data in Table 1 above, the method of this invention significantly outperforms the original rule comparison method and the ordinary deep learning audit identification method in multiple audit collaboration evaluation indicators. The original rule comparison method mainly relies on field consistency, process nodes, and material integrity for judgment, with an accuracy rate of 78.4% for identifying abnormal calibers, an abnormal event recall rate of 71.6%, and a fault location accuracy rate of 68.9%, indicating that its ability to identify matters where surface fields correspond but internal calibers cannot be closed is insufficient. The ordinary deep learning audit identification method has improved its overall feature extraction capability, with the accuracy rate of identifying abnormal calibers increasing to 85.7% and the abnormal event recall rate increasing to 82.1%. However, the accuracy rates for identifying shell company transactions, identifying caliber drift, and locating responsibility nodes are still 74.9%, 78.3%, and 81.4%, respectively, indicating that simply encoding multi-source audit data uniformly for risk identification still easily mixes misaligned inheritance relationships into ordinary feature associations. This invention utilizes a bundle of event categories, a matrix of category slots, and an improved EdgeNeXt network model for multi-category untangling processing. This achieves an accuracy rate of 93.6% for category anomaly identification, 91.5% for anomaly event recall, 88.7% for fault location, 87.6% for shell company identification, 89.2% for category drift identification, and 90.8% for responsibility node location. Simultaneously, the average collaborative processing time is reduced from 4.6 days using the original rule comparison method to 2.4 days, and the manual review rejection rate is reduced to 9.7%. This demonstrates that this invention not only improves anomaly identification capabilities but also enhances the efficiency of collaborative audit processing.

[0075] This embodiment extracts, crimps, and threads the audit flow records of agricultural scientific research achievement transformation and industrial collaboration projects, compressing the object traces, caliber traces, and collaborative actions originally scattered across different business systems into traceable audit clues. These are further structured into caliber bundles and caliber slot matrices, enabling unified management of multiple management calibers under the same audit item. Based on this, the caliber slot embedding layer in the EdgeNeXt network model is improved to slot-position caliber fibers and slot boundary pieces. The slot-locking attention layer uses a slot-locking untangling latching mechanism to divert normal, misaligned, and missing relationships into main residuals, isolated residuals, and fault residuals, preventing multiple caliber relationships from mixing in a unified propagation path. Subsequently, the untangling correction layer performs back-paste, attribution correction, and fault labeling on the three residual paths. The closed output layer transforms the caliber untangling status into collaborative verification objects, responsibility confirmation nodes, rectification chain positions, and review markers. Therefore, this embodiment can form a clearer chain of identification and handling for hidden problems such as miscalculation, misrepresentation of matters, shifting of standards, and gaps in responsibility, thereby improving the accuracy, closure, and traceability of digital audit management collaboration.

[0076] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based digital audit management collaboration method, characterized in that, Includes the following steps: S1. By meticulously analyzing and verifying the execution details of the audit flow records, the traces and collaborative actions of the audited entities are linked together to form audit clue fragments, thus creating a set of audit clue fragments. S2. Audit clue fragments are strung together along the subject matter to the execution position. The scattered audit clue fragments are superimposed to the subject matter anchor position, and the belonging gap fragments are pressed into the subject matter anchor position boundary to form a subject matter scope bundle. S3. The item diameter bundle is stripped into diameter fibers, and the notched piece is retained at the item anchor boundary. Adjacent diameter fibers are snapped together along the flow direction to form a groove boundary piece, forming a diameter groove matrix. S4. The diameter slot matrix is ​​fed into the improved EdgeNeXt network model. The diameter slot embedding layer presses the diameter fiber into slot segments and presses the inter-slot boundary sheet into the inter-slot receiving position to form the diameter slot embedding state. S5. Pay attention to the groove insertion state of the groove bearing interface and embed the groove buckle untangling latching mechanism. The groove retains the support to lock the groove segment. The groove inter-clamping support will clamp the groove bearing position that is in the same direction as the anchor position into the main residual. The reverse clamping peeling support will transfer the groove bearing position that is not in the same direction to the isolation residual. The missing groove segment will be transferred to the fault residual, forming three residuals. S6. The unwrapping correction layer receives the three residuals. The main residual is back-attached to the event anchor position. The isolation residual is corrected for attribution. The fault residual and the attribution gap are merged and calibrated to form the caliber unwrapping state. S7. Close the output layer interface untangling status, guide the corresponding item anchor position along the untangling status, lock the collaborative verification object, responsibility confirmation node and rectification chain position, backfill the review mark, and form the audit collaborative handling result. 2.The deep learning-based digitalized audit management collaboration method of claim 1, wherein, S1 specifically includes: S11. Perform field regularization and empty space masking on the audit flow records, retain the record segments that point to the audit object, management scope and collaborative actions, and remove redundant record segments that are irrelevant to the attribution of the matter to form a record segment to be extracted. S12. The record segment to be extracted is extracted along the continuous items of the audit object. Adjacent record segments are continuously verified by the chain of attribution. When the source relationship and the direction of flow both belong to the same audit object, they are connected as the same item thread. When the chain of attribution is broken, it is marked as a broken thread piece. S13. The item wire is connected to the management scope and the indentation is calibrated for the scope name, scope source and scope change traces to obtain the item wire with scope indentation. The broken wire piece is pressed into the boundary of the corresponding item wire. S14. The item wire with caliber indentation is encapsulated in the same position as the cooperating action execution, and the cooperating action is attached to the corresponding item wire's transfer position to form an audit clue piece; S15. Audit clue fragments are arranged in order of their subject matter, and broken thread fragments retain boundary markers along with their corresponding audit clue fragments, and are collected into an audit clue fragment set.

3. The deep learning-based collaborative digital audit management method according to claim 1, characterized in that, S2 specifically includes: S21. Audit clue fragments are set up according to the matter, with the audit object trace falling into the object position, the caliber indentation falling into the caliber position, and the coordinated action falling into the action position, thus creating a clue fragment to be threaded. S22. The thread pieces to be threaded are threaded along the object position in the execution sequence. The thread pieces to be threaded that are continuous in the object position and connected by the action position are strung together to form a matter line. The thread pieces to be threaded that have inconsistent names of the caliber position retain the original caliber indentation and are arranged with the matter line. S23. The item line is pressed into the item anchor position, the continuous thread piece falls into the main layer of the anchor position, and the dispersed diameter thread piece is attached to the side layer of the anchor position along the original diameter indentation. The main layer of the anchor position and the side layer of the anchor position are embedded to form the anchor position layer. S24. The thread piece to be threaded is peeled off into a belonging gap piece when the object position is disconnected, the action position is reversed, and the caliber position is missing. The belonging gap piece is pressed into the boundary of the anchor position and snapped to the edge of the anchor layer. S25. The main anchor layer, side anchor layer, and belonging gap piece within the anchor position are bound together in the order of threading to form the anchor diameter bundle.

4. The deep learning-based collaborative digital audit management method according to claim 1, characterized in that, S3 specifically includes: S31. The bundle of matters is spread out along the anchor position of the matter as the surface to be stripped. The audit clue pieces within the bundle fall into the corresponding bundle path according to the indentation of the matters, and the belonging gap pieces are docked at the boundary of the anchor position of the matter. S32. Audit clue pieces in the same bundle within the fiber bundle to be stripped are attached end to end according to the order of transfer, with a notch embedded at the break point to occupy the space, and stripped into the diameter of the fiber carrying the anchor position and the order of transfer. S33. Fibers of the same diameter are pressed into the diameter groove according to the diameter indentation. Fibers of the same diameter in the same groove are arranged in sequence along the flow order. Fibers of different diameters across grooves are aligned with the anchor position along the direction to form a diameter groove row with notches. S34. Adjacent diameter slots extract end-side fibers along the flow direction. When the end-side fibers are connected, they are pressed together to form a slot boundary piece. When the end-side fibers are disconnected, the notch piece pulls the notch to occupy the corresponding slot edge. S35, the diameter slot array, the inter-slot boundary plates and the slot edge notch occupants are assembled into a diameter slot matrix according to the anchor positions.

5. The deep learning-based collaborative method for digital audit management according to claim 1, characterized in that, S4 specifically includes: S41. The improved EdgeNeXt network model includes a slot embedding layer, a slot attention layer, an unwrapping correction layer, and a closed output layer. S42. The diameter groove matrix enters the diameter groove embedding layer. The diameter groove embedding layer spreads the matrix groove surface along the anchor position. The diameter fiber falls into the groove crimping position. The groove boundary piece falls into the groove receiving position candidate area. The notch occupant is left at the groove edge notch position along with the assigned notch piece. S43. The diameter fibers in the crimping position in the groove are arranged adjacently along the flow direction. The diameter fibers of the same anchor position are crimped end to end to form a segment in the groove. The diameter fibers of the anchor position that are broken stop at the notch position on the side of the groove and fit with the notch piece. S44. The inter-slot boundary piece in the inter-slot receiving position candidate area is projected to the end side of the adjacent inter-slot segment. The inter-slot boundary piece with the end side anchor position is pressed into the inter-slot receiving position. The inter-slot boundary piece with the end side anchor position is transferred to the slot edge notch position. S45. The segments inside the slot, the receiving positions between slots, and the notches on the edge of the slot are encapsulated in the order of the diameter slots. The segments inside the slot carry the position of the diameter inside the slot, the receiving positions between slots carry the connection position of the adjacent slots, and the notches on the edge of the slot carry the position of the notch to which they belong, thus forming a diameter slot embedded state.

6. The deep learning-based collaborative method for digital audit management according to claim 1, characterized in that, S5 specifically includes: S51. After the groove fastening layer bearing interface diameter groove is embedded, the unified transfer attention channel is disassembled. The segment inside the groove remains in the groove along the diameter groove and is retained in the groove. The groove bearing position is transferred into the groove fastening support along the adjacent groove edge. The groove edge notch position is transferred into the reverse fastening peeling support along the anchor boundary to form a parallel groove fastening attention surface. S52. The slot retainer performs slot transposition on the slot segment. The slot segment only has contact crimping with adjacent segments in the same diameter slot. The original diameter of the segment and the anchor position are locked in the slot retainer, thus forming a slot retainer piece. S53. The inter-slot fastening support performs double-end alignment on the inter-slot receiving position. The two ends of the inter-slot receiving position pull the retaining piece in the adjacent diameter slot. The two-end segments of the anchor position and the flow direction receiving are pressed into the same fastening position to form the inter-slot fastening piece. S54. The reverse stripping support performs reverse verification on the slot connection position that has not entered the fastening position. The segments with opposite anchor positions, reverse flow direction, and belonging gap are removed from the normal fastening path and enter the isolation channel. The slot retaining piece without adjacent end segment is pressed into the fault channel. S55, the slot buckle unwinding latching mechanism performs latching and merging on the slot retainer, slot interlocking clip, isolation channel and fault channel within the slot buckle attention layer. The slot interlocking clip is incorporated into the main propagation path, and the isolation channel and fault channel are isolated from the main propagation path. S56. The main propagation path carries the inter-slot interlocking piece back to the residual main slot. The segment that is separated from the normal interlocking path is pressed into the residual isolation slot through the isolation channel. The slot retaining piece that lacks the adjacent end segment is pressed into the residual fault slot through the fault channel. The residual main slot, residual isolation slot and residual fault slot converge into three residual paths.

7. The deep learning-based collaborative digital audit management method according to claim 6, characterized in that, Specifically, S55 includes: S551, the slot buckle unwinding latch mechanism is set at the three junctions of the slot buckle attention layer. A latch junction surface is opened along the diameter attention surface. The latch junction surface receives the slot buckle piece, the segment in the isolation channel and the segment in the fault channel, and introduces the buckle latch groove, the reverse buckle isolation groove and the fault retention groove respectively. The retaining piece in the groove is attached to the edge of the corresponding diameter groove as the end side verification piece. S552, the latching slot first clamps the two end segments of the inter-slot latching piece, and then pulls the end side verification piece in the two diameter slots. When the end segment and the end side verification piece are aligned and connected in the same position on the anchor position and in the flow direction, the inter-slot latching piece is embedded with the end side verification piece to form a latching closed piece, and the latching closed piece is connected to the main propagation path. S553. Fragments that have not been made into snap-fit ​​closures do not enter the main propagation path. They are diverted along the cause of failure by the snap-fit ​​latch groove. Fragments with incorrect anchor positions and reverse connection in the flow direction are cut into the reverse snap isolation groove. Fragments with missing end-side verification pieces are transferred into the tomographic retention groove. Only snap-fit ​​closures are retained in the main propagation path. S554. After the cut-in segment is received by the reverse-clamping isolation groove, the cut-in segment is clamped in the opposite direction along the original groove receiving position, the reconnection relationship between the cut-in segment and the normal fastening edge is stripped off, and then the misalignment cause mark is pressed into the cut-in segment to make a reverse-clamping isolation plate that is separated from the main propagation path. S555. After verifying the missing segment of the fault retention groove receiving end side, press the missing end segment along the groove edge notch. The missing end segment and the assigned notch segment are stacked on the same side. The disconnected groove receiving position is locked as the fault boundary, and a fault retention segment that does not participate in the groove fastening is made. S556, the residual outflow channel sequentially receives the snap-fit ​​closing plate, the reverse snap-fit ​​isolation plate and the fault retention plate. The snap-fit ​​closing plate is reinjected into the residual main channel along the main propagation path and pressed into the main residual. The reverse snap-fit ​​isolation plate is gathered into the residual isolation channel and pressed into the isolation residual. The fault retention plate is gathered into the residual fault channel and pressed into the fault residual. The residual isolation channel and the residual fault channel cut off the reinjection connection to the residual main channel.

8. The deep learning-based collaborative method for digital audit management according to claim 1, characterized in that, S6 specifically includes: S61. After the unwrapping correction layer receives the three residuals, a residual correction surface is opened. The main residual is attached to the main edge of the anchor position, the isolation residual falls into the side edge of the anchor position, and the fault residual is embedded in the edge of the anchor position gap. The three types of residuals are aligned under the same anchor position. S62. The main residual is corrected by backing along the main edge of the anchor position. The normally fastened slot connection position is backed to the adjacent diameter fiber end section. The slot segment returns to its original position with the original diameter slot to form the main closed piece. S63. The isolation residual is spread out along the side of the anchor position to form an isolation segment. The isolation segment is first stripped from the original groove connection position, and then transferred to the corresponding anchor position boundary along the cause of the misalignment, and a correction piece is made. S64. The fault residual is attached to the edge of the anchorage gap and the fault calibration piece is pressed into the same fault boundary to make the fault calibration piece. S65, the main closure piece, the attribution correction piece, and the fault calibration piece are merged along the anchor position of the matter, and the results of normal snapping, attribution correction, and fault retention are merged into the caliber unwrapping state.

9. The deep learning-based collaborative method for digital audit management according to claim 1, characterized in that, Specifically, S7 includes: S71. After the unwrapping state of the closed output layer bearing interface, unfold the closed treatment surface along the anchor position. The main closed piece that has been reattached is incorporated into the closed chain position, and the correction piece is assigned to the traction verification slot. The fault calibration piece is assigned to the traction verification slot, and the treatment surface is divided into slots. S72. The closed chain position is approved along the anchor position of the matter. The slot receiving position after the backing is completed is locked with the corresponding slot segment to form a continuous diameter chain. The continuous diameter chain is pressed into the bottom slot of the collaborative treatment. S73. Verify the slot receiving the correction sheet, lock the collaborative verification object along the deviation position in the correction sheet, and attach the collaborative verification object to the corresponding item anchor position. S74. Verify the fault calibration plate of the trench, locate the missing acceptance position along the fault boundary, assign the missing plate to the responsibility confirmation node, and use the missing acceptance position to guide the rectification chain position. Verify the backfilling of the mark to the boundary of the anchor position. S75. The collaborative handling of the bottom trench closing chain position, verification trench position and review trench position handling pieces are packaged into audit collaborative handling results according to the item anchor position.

Images