A method for review feedback and process rule evolution for spray drying abnormal candidate events

Abstract

The application discloses a kind of spray drying abnormal candidate event-oriented review feedback and process rule evolution method, it is related to industrial process review and process knowledge evolution technical field, receives abnormal candidate event: receives spray drying abnormal candidate event;Generation review task: according to the abnormal candidate event generation review task, the review task is divided into problem review and production review;When multiple the abnormal candidate event belongs to same equipment, same candidate type and same rule source and time interval satisfies merging condition, merge to same event thread.The application realizes the whole life cycle management of abnormal event, artificial experience structured sedimentation and the continuous evolution of process rule safe and auditable, effectively reduces false alarm rate, improves the abnormal identification ability and knowledge reuse level of spray drying production.

Description

Technical Field

[0001] This invention relates to the field of industrial process retrospective and process knowledge evolution technology, specifically to a retrospective feedback and process rule evolution method for candidate abnormal events in spray drying. Background Technology

[0002] Spray drying production sites are characterized by strong coupling, strong lag, and multiple disturbances. Traditional automated control and single-point threshold alarms can handle some limit exceedance issues, but it is difficult to continuously accumulate each candidate warning, manual judgment, on-site handling, false alarm cause, and production review into reusable process knowledge.

[0003] In actual production, on-site personnel often rely on experience to determine whether an anomaly is valid. For example, an increase in outlet air temperature may be due to insufficient material supply, or it may be caused by a synchronous increase in inlet air temperature, raw material switching, changes in liquid temperature, changes in the material supply stage, or the state of the drying tower. If the system only records alarms without recording on-site confirmation, counter-evidence, and review conclusions, the rules will be difficult to continuously optimize, and false alarms and missed alarms will recur.

[0004] Existing industrial anomaly detection solutions typically focus on model detection, alarm output, or control optimization, with less attention paid to the subsequent lifecycle of anomaly candidate events on the production site: how candidate events are manually confirmed, how false alarms are archived, how problem reviews are generated, how similar experiences are merged, and how they are entered into the formal rule base after playback and manual approval.

[0005] Therefore, a method for review feedback and process rule evolution for abnormal candidate events is needed to connect candidate warnings, manual feedback, problem review, production review, rule suggestions, source playback, accidental damage inspection, manual approval and version management into an auditable closed loop. Summary of the Invention

[0006] This invention aims to solve at least one of the technical problems existing in the prior art. To this end, this invention proposes a method for review and feedback and process rule evolution for candidate abnormal events in spray drying. This method realizes full life cycle management of abnormal events, structured accumulation of human experience, and continuous evolution of process rules that are safe and auditable. It effectively reduces the false alarm rate and improves the ability to identify abnormalities and the level of knowledge reuse in spray drying production.

[0007] To achieve the above objectives, a first aspect of the present invention provides a method for post-mortem feedback and process rule evolution for candidate anomalies in spray drying, comprising:

[0008] Step S1: Receive abnormal candidate events. The system receives spray drying abnormal candidate events generated by abnormal candidate identification signals. The abnormal candidate events include event number, event thread number, occurrence time, associated equipment, associated variables, current operating stage, trend evidence, gating result, candidate level, on-site inspection items, and non-final fault characterization markers. The abnormal candidate events are not final fault conclusions and are only used as evidence entry points for review, feedback, and rule evolution processes.

[0009] Step S2: Generate a review task. The system generates a review task based on the candidate abnormal events. The review task is divided into two categories: problem review and production review. The problem review targets a single candidate event or event thread, while the production review targets a complete production cycle, summarizing the stage switching, event distribution, recovery status, false alarm statistics, and rule suggestions within the cycle. When multiple candidate abnormal events belong to the same device, the same candidate type, and the same rule source, and the time interval meets the preset or adaptive merging conditions, the system merges them into the same event thread. The event thread records the start time, most recent trigger time, duration, number of reminders, feedback status, and recovery status.

[0010] Step S3: Read time-series evidence and generate a deterministic fact summary. After the review task is generated, the system reads the PLC time-series evidence, process feature blocks, process interpretation relationships, gate records, and historical similar events within the time window corresponding to the abnormal candidate event. Before entering manual review, the system first generates a deterministic fact summary. For material supply-related candidate events, the fact summary includes at least the inlet air temperature, outlet air temperature, tower pressure, feed pump setpoint and feedback value within the same process feature block or an alignable time window, and marks the trend correspondence of inlet and outlet air temperatures. The deterministic fact summary is used to avoid the review only repeating the alarm text and ignoring the evidence correlation between the original parameters.

[0011] Step S4: Receive and normalize manual feedback. The system receives manual feedback from on-site operation pages, enterprise internal messaging terminals, management backend pages, or manual input. Feedback types include confirmed receipt, confirmed anomaly, obvious false alarm, processed, continued observation, uncertain, and supplementary on-site facts. The system normalizes manual feedback into fixed feedback types, standard object names, and standardized semantics. The imperative statements in the manual feedback are only stored in the ledger as factual descriptions, do not trigger automatic control actions, do not bypass the manual approval process, and do not automatically publish rules. The manual feedback is written to an independent feedback ledger, does not modify the PLC original sampling data table, and the feedback subject is recorded as a role-based account identifier.

[0012] Step S5: Generate a structured debriefing record. The system generates a structured debriefing record based on the aforementioned candidate anomalies, PLC timing evidence, process feature blocks, human feedback, and supplementary facts from the field. The structured debriefing record includes a debriefing number, source candidate event number, debriefing type, list of factual evidence, timeline, list of candidate explanations, list of counter-evidence, list of missing variables, human feedback record, debriefing conclusion, rule impact items, and debriefing status. The debriefing status includes draft, awaiting human review, confirmed, archived, rejected, and replaced by a subsequent version.

[0013] Step S6: False Alarm Archiving and Sample Accumulation. When the abnormal candidate event is manually marked as an obvious false alarm, the system generates a false alarm archive record, recording the cause of the false alarm, the corresponding source rule, the applicable stage, the evidence summary, and the manual remarks. A single false alarm does not directly block the corresponding rule, but only generates a false alarm candidate sample. When similar false alarms meet the preset repeated triggering conditions under the same operating stage and the same process interpretation relationship, the system generates a rule downgrade or suppression suggestion, which is then entered into the subsequent verification and approval process.

[0014] Step S7: Generate rule evolution suggestions and extract rule signatures. The system generates process rule evolution suggestions based on the structured review record. Suggestion types include new rules, modified rules, downgraded rules, suppression rules, archived rules, and unchanged rules. The system extracts rule signatures to avoid duplicate rule suggestions. The rule signature includes equipment object, result variable, explanatory variable, applicable stage, trend direction, gating action, and inspection item.

[0015] Step S8: Rule Verification Gate Verification Rules Recommendation: Before official release, the system should perform at least three verifications:

[0016] Source replay: Apply the suggested rules to the core time window corresponding to its source replay to verify whether the rules can hit the original abnormal event or correctly degrade the original false alarm;

[0017] Historical false alarm check: Apply the suggested rules to the historical production cycle dataset to check whether the rules will increase false alarms and whether they will cause false hits during the start-up phase, water operation phase, raw material switching phase, material shutdown and cooling phase, or drying tower phase.

[0018] Phase Adaptation Check: Check whether the applicable phase setting of the rule is consistent with the operation phase in the source review scenario and historical production cycle; if any verification fails, the rule is recommended to be returned for revision, downgraded to an observation experience item, or retained as a non-executable prompt message.

[0019] Step S9: After the rule evolution proposal is manually approved and a manual approval record is generated, it is entered into the official rule base. The published rule evolution proposal carries a version identifier, generates versioned rules, and retains all historical versions to form an auditable rule evolution chain. The approval record includes the approval entity identifier, approval time, approval conclusion, approval remarks, and verification report reference. The published rules carry a unique version identifier. When rules are revised, deactivated, archived, or restored from old versions, historical versions are retained in their entirety without overwriting the historical audit chain. The records in the versioned rule base include the rule version number, source review number, approval entity identifier, approval time, verification report reference, effective time, and reference relationship with old versions.

[0020] Step S10: The rules are fed back to the anomaly candidate identification. The versioned rules are used for subsequent anomaly candidate identification signals or gating signals. The review conclusions and rule evolution results are used in reverse to affect the anomaly candidate identification logic and gating judgment, forming a complete knowledge evolution closed loop.

[0021] Furthermore, the method also includes an abnormal candidate event evidence curve playback step: In the debriefing task, the system loads the multi-timescale evidence curves corresponding to the abnormal candidate events. The evidence curves are displayed by selecting result variables, explanatory variables, and background variables according to the candidate type, and human feedback nodes, recovery nodes, rule suggestion nodes, or approval nodes are superimposed on the same event thread timeline. The system compares the changes in variable relationships before, during, and after the candidate event occurs, and writes the comparison results into the structured debriefing record or rule evolution suggestion.

[0022] Furthermore, the method also includes a field prompt library evolution step: the system generates field prompt library suggestions based on the review records. The field prompt library stores field natural language terms, synonymous equipment terms, equipment number families, low-cost inspection items, items to avoid misjudgment, and review prompts. The field prompt library does not directly trigger scanning rules, but only provides contextual support during review analysis, rule suggestion generation, and manual input normalization.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] Construct a complete lifecycle for abnormal events. Expand candidate abnormal events from one-time alarms to traceable, replayable, and archiveable full lifecycle management objects. Use an event thread mechanism to connect initial signals, repeated hits, status escalation, manual feedback, recovery closure, and rule evolution to form a complete auditable chain of evidence.

[0025] Achieve structured accumulation of human experience. Through deterministic fact summaries, standardized feedback types, and structured debriefing records, transform the natural language experience of field personnel into parsable, traceable, and reusable digital knowledge, solving the problems of industrial knowledge dependence on individuals and high risk of loss.

[0026] Establish a progressive false positive management mechanism. A single false positive will not directly block the rule to avoid overreaction; only when similar false positives accumulate to a preset condition will a downgrade or suppression suggestion be triggered, which balances the timeliness and stability of rule optimization, effectively reducing the false positive rate while avoiding the risk of introducing false negatives.

[0027] A triple-verification gate ensures the secure evolution of rules. Before a rule is officially released, it must undergo source replay, historical error checking, and phase adaptation checking to verify its effectiveness from three dimensions: source, history, and phase. This significantly reduces the secondary risks brought about by rule modifications.

[0028] Versioned rule management with full-chain auditability. All rule changes are manually approved, historical versions are fully preserved, and audit traceability and rollback of old versions are supported, ensuring that the rule evolution process is transparent, controllable, and traceable.

[0029] A secure human feedback mechanism. Human feedback is standardized and normalized into standard semantics, stored as facts in an independent ledger, without modifying original production data or directly triggering control actions, thus ensuring the safety of the production system while meeting industrial data auditing requirements.

[0030] Dual-loop re-analysis drives continuous knowledge evolution. Problem re-analysis and production re-analysis are integrated into the process rule evolution system. Micro-level re-analysis of individual events is combined with macro-level re-analysis of the production cycle to continuously optimize anomaly identification capabilities and form a self-evolving industrial knowledge system. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the overall closed-loop process of post-mortem feedback and rule evolution in this invention.

[0033] Figure 2 This invention provides a schematic diagram of the pre-write verification gate for the rule recommendations.

[0034] Figure 3 This is a lifecycle state diagram of abnormal candidate events in this invention;

[0035] Figure 4 This is a timeline diagram illustrating the abnormal process recap of the dehumidifier of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1: Obvious False Alarm Archiving and Rule Suppression Evolution

[0038] like Figure 1-4 This embodiment uses the false alarm management of candidate events for spray drying feed effectiveness as an example to explain in detail the implementation process of the present invention.

[0039] A food company's spray drying production line is equipped with an anomaly candidate identification system. The system monitors parameters such as inlet air temperature, outlet air temperature, tower pressure, and feed pump frequency in real time. During production on a certain day, at 10:15, the system generated an anomaly candidate event for feed effectiveness, event number EVT-20260415-003, with a candidate level of Level 2, indicating that the continuously rising outlet air temperature may indicate a risk of insufficient feed.

[0040] After receiving the candidate event, the system executes step S2 to generate a problem review task, task number REV-20260415-003, with the review type being problem review. Since the time interval between this candidate event and a similar candidate event on the same device 10 minutes prior is less than the preset merging threshold of 15 minutes, the system groups both into the same event thread THD-20260415-002, and the thread status is updated to pending feedback.

[0041] In step S3, the system reads the PLC timing data within a ±5 minute time window, extracts process feature blocks, and generates a deterministic fact summary: From 10:10 to 10:20, the outlet air temperature rose from 87.2℃ to 89.5℃ at a rate of 0.23℃ / min; the inlet air temperature simultaneously rose from 183.8℃ to 186.1℃, showing a positive correlation; the feed pump frequency remained stable at 35Hz, with no significant decrease in the feedback value; the tower pressure remained around -225Pa without significant fluctuations. The fact summary clearly indicates that the inlet and outlet air temperatures rose simultaneously, suggesting that this temperature increase may be caused by inlet air adjustment rather than a feeding problem.

[0042] After viewing the candidate event and fact summary on the central control page, the on-site operator submitted manual feedback based on the on-site inspection. The feedback type was "obvious false alarm," and the remarks were "the air inlet temperature setting was increased, which is a normal adjustment, and the material supply is normal."

[0043] In step S4, the system normalizes the manual feedback: the feedback type is mapped to "obvious false alarm," the associated object is standardized as "feed pump-1," and the feedback semantics are parsed as "the increase in inlet air temperature setting caused a synchronous increase in outlet air temperature, which is not a feeding anomaly." The system writes this feedback into a separate feedback log, records the feedback role as "on-duty operator," does not include employee name, employee number, or other identity information, and does not modify any PLC original data tables.

[0044] After executing step S5, the system generates a structured debriefing record, with the following core fields:

[0045] review_id:REV-20260415-003

[0046] source_event_id:EVT-20260415-003

[0047] review_type:issue

[0048] Facts: ["Outlet air temperature rose from 87.2℃ to 89.5℃ between 10:10 and 10:20", "Inlet air temperature rose simultaneously from 183.8℃ to 186.1℃", "Feed pump frequency remained stable at 35Hz", "Tower pressure remained stable at -225Pa"]

[0049] Explanation: ["Raising the intake air temperature setting causes the exhaust air temperature to rise simultaneously"]

[0050] counter_evidence: ["Feed pump frequency is stable", "Tower pressure shows no abnormal fluctuations"]

[0051] human_feedback:["Clearly a false alarm, air intake is normally adjusted"]

[0052] Conclusion: This is clearly a false alarm; the cause is that the intake air temperature setting was increased.

[0053] rule_implications: ["The rules for ensuring the effectiveness of material supply need to consider the interpretation conditions of synchronous changes in air intake"]

[0054] review_state: Confirmed

[0055] In step S6, the system generates a false alarm archive record, corresponding to the RULE-FEED-007 material supply validity detection rule, applicable to the formal production stage. System statistics show that similar false alarms (false alarms caused by synchronous increase in air intake leading to material supply candidate errors) have occurred 5 times in the formal production stage within the past 30 days, meeting the preset condition of "≥3 similar false alarms triggering rule optimization suggestions".

[0056] In step S7, the system generates a rule correction suggestion, categorized as "Rule Correction." The suggestion adds an inlet air temperature interpretation condition to the material supply validity rule. When both the outlet air temperature and inlet air temperature rise simultaneously, the system suppresses abnormal material supply candidate outputs or downgrades them to the observation level. The system extracts the rule signature: Equipment Object = Drying Tower A, Result Variable = Outlet Air Temperature, Explanatory Variables = Inlet Air Temperature, Material Supply Pump Frequency, Applicable Stage = Formal Production, Trend Direction = Upward, Gating Action = Suppress. Rule signature comparison confirms this suggestion has not been generated before, avoiding duplicate suggestions.

[0057] In step S8, the system suggests performing triple verification for this rule:

[0058] Source replay: The revised rules were applied to the 10:10-10:20 time window corresponding to this review, and the verification results showed that the rules correctly suppressed the candidate outputs, which was in line with expectations;

[0059] Historical false alarm check: The revised rule was applied to historical data of 12 production batches over the past 90 days. The check results showed that the rule did not miss any confirmed real material supply anomalies and successfully suppressed 7 similar false alarms, with a false alarm rate of 0.

[0060] Phase Adaptation Check: Confirmed that the rule modification only applies to the formal production phase, consistent with the operational phase of the source event, and will not affect other phases such as startup and water operation. All three verifications passed, verification report number VAL-20260415-002.

[0061] Execute step S9, and the rule is recommended to be submitted to the process engineer for approval. The engineer reviews the verification report and approves it at 14:30 on April 15, 2026, with the approval note "Agreed to add the air intake explanation conditions." The system releases a new version of the rule, RULE-FEED-007v2.1, recording the version number, source review number, approval entity identifier, approval time, verification report citation, and effective time, while retaining the original version v2.0, ensuring a complete historical audit chain.

[0062] In step S10, the new version of the rules is synchronized to the rule base of the anomaly candidate identification module. Subsequent material supply validity candidate identification will automatically apply the new interpretation conditions, and similar false alarms will be effectively suppressed.

[0063] Example 2: Stage-Adaptive Rule Revision

[0064] This example illustrates the differentiated adaptation and evolution process of rules at different production stages.

[0065] The tower pressure anomaly candidate rules of a certain spray drying production line performed well during the formal production stage, effectively identifying real anomalies such as airflow blockage. However, they frequently triggered false alarms during the water operation stage. The water operation stage is the preheating and cleaning stage after the production line is started. There is only water and hot air in the tower, without actual liquid material, and the fluctuation characteristics of process parameters are different from those during formal production.

[0066] The system's production review revealed that, across nearly 10 production batches, 23 tower pressure anomaly candidates were triggered during the water operation phase. Of these, 21 were manually flagged as obvious false alarms, resulting in a false alarm rate exceeding 90%. The system automatically generated a rule correction suggestion: split the tower pressure anomaly candidate rule into phase versions. Maintain the original warning level during the formal production phase, and downgrade it to the observation level during the water operation phase, without outputting any warning candidates.

[0067] The rules recommend entering the verification gate: Source playback verification shows that all historical false alarm samples during the water operation phase were correctly downgraded; historical false alarm checks show that real anomalies during the formal production phase are unaffected and all hit normally; phase adaptation checks confirm that the boundary between the two phases is clear and consistent with the actual production phase switching logic.

[0068] After manual approval, the rule was released as a two-stage version RULE-PRES-012v3.0, with the stage mapping relationship retained in the rule base. After going live, false tower pressure alarms during the water operation stage were reduced by more than 90%, while the anomaly detection capability during the formal production stage remained unaffected.

[0069] Example 3: Evolution of Production Retrospective and On-site Notification Library

[0070] This embodiment illustrates the implementation method of production review and its optimization effect on the on-site prompt library.

[0071] After a production cycle ends, the system automatically triggers a production review task, summarizing the entire process data for this batch, from startup, water operation, formal production, raw material switching, and shutdown / cooling. The production review record includes: 17 candidate abnormal events were generated in this batch, of which 3 were confirmed abnormalities, 8 were obvious false alarms, and 6 require continued observation; false alarms mainly focused on temperature fluctuations during the raw material switching phase and tower pressure changes before shutdown; two reusable lessons were extracted: first, temperature fluctuations of ±2℃ during the raw material switching phase are normal background fluctuations and should not trigger temperature abnormality candidates; second, rapid tower pressure changes at the initial stage of shutdown / cooling are normal processes, and a ventilation system status check item should be added as an auxiliary judgment.

[0072] Based on the production review conclusions, the system generates two rule suggestions and three updates to the on-site prompt library. The rule suggestions supplement the explanation conditions for temperature fluctuation suppression during raw material switching and tower pressure during material shutdown. The newly added content to the on-site prompt library includes: reference values ​​for temperature fluctuations during raw material switching, characteristics of tower pressure changes during material shutdown, and key points for indirect judgment of dehumidifier tripping. After the on-site prompt library is updated, it automatically provides contextual references during subsequent review analysis and manual input normalization, improving review efficiency and accuracy.

[0073] Example 4: Indirect Review of Dehumidifier Malfunctions

[0074] This embodiment uses the example of a device malfunction that is not directly connected to the data acquisition system to illustrate the indirect evidence review capability of the present invention.

[0075] The dehumidifier on a spray drying production line was not connected to the PLC data acquisition system, making it impossible to directly read its operating status and trip signals. During production one day, the system detected an abnormal combination of changes in tower pressure and outlet air temperature: the tower pressure suddenly dropped from -230Pa to approximately -280Pa, followed by a slow rise in outlet air temperature from 88.5℃ to approximately 90℃, while the parameters of key equipment such as inlet air temperature, feed pump, and atomizer showed no significant adjustment. The system generated a candidate event for a cold air link anomaly.

[0076] After on-site investigation, it was found that the dehumidifier had tripped. Once power was restored, the parameters gradually returned to normal. During the review, the system constructed a timeline based on indirect evidence: 12:27 Tower pressure began to drop rapidly; 12:35 Outlet air temperature began to rise; 15:30 Dehumidifier tripping was discovered on-site; 15:40 After power was restored, tower pressure and outlet air temperature gradually decreased. The system correlated the changes in indirect parameters with the confirmed on-site facts, forming a complete chain of evidence of anomalies.

[0077] Based on this review record, the system generated two results: First, it added indirect identification rules for dehumidifier malfunctions, identifying dehumidifier faults through a combination of tower pressure surges, outlet temperature increases, and lack of main equipment adjustment. Second, it supplemented the on-site alert database with indirect judgment features and inspection paths for dehumidifier tripping. The rule recommendations were released and implemented after historical playback verification and manual approval, achieving the ability to indirectly identify malfunctions in equipment not directly sourced.

[0078] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for post-mortem feedback and process rule evolution for candidate abnormal events in spray drying, characterized in that, include: Receive abnormal candidate events: Receive abnormal candidate events in spray drying. The abnormal candidate events include event number, event thread number, occurrence time, associated equipment, associated variables, current operating stage, trend evidence, gating result and candidate level. The abnormal candidate events are not final fault conclusions. Generate debriefing tasks: Generate debriefing tasks based on the abnormal candidate events. The debriefing tasks are divided into problem debriefing and production debriefing. When multiple abnormal candidate events belong to the same device, the same candidate type, and the same rule source, and the time interval meets the merging condition, they are merged into the same event thread. Read evidence and generate fact summary: Read the PLC timing evidence, process feature block and gate record within the time window corresponding to the abnormal candidate event, and generate a deterministic fact summary, which includes the correspondence between the numerical changes and trends of the related variables; Receive and normalize human feedback: Receive human feedback and normalize it into a fixed feedback type and standardized semantics. The human feedback is written into an independent feedback ledger, without modifying the original sampling data of the PLC, and without directly triggering automatic control. Generate a structured debriefing record: Based on the aforementioned abnormal candidate events, PLC timing evidence, and human feedback, generate a structured debriefing record, which includes factual evidence, timeline, candidate explanations, counter-evidence, debriefing conclusions, and rule impact items; False alarm archiving and sample accumulation: Candidate events marked as obvious false alarms will generate false alarm archive records. A single false alarm will not directly block the rule. When the accumulation of similar false alarms meets the preset conditions, suggestions for rule downgrading or suppression will be generated. Generate rule evolution suggestions: Generate rule evolution suggestions based on the structured retrospective records, and extract the rule signatures containing device objects, result variables, explanatory variables, and applicable stages to avoid duplicate suggestions; Triple verification gate check: Before the rule evolution recommendation is released, at least source playback, historical false alarm check and stage adaptation check should be performed. If the verification fails, it should be rolled back for revision or downgraded to an observation item. Manual approval and versioned release: The rule evolution suggestions are manually approved and a manual approval record is generated. The approved rule evolution suggestions are then entered into the official rule library. After release, the rule evolution suggestions carry a version identifier and generate versioned rules. Historical versions are fully preserved to form an auditable rule evolution chain. Rule feedback application: The versioned rules are used for the identification of abnormal candidate events, and the review conclusions are applied in reverse to the abnormal identification logic, forming a knowledge evolution closed loop.

2. The method for review and feedback and process rule evolution for candidate abnormal events in spray drying according to claim 1, characterized in that, The event thread records the start time, most recent trigger time, duration, number of reminders, feedback status, and recovery status, which is used to connect multiple triggers, manual interventions, and recovery processes of the same type of exception.

3. The method for review and feedback and process rule evolution for candidate abnormal events in spray drying according to claim 1, characterized in that, The deterministic fact summary for material supply-related candidates includes at least the inlet air temperature, outlet air temperature, tower pressure, feed pump setpoint and feedback value within the same time window, and marks the trend correspondence between inlet and outlet air temperatures.

4. The method for review and feedback and process rule evolution for candidate abnormal events in spray drying according to claim 1, characterized in that, The types of human feedback include confirmed receipt, confirmed anomaly, obvious false alarm, processed, continued observation, uncertain, and supplementary on-site facts.

5. The method for review and feedback and process rule evolution for candidate abnormal events in spray drying according to claim 1, characterized in that, The types of rule evolution suggestions include adding rules, modifying rules, downgrading rules, suppressing rules, archiving rules, and not changing rules.

6. The method for review and feedback and process rule evolution for candidate abnormal events in spray drying according to claim 1, characterized in that, The source playback involves applying the suggested rules to the core time window corresponding to the source replay to verify whether the original anomaly can be hit or the original false alarm can be correctly downgraded; the historical false alarm check involves applying the suggested rules to the historical production cycle to check whether false hits occur during startup, water operation, raw material switching, material shutdown and cooling, or drying tower stages; the stage adaptation check involves checking whether the stage to which the rules are applied is consistent with the operation stage of the source replay and historical production.

7. The method for review and feedback and process rule evolution for candidate abnormal events in spray drying according to claim 1, characterized in that, The manual approval record includes the approval entity identifier, approval time, approval conclusion, approval remarks, and verification report reference; the record in the formal rule base includes the rule version number, source review number, approval entity identifier, approval time, verification report reference, effective time, and old version reference relationship.

8. The method for review and feedback and process rule evolution for candidate abnormal events in spray drying according to claim 1, characterized in that, It also includes an evidence curve replay step: loading multi-timescale evidence curves in the debriefing task, displaying result variables, explanatory variables and background variables according to the candidate type, superimposing manual feedback nodes, recovery nodes and rule suggestion nodes on the time axis, and comparing the changes in variable relationships before and after the anomaly occurred.

9. The method for post-mortem feedback and process rule evolution for candidate abnormal events in spray drying according to claim 1, characterized in that, It also includes the evolution steps of the on-site prompt library: generating on-site prompt library suggestions, saving on-site natural language terms, synonymous equipment terms, low-cost inspection items and items to avoid misjudgment, and the on-site prompt library provides contextual support during retrospective analysis and normalization of manual input.

Images