An industrial mechanism generation and version evolution method based on artificial correction difference

Abstract

The application discloses an industrial mechanism generation and version evolution method based on artificial correction difference. The original output and the artificial final output of the industrial intelligent agent are compared on the factory side to form a structured correction difference and a context fingerprint, and are bound with a baseline mechanism snapshot used by a task as a desensitization backflow record; the backflow record is idempotently deduplicated, constraint-partitioned and mixed-similarly clustered to construct positive examples, negative examples and maturity vectors; representative conditions capable of distinguishing positive and negative examples are selected to generate a candidate mechanism containing conditions, conclusions, exceptions, evidence root abstracts and baseline relationships; immutable mechanism versions and mechanism snapshots are created after conflict detection and versioned playback. The original snapshot is continuously used for the started task, and the new snapshot is used for the new task; the new version is paused or rolled back when an abnormality occurs. The application can reduce the situation that incidental corrections are solidified into mechanisms, and keep the mechanism evolution process traceable and the task running results reproducible.

Description

Technical Field

[0001] This invention relates to the fields of industrial intelligent agents, knowledge engineering, rule engines, data version governance, and human-machine collaborative feedback processing. Specifically, it relates to a method, system, and computer program product for forming candidate mechanisms from the structured differences between the original output of industrial intelligent agents and the final output of humans, and for realizing the continuous evolution of mechanisms through evidence aggregation, conflict verification, and version release. Background Technology

[0002] Industrial intelligent agents can generate business outputs based on skills, rules, computational models, knowledge documents, and on-site data in scenarios such as equipment diagnostics, processing parameter recommendation, quality assessment, and R&D expense compliance. To control risks, some outputs will be approved, corrected, rejected, or transferred to manual processing by business personnel or domain experts. The results of manual processing include error correction information for the automatically generated outputs, which is an important source for subsequent improvement rules and mechanisms.

[0003] Existing technologies typically use human feedback for model training, statistical analysis, manual rule maintenance, or supplementing the knowledge base. Some solutions also offer rule suggestions based on the differences before and after editing, or organize feedback samples through clustering. These solutions can address some rule discovery issues, but when used in continuously operating industrial intelligent agents, they still struggle to simultaneously solve problems related to the reliability of feedback samples, data out-of-domain restrictions, verification of mechanistic conflicts, and consistency of operating versions.

[0004] First, a single manual correction might be an isolated case, a missing input, or a matter of personal preference. Generating formal rules based solely on frequency of occurrence or a single correction can easily solidify noise and exceptions into the underlying mechanism, leading to subsequent tasks being incorrectly intercepted or adjusted. Second, manual corrections lack a stable connection to the then-current mechanism version, digital employee version, task context, and evidence, making it difficult to subsequently reproduce the rule's origin and applicable conditions.

[0005] Secondly, original customer drawings, documents, process parameters, and identification information typically cannot leave the factory. Uploading complete cases to the cloud for clustering and rule extraction would compromise industrial data boundaries; uploading only natural language summaries makes reliable difference comparisons and playback verification difficult. Furthermore, candidate rules may exhibit overlapping conditions, contradictory conclusions, priority conflicts, circular dependencies, or unreachable branches compared to existing rules, necessitating technical verification before release.

[0006] There is also the issue of version evolution consistency. If a task that has already started reads a newly released rule during execution, it will result in the same task using different mechanisms before and after, making it impossible to reproduce the issue. If a new version experiences a high rate of manual correction or major errors after running, and cannot be paused or rolled back, it will also amplify the impact. Therefore, a technical solution is needed that can identify stable patterns from manual correction discrepancies, establish positive and negative evidence, verify candidate mechanisms, and form an immutable version while preserving the customer's original data at the factory. Summary of the Invention

[0007] The technical problems to be solved by this invention are: how to transform the manual corrections generated during the operation of industrial intelligent agents into computable, clusterable and verifiable difference data; how to distinguish between stable patterns and occasional special cases; how to form candidate mechanisms with evidence chains without uploading the original business data of customers; and how to make the new mechanism version only affect newly started tasks and can be paused or rolled back when the operation is abnormal.

[0008] To address the aforementioned technical problems, this invention provides a method for generating and evolving industrial mechanisms based on manually correcting differences.

[0009] The factory side obtains the original output reference, the final output reference, the manual action, the manual reason, the task context, and the baseline mechanism snapshot version locked by the task for the industrial intelligent agent task; compares the original output with the final output according to the predetermined output mode, generates a structured correction difference object containing field path, operation type, before modification summary, after modification value or summary, unit and action direction, and performs field standardization, unit normalization and sensitive field processing on the task context to generate a context fingerprint.

[0010] On the factory side, the corrected difference object, context fingerprint, task type, professional domain, baseline mechanism snapshot version, model confidence information, evidence reference, and event time are bound to form a reflow record. The original customer files and reversible identification information are retained in the factory; the reflow record sent to the mechanism library only contains anonymized features, local object references, and irreversible summaries. The reflow record forms an idempotent key using task identifier, execution round, and disposal version, and forms an event summary based on the normalized difference, context fingerprint, and baseline mechanism version to avoid duplicate submissions and increase sample weight.

[0011] Information increments are distributed among the returned records based on manual processing actions, whether effective corrections have occurred, and the automatically generated confidence status. Records that pass directly without field correction and with high confidence are used for confidence calibration statistics instead of directly entering candidate clusters; records that pass with low confidence, pass after correction, are returned for redo, are transferred to manual management, and are in the capability blind spot are entered into the candidate pool for extraction; special cases marked by experts, records that pass directly under similar conditions, and records that fail to replay can be used as counterexamples or exception samples.

[0012] The candidate data for extraction is first partitioned using hard constraints based on professional domain, task type, baseline mechanism family, and main version of the data pattern. Then, a mixed similarity is calculated based on the overlap of modified field paths, similarity of contextual features, consistency of human action direction, semantic similarity of causes, and overlap of evidence citations, forming candidate clusters within each partition. Hard-constrained partitioning prevents records with different professional meanings or incompatible data patterns from being merged due to textual similarity.

[0013] For each candidate cluster, a set of positive examples and a set of negative examples are constructed. A maturity vector is calculated, including factors such as source diversity, time span, consistency of corrections, completeness of evidence, degree of impact, model confidence bias, proportion of exceptions, and proportion of conflicts with existing mechanisms. The maturity determination strategy has a version identifier; when source diversity, completeness of evidence, or risk thresholds are not met, the candidate remains in a state of pending further evidence and no formal mechanism is generated.

[0014] Based on the distinguishing ability of standardized contextual features in positive and negative examples, representative conditional features capable of separating positive and negative examples under predetermined complexity constraints are selected to form structured conditional predicates. These are then combined with manually revised directions and the final output to form candidate conclusions or actions, generating exception predicates, scope of application, positive and negative sample citations, source citations, and root evidence summaries. Candidate mechanisms are also tagged with their relationship to existing mechanisms as addition, merging, substitution, conflict, or simply retained as cases.

[0015] Static analysis is performed on the candidate mechanism and target mechanism sets to detect condition coverage, contradictory conclusions, uncertain priorities, circular dependencies, and unreachable branches. Replay is performed in an isolated execution environment using a versioned test set, which includes at least positive examples, negative examples, boundary examples, and abnormal examples. The test set is weighted and evaluated for false positives, false negatives, and false negatives based on task risk.

[0016] When a candidate mechanism meets the release thresholds in terms of source, structure, conflict, testing, data boundaries, and professional review results, an immutable mechanism asset version is created. This version records at least the parent version, candidate cluster, maturity strategy version, root evidence summary, test report summary, and review results, and is parsed into a new mechanism snapshot. Tasks already started continue to use the original mechanism snapshot locked at startup; only newly started tasks use the new mechanism snapshot.

[0017] According to the mechanism version, the de-identified operation indicators returned from the factory side include hit, miss, manual correction, transfer to manual, downgrade, error, and latency. When the risk indicators of the new mechanism version exceed the threshold or a major defect occurs, the version is put into a paused or revoked state, and the new task is rolled back to the previous stable snapshot; the corresponding failure record is used as a counterexample or a new reflow candidate for subsequent evolution.

[0018] The present invention also provides an industrial mechanism evolution system, computer equipment, computer-readable storage medium, and computer program product for implementing the above methods.

[0019] This invention transforms manual corrections into structured differences and contextual fingerprints corresponding to baseline mechanism versions, allowing the original business data to remain on the factory side while the desensitized features can be used for clustering and verification; it identifies candidate mechanisms with stable applicability conditions through positive examples, negative examples, and maturity thresholds; and it makes the formation, use, and rollback of mechanism versions traceable through conflict detection, isolated playback, immutable versions, and task snapshot locking. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall system structure of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure for reflow recording and correction of differences.

[0022] Figure 3 This diagram illustrates incremental information routing and idempotent deduplication.

[0023] Figure 4 This is a flowchart of the overall process of the method of the present invention.

[0024] Figure 5 This is a schematic diagram of constrained clustering and maturity vectors.

[0025] Figure 6 This is a schematic diagram of the candidate mechanism and evidence summary chain.

[0026] Figure 7 This diagram illustrates conflict detection, replay verification, and version release.

[0027] Figure 8 This is a schematic diagram of the mechanism version's runtime feedback and rollback closed loop. Detailed Implementation

[0028] The embodiments of the present invention will be further described below with reference to the accompanying drawings. These embodiments are used to explain the present invention and not to limit the scope of protection. The industrial intelligent agent can be deployed on factory edge nodes, enterprise private clouds, or local servers isolated from the factory network; the mechanism library construction service can be deployed in the cloud or enterprise private clouds.

[0029] like Figure 1 As shown, the system includes a factory operation engine and coordinating console, a factory backflow processor, a mechanism library extraction pool, a candidate mechanism generator, a quality validator, a mechanism version manager, a mechanism snapshot publisher, and an operation feedback aggregator. The factory operation engine and coordinating console generate manual processing results; the backflow processor is responsible for generating structured differences, de-identifying, and calculating summaries before the data leaves the factory; the cloud component only processes the de-identified backflow records.

[0030] The mechanisms in this implementation include executable rules, review points, parameter constraints of the computational model, applicable conditions for knowledge retrieval, anomaly patterns, and constraints on data capability usage. Mechanism assets consist of an asset master record and an immutable version record. A mechanism snapshot identifier is written when a task starts; if the task subsequently retryes, resumes, or requires manual collaboration, the snapshot corresponding to this identifier is still read, and the execution process does not switch to a new version.

[0031] like Figure 2 As shown, the factory backflow processor performs a structured comparison of the original output and the final manual output according to the task output mode. Changes in scalar fields are recorded as replacement operations, changes in set fields are recorded as addition, deletion, or order adjustment operations, and object fields are recursively compared to the smallest editable path; for large files or binary objects, only the content summary, metadata, and pre-allowed extracted features are compared, and the file body is not uploaded.

[0032] Each correction difference record must include at least the difference identifier, field path, operation type, summary of the content before modification, the value after modification or its summary, data type, unit, correction direction, reason code, and impact type. The modified plaintext value is saved only when the corresponding field has passed the de-identification strategy and belongs to the feature that allows out-of-domain access; otherwise, only the irreversible summary and the factory-side object reference are saved.

[0033] Context normalization is performed according to a versioned context mode. Numeric fields are converted to a unified unit, enumerated values ​​are converted to standard codes, technical terms are normalized using a versioned thesaurus, and continuous numerical values ​​are mapped to ranges according to professional strategies. Organization, customer, equipment serial number, and personal identification fields are deleted, generalized, or replaced with irreversible identifiers. The normalized fields are sorted by field path, serialized using a unified character encoding and null value representation, and then the context fingerprint is calculated.

[0034] The reflow record includes at least the reflow identifier, idempotent key, event summary, tenant boundary identifier, professional domain, task type, baseline mechanism snapshot, baseline mechanism family, run release package, original output reference, final output reference, set of corrected differences, normalized context features, context fingerprint, manual action, cause code, model confidence information, evidence reference, event time and data schema version.

[0035] Idempotent keys can be calculated using the task identifier, execution round, manual processing result version, and source system identifier; event summaries can be calculated using the baseline mechanism snapshot, normalized correction difference set, context fingerprint, processing action, and event time bucket value. The receiving end sets a unique constraint on the idempotent key and checks for near-duplicate records based on the event summaries. Duplicate messages with the same idempotent key do not add new valid samples; records with the same event summaries but different message sources are only merged to receive audit information, without increasing cluster weights.

[0036] like Figure 3 As shown, the incremental information router determines the purpose of a record based on manual processing actions, whether the correction difference set is empty, the model confidence state, and the routing policy version. Records that pass directly with high confidence and have no effective correction are entered into the confidence calibration statistics; records that pass with low confidence, pass after correction, are returned for redo, are transferred to manual processing, or have key constraints supplemented manually are entered into the extraction pool; records marked by experts as exceptions, insufficient evidence, or unrepresentative are entered into the counterexample and exception pool.

[0037] Directly passing records is not always necessary. When a record's context is similar to that of a candidate cluster but the conclusion has not been manually corrected, it can be selected as a counterexample for that cluster to prevent the candidate criteria from being too broad. Low-confidence records can also be used to detect whether the automated system is overly conservative; high-confidence records that have been significantly corrected are used to detect overconfidence and form the model confidence bias feature.

[0038] After deduplication, the records are first partitioned according to a hard partition key. The hard partition key consists of at least a professional domain, task type, baseline mechanism family, and major version of the data model; two baseline mechanism versions are only allowed to enter the same partition if they are explicitly marked as comparable in the compatibility table, otherwise cross-version clustering is not allowed.

[0039] Within the same partition, the overlap of corrected field paths, contextual feature similarity, consistency of human action direction, causal similarity, and evidence citation overlap are calculated separately. Each component is normalized to between zero and one, and then the mixed similarity is obtained according to the weights of the clustering strategy version records. The overlap of corrected field paths can be expressed as the ratio of intersection to union; the proportion of consistent terms is used for enumerated contexts, and the distance function after dimension normalization is used for numerical contexts. For security or compliance tasks, the clustering strategy increases the weights of key field paths and action directions.

[0040] This embodiment employs density clustering: mixed similarity is converted to distance, and clusters are generated according to the neighborhood threshold and minimum sample size recorded in the clustering strategy version. Records that do not meet the minimum sample size are marked as outliers. Other implementations may use hierarchical clustering or incremental clustering. Each clustering session saves the strategy version, member event summary, representative samples, boundary samples, outliers, and generation time; a single event cannot directly form a publishable candidate.

[0041] like Figure 5As shown, maturity vectors are calculated for candidate clusters. Source diversity can be formed by de-identification counts from different tasks, time windows, different human processors, or different organizations; time span is the interval between the earliest and latest events; correction consistency measures the degree of consistency between correction directions and results in similar contexts; evidence completeness measures the completeness of output summaries, rationales, mechanistic citations, and evidence citations; impact level comes from the level of monetary, quality, security, compliance, or delivery risk.

[0042] The maturity vector also includes model confidence bias, exception ratio, and conflict ratio. The exception ratio equals the proportion of records marked as special cases by experts or similar to candidate conditions but requiring no correction; the conflict ratio equals the proportion of test samples where candidate conclusions contradict the current set of mechanisms. The versioned maturity strategy specifies hard thresholds and combined scores; in cases of high-risk conflicts, unknown sources, or data exceeding limits, the hard threshold is not offset by the combined score.

[0043] The set of positive examples consists of records that support the same direction of correction and meet the evidentiary threshold; the set of negative examples may include records that pass directly in similar contexts, exceptions confirmed by experts, valid manual actions contrary to candidate conclusions, and records of failed sandbox replays. Each positive and negative example retains an event summary, contextual features, baseline mechanism version, and source type.

[0044] Representative criteria selection follows a step-by-step screening process. The initial criterion set is empty. For each candidate feature, the changes in positive example coverage, negative example exclusion rate, and high-risk false positive rate before and after adding the feature are calculated, and the condition complexity penalty is deducted. The feature with the largest net gain is then added to the criterion set. The process stops when the net gain of a new feature falls below a threshold, the number of criteria reaches the upper limit, or the positive example coverage reaches the target. Fields with mandatory professional constraints are pre-added as fixed criteria.

[0045] Candidate mechanism objects such as Figure 6 As shown, it should include at least the candidate identifier, structured conditional expression, conclusion or action, scope of application, exception conditions, positive example citations, negative example citations, source citations, relationship type with baseline mechanism, clustering strategy version, maturity strategy version, root evidence summary, and candidate content summary. The structured conditional expression uses restricted operators, explicit data types, and units. Natural language descriptions are only for review purposes and are not the sole executable content.

[0046] Root evidence summaries can be obtained by constructing a summary tree by arranging member event summaries in a stable order, or by using a sequenced summary chain. When member evidence, sets of positive and negative examples, candidate conditions, or candidate conclusions change, the root evidence summary and candidate content summary change synchronously. Professional review records and playback test reports both reference candidate content summaries to ensure that the reviewed, tested, and published content is the same candidate.

[0047] The relationship between candidate mechanisms and existing mechanisms is determined through condition coverage and conclusion comparison. Candidate conditions that do not largely overlap with existing conditions are marked as new; candidates that supplement existing conditions or provide exceptions are marked as merged; candidates that modify old conclusions within the same scope of application are marked as substitutes; two conclusions that cannot be simultaneously valid under overlapping conditions are marked as conflicting; and conclusions that can only explain a single case and cannot form stable conditions are marked as cases.

[0048] like Figure 7 As shown, the static analyzer adds candidate mechanisms to the temporary dependency graph corresponding to the version of the target mechanism set, detecting conditional intersections, priorities, mutual exclusions, circular dependencies, and unreachable branches. For the rule decision table, it also checks for condition coverage and condition gaps in each row. High-risk conflicts that fail to determine priority block releases.

[0049] Replay validation uses a versioned test set. Test samples include candidate positive examples, negative examples, boundary values, anomalous inputs, existing rule regression samples, and manually constructed safe samples. The validator executes candidate rules and the target mechanism set in an isolated environment, recording rule hits, outputs, execution latency, and errors. A risk-weighted error metric assigns a higher weight to falsely released high-risk samples than to falsely blocked ordinary samples.

[0050] The release access control must include at least the following: valid source and usage authorization; complete conditions, conclusions, exceptions, and data types; no unresolved high-level conflicts; key tests meeting thresholds; no original customer data, credentials, or reversible identity information; valid professional audit results; and frozen asset versions, test sets, and policy versions referenced. Model-generated candidate descriptions or draft conditions must pass through the same access control and cannot be directly converted into released mechanisms.

[0051] After access is granted, the version manager creates a mechanism asset version record, saving the asset identifier, version number, parent version identifier, candidate identifier, root evidence summary, test report summary, professional review result, scope of effect, status, and content summary. Released versions cannot be modified in situ; any content changes will create a new version record.

[0052] The publisher parses the mechanism set, knowledge space, data capability contract, and model dependencies to form a mechanism snapshot package; the snapshot list records each asset version and its content summary. New tasks from digital employees are bound to the current recommended snapshot upon startup; tasks that have already started, are being retried, are being restored, or are in a state of human collaboration continue to use the mechanism snapshot identifier written when the task started.

[0053] Example 1 illustrates the evolution of the compliance mechanism for R&D expense collection. In multiple tasks, the automated system would include certain expenses in R&D expenditures when project-related evidence was lacking, leading the finance manager to repeatedly modify the entries to require supplementary documentation or remove them altogether. The feedback processor only uploads the expense category, the existence of evidence type, the project-related status, the modification action, and a summary; it does not upload payroll details or vouchers. The system generates a candidate review rule that "expenses that meet the predetermined category but lack project-related evidence should not be directly collected," and verifies this rule using positive and negative examples from different months, different projects, and direct samples.

[0054] Example 2 illustrates the evolution of the equipment anomaly diagnosis mechanism. Digital operators repeatedly identified vibration characteristics as bearing failure under high-speed conditions. Experts, combining load, temperature, and maintenance records, revised the conclusion to process resonance. The system generates candidate conditions based on vibration frequency band, speed range, load range, and correction direction, and uses cases with similar frequency bands but indeed bearing failures under low load as counterexamples, thus generating a diagnostic mechanism with exceptional conditions.

[0055] Example 3 illustrates the evolution of the machining parameter recommendation mechanism. For the same material and tool family, multiple reviewers reduce the feed rate only under conditions of deep cavity and long overhang. The system selects to retain the depth-to-diameter ratio and overhang ratio based on representative conditions, and removes features such as work group, order, and equipment number that have low distinguishability between positive and negative examples; candidate rules are released as new versions after passing simulation and historical case playback.

[0056] like Figure 8 As shown, the factory operation service aggregates hit counts, miss counts, manual correction rate, manual conversion rate, degradation rate, error category, and latency quantiles by mechanism version, and only sends back anonymized statistics and version identifiers. When a new version exceeds the risk threshold within the observation window, the version manager generates a pause or rollback event, causing subsequent new tasks to revert to the previous stable snapshot; during factory network outages, the locally activated stable snapshot continues to be used.

[0057] Pausing or revoking does not delete historical versions. Historical tasks can still be reproduced based on the version, event summary, root evidence summary, and test report to determine the original judgment. Failure samples that led to the pause are anonymized and used as counterexamples or new re-entry candidates, entering the next candidate cluster to form a traceable version evolution chain.

[0058] Each module of this invention can be implemented by a computer program executed by one or more processors. Object identity, version, status, and reference relationships are stored in a relational database; original files or large objects can be stored in object storage; full-text indexes and vector indexes are used for retrieval and can be reconstructed from authoritative data.

Claims

1. A method for generating and evolving industrial mechanisms based on manually correcting differences. A method for generating and evolving industrial mechanisms based on manually corrected differences, characterized in that: include: Obtain the raw output of the industrial intelligent agent task, the final output after manual correction, the manual action, the task context, the evidence references related to the manual action, and the baseline mechanism snapshot identifier used by the task; The original output and the final output are compared according to a predetermined output pattern to generate a correction difference record. The correction difference record includes the correction field path, operation type, content summary before modification, and content or summary after modification. The task context is standardized and sensitive information is processed, and the processed task context, correction difference records, manual actions, evidence citations and baseline mechanism snapshot identifiers are associated. Related data are deduplicated and grouped according to professional domain, task type, baseline mechanism family, and data pattern version; within the same group, the similarity between records is calculated based on the corrected field path, standardized task context, manual handling action, and evidence citation to form candidate sample groups; Based on the candidate sample group, construct a set of positive examples and a set of negative examples. Based on the degree of separation between positive examples and negative examples in the task context features and the degree of evidence completeness, determine the conditional features used to distinguish the set of positive examples and the set of negative examples. Candidate mechanism rules are generated based on the conditional features and the direction of manual correction. The candidate mechanism rules include conditional expressions, conclusions or actions, exception conditions, positive and negative sample citations, and evidence summaries. The candidate mechanism rules are checked against the existing mechanism rules for condition coverage, conclusion conflict, priority, and dependency, and replay tests are performed in an isolated execution environment using a test set with version identifiers. When the professional review is approved and the playback test results meet the preset release conditions, a new mechanism asset version is created and a new mechanism snapshot is written; when the task starts, the mechanism snapshot identifier used is recorded so that the already started tasks continue to use the mechanism snapshot recorded at the time of startup, and the newly started tasks use the new mechanism snapshot. Summarize operational metrics according to the mechanism asset version; when the operational metrics meet the preset rollback conditions, set the corresponding mechanism asset version to a paused or canceled state, and make subsequent newly started tasks use the rollback target mechanism snapshot.

2. The method according to claim 1, characterized in that, The field standardization and sensitive information processing of the task context include: The fields in the task context are standardized, units are normalized, and sensitive fields are processed according to the context mode with version identifiers. The processed fields are then serialized according to a fixed field order to generate a context fingerprint. The original customer files, credentials, equipment identification and reversible identity information are kept on the factory side, and only correction difference records, standardized task context features, local object references and irreversible summaries are sent to the mechanism library; When the modified content does not conform to the preset data out-of-domain rules, the irreversible digest of the modified content is saved in the correction difference record, but the plaintext value of the modified content is not saved.

3. The method according to claim 2, characterized in that, The process before forming the candidate sample group also includes: Records that pass directly without valid field correction and whose model confidence reaches the preset threshold are used for confidence calibration and are not considered positive examples; records that pass with model confidence below the preset threshold, records that pass after correction, records that are returned for reprocessing, or records that are transferred to manual processing are considered candidate records; records marked as exceptions by professionals, records that pass directly in similar task contexts, or records that fail during replay testing are considered negative examples or exception samples. An idempotent key is generated based on the task identifier, execution round, and manual handling version, and an event summary is calculated based on the correction difference record, context fingerprint, and baseline mechanism snapshot identifier. For records with the same idempotent key, only one valid record is retained. Records with the same event summary are treated as one record when calculating the sample size and sample weight.

4. The method according to claim 1, characterized in that, Forming the candidate sample group includes: Only records that are compatible with the professional domain, task type, baseline mechanism family, and data model version are grouped into the same group; Within the group, the similarity between records is calculated based on at least four of the following: overlap of the modified field path set, similarity of standardized task context features, consistency of human action direction, semantic similarity of human correction reasons, and overlap of evidence citations. Candidate sample groups are formed according to the weights and thresholds specified by the clustering parameters with version identifiers, and records with fewer than the preset minimum number of samples are marked as outliers.

5. The method according to claim 1, characterized in that, Constructing the set of positive examples and the set of negative examples, and determining the conditional features, includes: Candidate records that support the same direction of manual correction and whose evidence completeness rate reaches a preset threshold will be used as positive examples. Records with similar task contexts to the positive examples but do not require manual correction, exception records confirmed by professionals, effective disposal records that contradict the candidate conclusions, and replay test failure records in isolated execution environments will be used as negative examples. The number of sample sources, sample time span, consistency rate of manual correction direction, evidence completeness rate, proportion of exception samples, model confidence bias, and proportion of conflict with existing mechanism rules are calculated respectively, and the calculation results are compared with the judgment threshold with version identifier; if any mandatory threshold is not reached, the generation of publishable candidate mechanism rules is stopped. Using positive example coverage and negative example exclusion rate as evaluation indicators, task context features or feature intervals are selected one by one until the addition of new features causes the increment of the evaluation indicators to be lower than a preset threshold, the number of conditions to reach a preset upper limit, or the positive example coverage to reach a preset target value.

6. The method according to claim 3, characterized in that, Generating the evidence summary and checking the relationship between the candidate mechanism rules and existing mechanism rules includes: The event summaries in the positive and negative example sets are arranged in a predetermined order, and an evidence summary is calculated from the arranged event summaries; The professional review results and replay test reports are associated with the content summary of the candidate mechanism rules; when the set of positive examples, the set of negative examples, the conditional expression, or the conclusion changes, the professional review and replay test are performed again. Based on the condition coverage relationship and conclusion compatibility relationship between the candidate mechanism rules and the existing mechanism rules, determine the processing method for adding, merging or replacing versions; when the applicable conditions of the two overlap, the conclusions cannot be simultaneously valid and the execution order cannot be determined according to priority, set a conflict flag and stop publishing the candidate mechanism rules.

7. The method according to claim 1, characterized in that, The versioned test set includes at least positive examples, negative examples, boundary examples, outlier examples, and existing mechanism regression samples; The replay test records the hit results, output results, execution errors, and execution latency of the candidate mechanism rules in an isolated execution environment; Calculate the error pass rate and error block rate based on the number of error passes, the number of error blocks, and the risk weight of the corresponding task, respectively. If the applicable conditions of the candidate mechanism rule overlap with those of the existing mechanism rule, the conclusions cannot be simultaneously valid and the execution order cannot be determined according to priority, any specified test indicator fails to reach the release threshold, the candidate mechanism rule contains data that is prohibited from being output by the preset data out-of-domain rule, or the professional review conclusion is not in the pass state, then the preset release conditions are determined to be unmet.

8. The method according to claim 1, characterized in that, The mechanism asset version shall at least record the parent version identifier, candidate sample group identifier, clustering parameter version, judgment threshold version, evidence summary, test report summary, professional review result and content summary; The published mechanistic asset version remains unchanged; when the condition expression, conclusion, action, exception condition, or evidence changes, a new mechanistic asset version is created. When a task is started, the mechanism snapshot identifier is written. During subsequent retries, recovery, or manual collaboration, the task still loads the mechanism asset version based on the mechanism snapshot identifier, instead of switching to the new mechanism snapshot after release.

9. The method according to claim 1, characterized in that, The operational metrics include at least one of the following: number of hits, number of misses, number of manual corrections, number of times the process was transferred to manual handling, number of times the process was downgraded, number of execution errors, and execution latency. When the error rate, manual correction rate, or rate of transfer to manual processing of the new mechanism asset version exceeds the corresponding threshold, or when a preset type of execution error occurs, the new mechanism asset version is put into a paused or canceled state, so that the newly started tasks use the mechanism snapshot of the most recent corresponding running indicator that has not exceeded the threshold and is not in a paused or canceled state, while retaining the reference of the new mechanism asset version to the already started tasks. The records of tasks that resulted in suspension or cancellation will be used as counterexamples or new candidate records for the formation of candidate sample groups in the future.

10. A system for generating and evolving industrial mechanisms based on manually corrected differences, comprising a processor, a memory, and a database, wherein the memory stores a computer program executable by the processor, characterized in that... When the processor executes the computer program, the system: The system acquires the original output of the industrial intelligent agent task, the final output after manual correction, the manual actions, and the evidence references related to the manual actions. It generates correction difference records, performs field standardization and sensitive information processing on the task context, and associates the processed task context, correction difference records, manual actions, evidence references, and baseline mechanism snapshot identifiers. The related data is deduplicated and grouped. Within the same group, the similarity between records is calculated based on the corrected field path, the standardized task context, the manual handling action, and the evidence reference to form a candidate sample group. A set of positive examples and a set of negative examples are constructed based on the candidate sample group. Conditional features are determined based on the degree of separation between positive and negative examples in the task context features and the completeness of evidence. Candidate mechanism rules are generated based on the conditional features and the direction of manual correction. The candidate mechanism rules are checked against the existing mechanism rules for condition coverage, conclusion conflict, priority, and dependency, and replay tests are performed in an isolated execution environment using a test set with version identifiers. When the professional review is approved and the playback test results meet the preset release conditions, a new mechanism asset version is created and a new mechanism snapshot is written, so that the already started tasks continue to use the mechanism snapshot recorded at startup, and the newly started tasks use the new mechanism snapshot. The system summarizes the operating metrics according to the mechanism asset version, and when the operating metrics meet the preset rollback conditions, the corresponding mechanism asset version is set to a paused or canceled state, so that subsequent newly started tasks use the rollback target mechanism snapshot.

Images