A large model assisted power generation working condition data intelligent labeling method

Abstract

The application discloses a large model assisted power generation working condition data intelligent labeling method, relates to the technical field of data labeling, and aims to solve the technical problems of low labeling quality and model training failure caused by the misplacement of log time and physical time, the inability of a fixed window to match dynamic evolution in the prior art labeling method, and comprises the following steps: according to an auxiliary machine object in a structured trip event semantic framework generated by a large model, a DCS candidate curve set corresponding to the auxiliary machine object is matched; and taking a trip event time recorded in a log as initial positioning basis, at least one curve change evidence in device running state flip, related flow change, standby device response change and main parameter disturbance is extracted from the DCS candidate curve set. The application replaces a fixed time window by taking the adjacent stage evidence deviation degree, the interlocking response chain node fitting degree and the stage confusion degree as core criteria based on a dynamic labeling boundary determination formula, so that the boundaries of each stage are matched with the evolution of a physical process.

Description

Technical Field

[0001] This invention relates to the field of data annotation technology, and more specifically, to a method for intelligent annotation of power generation operating condition data assisted by a large model. Background Technology

[0002] As the level of automation control and operation monitoring of generator sets continues to improve, multi-source operation data such as operation logs, alarm records, operation records, interlock action records, and DCS historical curves are continuously generated during the operation of the units. Intelligent annotation of power generation condition data refers to the identification, classification, and assignment of abnormal event objects, event triggering locations, curve change stages, operation impact objects, and label sources in the above-mentioned multi-source operation data. This transforms the originally scattered log text, alarm information, and curve fragments into structured annotation results that can be used for the construction of abnormal operating condition sample libraries, fault diagnosis model training, operation review analysis, and operating condition identification model training.

[0003] However, in the data annotation scenario of generator auxiliary equipment tripping conditions, existing annotation methods mostly directly adopt the unified timestamp of log records, ignoring log entry lag, system clock deviation, and manual operation delays. This results in a time gap between the text record time and the physical tripping moment reflected by the curve. Adding to this the complex interlocking response logic and media transmission lag of different auxiliary equipment types (pumps, fans, etc.), it becomes impossible to accurately correspond to the actual physical evolution process of symptoms, triggers, responses, disturbances, and recovery by simply relying on operating condition segments captured by fixed time windows. It also fails to accurately capture the subtle features at the moment of tripping and the dynamic chain of subsequent interlocking propagation. Once this low-quality annotated data is used for model training, it not only fails to improve fault identification accuracy but also leads to systematic biases in the operating condition extrapolation results. Therefore, we propose a large-model-assisted intelligent annotation method for generator operating condition data. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology, adapt to the needs of reality, and provide a large model-assisted intelligent annotation method for power generation operating data. This method aims to solve the technical problems of low annotation quality and model training failure caused by the misalignment of log time and physical time and the inability of fixed windows to match dynamic evolution in existing annotation methods.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for intelligent annotation of power generation operating condition data assisted by a large model, comprising the following steps: Based on the auxiliary machine objects in the structured trip event semantic framework generated by the large model, match the DCS candidate curve set corresponding to the auxiliary machine objects; Using the trip event time recorded in the log as the initial location basis, extract evidence of at least one curve change from the DCS candidate curve set, including equipment operating status reversal, drive current sudden drop, outlet pressure change, related flow change, standby equipment response change, and main parameter disturbance. The auxiliary machine objects and tripping actions identified by the large model are used as semantic gating conditions. Combined with direct tripping evidence, delayed response evidence, timing consistency and conflict penalty, the actual tripping trigger position after curve alignment correction is determined. Using the actual trip trigger location as the auxiliary machine trip node, and based on the semantic framework of the structured trip event and the relationship between the curve changes in the DCS candidate curve set, an interlocking response chain from trip trigger to recovery of stability is constructed. Based on the degree of deviation of evidence between adjacent stages, the degree of fit of interlocking response chain nodes, and the degree of stage confusion, the continuous curve segment corresponding to the auxiliary machine tripping is dynamically divided into multiple time-progressive operating condition stages.

[0006] Preferably, the DCS candidate curve set is jointly determined by the auxiliary machine object, auxiliary machine type, interlock response description, and affected operating parameters in the structured trip event semantic framework; For different types of auxiliary equipment, the DCS candidate curve set should at least include the body monitoring dimension curves that reflect the operating status or changes in drive parameters of the auxiliary equipment, the process parameter dimension curves that reflect changes in process medium pressure, flow rate, air volume, temperature or water level, the interlock response dimension curves that reflect the activation of standby equipment and the action of valves, guide vanes or baffles, and the system impact dimension curves that reflect the disturbance of the unit's operating status.

[0007] Preferably, in the process of determining the actual trip trigger position after curve alignment correction, the matching result of the auxiliary machine object identified by the large model is used as the semantic gating factor of the auxiliary machine object, and the matching result of the curve change direction and the trip action is used as the semantic gating factor of the trip action. The normalized disturbance weighted value of the direct trip evidence measurement point set and the normalized disturbance weighted value of the delayed response evidence measurement point set are respectively gating and verified. Combined with the timing consistency factor, the trigger positioning stability factor and the conflict penalty factor, the candidate moment with the highest comprehensive score is selected as the actual trip trigger position.

[0008] Preferably, the actual trip trigger position is determined according to the following formula: ; in, This indicates the actual trip trigger position determined after curve alignment correction; This refers to auxiliary machine objects identified by the large model from operation logs, alarm records, or operation records; Indicates the candidate trigger time; Indicates the auxiliary object The corresponding set of candidate trigger times; This represents the semantic gating factor of the auxiliary machine object; Indicates the semantic gating factor of the tripping action; Indicates the auxiliary object The set of measurement points directly corresponding to the tripping evidence; Represents the set of measurement points for direct tripping evidence. Any measuring point in the; Indicates the measurement point for direct tripping evidence Weight of evidence; Indicates the measurement point for direct tripping evidence At the candidate trigger time The normalized perturbation at the location; Indicates the auxiliary object The set of indirect corresponding delay response evidence measurement points; Represents the set of evidence points for delayed response. Any measuring point in the range; Indicates the evidence measurement points for delayed response Weight of evidence; Indicates the evidence measurement points for delayed response The response lag relative to the candidate trigger moment; Indicates the evidence measurement points for delayed response The normalized perturbation at the response lag time; Indicates the candidate trigger time Temporal consistency factor at the location; Indicates the candidate trigger time Conflict punishment factor at the location; This indicates the triggering of the positioning stabilization factor.

[0009] Preferably, in the process of dynamically dividing the operating condition into multiple time-sequential progressive stages, the operating condition stages include at least the pre-trip symptom stage, the trip trigger stage, the interlock response stage, the main parameter disturbance stage, the manual handling stage, and the recovery and stabilization stage. The dynamic annotation boundary between adjacent working condition stages is determined by the difference in the strength of dominant evidence between adjacent stages, the minimum distinguishable factor of the stage boundary, the stability factor of the boundary calculation, the stage confusion degree, and the degree of fit of the interlocking response chain nodes. The candidate time point with the smallest boundary evaluation value is selected as the dynamic annotation boundary.

[0010] Preferably, before determining the actual trip trigger location, the curve changes of each measuring point in the DCS candidate curve set are first converted into dimensionless normalized disturbance quantities. The normalized disturbance is calculated based on the actual sampled value of the measuring point at the candidate time point, the local stability benchmark before the candidate time point, the local fluctuation amplitude, and the minimum resolvable change of the measuring point.

[0011] Preferably, the normalized perturbation is determined according to the following formula: ; in, Indicates the first Each DCS measurement point at the candidate time point The actual sampled value at that location; Indicates the first Each DCS measurement point at the candidate time point Previous local stability benchmark; Indicates the first Each DCS measurement point at the candidate time point The previous local fluctuation range; Indicates the first The minimum resolvable change at each DCS measurement point; Indicates the first Each DCS measurement point at the candidate time point The normalized perturbation at the location.

[0012] Preferably, the interlocking response chain is formed based on the interlocking description in the structured trip event semantic framework and the sequence of curve changes in the DCS candidate curve set, and includes at least the auxiliary machine tripping node, the standby equipment start-up node, the relevant valve action node, the medium pressure or flow recovery node, the main parameter disturbance node, the manual handling node, and the recovery stabilization node. Among them, the auxiliary equipment tripping node is determined by the actual tripping trigger position; the standby equipment start-up node is determined by the response change of the standby equipment operating status curve or standby equipment current curve; the relevant valve action node is determined by the state change of the valve action curve; the main parameter disturbance node is determined by the continuous response of steam pressure, water level, air volume, furnace negative pressure, load or combustion parameters; the manual handling node is determined by the operation record and the corresponding curve change; and the stabilization recovery node is determined by the position where the affected main parameter returns to a stable operating state.

[0013] Preferably, before matching the DCS candidate curve set, the process further includes: The system collects operation logs, alarm records, operation records, interlock action records, and DCS historical curve data generated during the operation of the generator set. It also standardizes and unifies the equipment name, measurement point name, event time, and curve time in the collected data to form unified operating condition data to be labeled. The large language model, fine-tuned by text in the power generation field, is invoked to perform semantic parsing on the operation logs, alarm records, and operation records. From the parsing results, auxiliary equipment objects, tripping actions, interlocking descriptions, standby equipment activation descriptions, manual handling descriptions, and recovery descriptions are identified, and the structured tripping event semantic framework is generated based on the identification results.

[0014] Preferably, after the work condition stages are divided, corresponding event main labels, stage labels and impact labels are generated for each work condition stage, and structured annotation results containing the labels and their corresponding time interval information are output. Among them, the main event label is used to characterize the type of auxiliary machine tripping event, the stage label is used to characterize the development stage of the curve segment in the entire process of auxiliary machine tripping, and the impact label is used to characterize the object of the impact of auxiliary machine tripping on the generator set operating parameters. The structured annotation results should include at least the event number, auxiliary machine object, actual trip trigger location, start and end positions of the curve segment, main event label, stage label, impact label, label source, and review status.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention automatically identifies key links such as auxiliary machine tripping nodes, standby equipment start-up nodes, valve action nodes, and main parameter disturbance nodes based on the structured tripping event semantic framework and curve change sequence, forming a complete event propagation path. Based on the dynamic annotation boundary determination formula, it uses the degree of deviation of evidence between adjacent stages, the degree of fit of interlocking response chain nodes, and the degree of stage confusion as the core criteria to replace the fixed time window, realizing the adaptive division of the pre-tripping symptom stage, tripping trigger stage, interlocking response stage, main parameter disturbance stage, manual handling stage, and recovery and stabilization stage. This makes the boundaries of each stage match the evolution of the physical process, solving the problems of low annotation quality and model training failure caused by the misalignment of log time and physical time and the inability of fixed windows to match dynamic evolution in existing annotation methods.

[0016] 2. This invention also transforms measurement data of different physical units such as operating status, current, pressure, and flow rate into dimensionless disturbance evidence through a normalized disturbance calculation method. Combined with the formula for determining the actual trip trigger position, it introduces auxiliary machine object semantic gating, trip action semantic gating, direct trip evidence weighting, delayed response evidence weighting, timing consistency factor, and conflict penalty factor to achieve multi-dimensional verification of log time and curve changes. This avoids the labeling lag or misjudgment caused by simply relying on log time, improves the accuracy of trip trigger position positioning, and makes the labeling starting point consistent with the actual physical process.

[0017] 3. This invention also achieves the effect of mutual verification between text semantics and curve evidence by designing a structured trip event semantic framework and a matching method with the DCS candidate curve set. Through the structured trip event semantic framework, auxiliary machine objects, trip actions, interlock descriptions, standby equipment activation descriptions, manual handling descriptions, and recovery descriptions are identified from the operation log, alarm records, and operation records. Based on this, the DCS candidate curve set related to the current trip event is matched, so that the operating status curve, drive current curve, pressure and flow curve, standby equipment response curve, and main parameter curve can jointly participate in trigger positioning and stage division, thereby improving the reliability of the annotation results. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a flowchart of the multi-source data acquisition and standardization processing of the present invention; Figure 3 This is a flowchart of the large-scale model semantic parsing and event framework construction process of the present invention; Figure 4 This is a flowchart illustrating the trip time alignment and actual trigger position correction of the present invention. Figure 5 This is a flowchart illustrating the construction of the interlocking response chain and the dynamic division of operating phases in this invention. Figure 6 This is a flowchart of the multi-level structured annotation result output of the present invention. Detailed Implementation

[0019] Example: Figures 1 to 6 As shown, the present invention relates to a large-model-assisted intelligent annotation method for power generation operating condition data, comprising the following steps: S100, Multi-source data acquisition and standardized processing: By collecting logs, alarms, operations, interlocking records and DCS curve data during generator set operation, and unifying equipment naming and time reference, standardized data to be labeled is formed; S101. Collect the operation log, alarm record, operation record, interlock action record and DCS historical curve data generated during the operation of the generator set; The operation log is used to record the equipment status, abnormal phenomena and handling process by the operators; the alarm log is used to record the alarm information of auxiliary machine tripping, abnormal equipment status or abnormal related parameters; the operation log is used to record the operation behavior performed by the operators on the auxiliary machine, standby equipment, valve or related controlled object; the interlock action record is used to record the equipment action process triggered by the interlock logic after the auxiliary machine trips; and the DCS historical curve data is used to record the actual process of the auxiliary machine status, auxiliary machine current, outlet pressure, related flow, standby equipment response and the main parameters of the unit changing over time.

[0020] S102. Standardize and unify the equipment name, measurement point name, event time, and curve time in the collected data to eliminate format differences and time series reference deviations of multi-source data and form unified working condition annotation data. Equipment name standardization: Unify the mapping of equipment aliases, abbreviations, and job-specific terms (such as "#1 induced draft fan", "No. 1 induced draft fan", and "A induced draft fan") in operation logs, alarm records, operation records, and interlock action records, establish an equipment alias mapping table, and convert them into equipment object names (such as "ID_FAN_1A") that are standardized with DCS measurement point names.

[0021] Measurement point attribution and organization: The state variables, analog variables, and operational variables in the DCS historical curves are merged according to auxiliary machine objects. For example, for "Water Pump A", its operating status curve, current curve, outlet pressure curve, inlet flow curve, outlet valve opening curve, etc. are all merged under the same event to be labeled, forming a logically related curve cluster.

[0022] Unified time standard: Ensure that the time stamps of all data sources use the same time system (such as Beijing time, accurate to milliseconds). For systems with time drift, perform offline calibration by synchronizing the time offset recorded by NTP to make the log recording time comparable to the DCS curve time.

[0023] S200, Large Model Semantic Analysis and Event Framework Construction: Utilize a large language model to perform semantic analysis on logs, alarms, and operation texts, identify event elements related to auxiliary machine tripping, and generate a structured tripping event semantic framework; S201. The open-source large language model LLaMA-3-8B-Instruct is used as the base model and deployed on a local server. Considering the characteristics of log text in the power generation field, historical logs, alarms, and operation records are collected to form training corpora. Supervised fine-tuning (SFT) is performed on the model to improve its recognition accuracy of professional terms and expressions such as auxiliary equipment, tripping actions, and interlocking logic. The fine-tuned large language model is then used to perform semantic parsing on the standardized operation logs, alarm records, and operation records. The large language model is used to process unstructured text content in operation logs, alarm records, and operation records. Its role is to identify semantic information in the text related to auxiliary machine tripping events, such as objects, actions, interlocks, handling, and recovery, rather than directly generating final operating condition labels.

[0024] S202. Identify the auxiliary machine object, tripping action, interlocking description, standby equipment activation description, manual handling description, and recovery description in the auxiliary machine tripping event from the analysis results; In step S202, when the large model performs semantic parsing, it first extracts event elements from the operation log, alarm record, and operation record. The event elements include event object, event action, event occurrence description, interlock response description, standby equipment action description, operator handling description, abnormal impact description, and recovery status description. Then, a structured trip event semantic framework is generated based on the event elements, so that the trip event in the unstructured text is transformed into structured data that can be matched and verified with DCS curve data. For example, in a "#1 induced draft fan trip" event, the operation log records "#1 induced draft fan A tripped, operators adjusted furnace negative pressure," the alarm record records "#1 induced draft fan A trip alarm, #1 induced draft fan B interlock start alarm," and the operation record records "adjusted induced draft fan blade opening or furnace negative pressure control setting." After semantic parsing the above text, the large model identifies the event object as ID_FAN_1A, i.e., #1 induced draft fan A; identifies the event action as tripping or abnormal shutdown; identifies the interlock response description as #1 induced draft fan A tripping triggers #1 induced draft fan B interlock start; identifies the standby equipment action description as #1 induced draft fan B starting up; identifies the operator's handling description as operators adjusting induced draft fan blade opening, inlet damper opening, or furnace negative pressure control setting; identifies the abnormal impact description as furnace negative pressure disturbance, airflow disturbance, or combustion state disturbance; and identifies the recovery state description as furnace negative pressure or related main parameters returning to stability after a recovery process.

[0025] S203. Generate a structured semantic framework for the tripping event based on the identified information; The structured trip event semantic framework includes at least auxiliary equipment objects, trip actions, interlock descriptions, standby equipment activation descriptions, manual handling descriptions, and recovery descriptions. This structured trip event semantic framework transforms previously scattered and unstructured log and alarm texts into the basic event information required for subsequent DCS candidate curve matching, trip trigger location correction, and interlock response chain construction.

[0026] S300, Alignment Correction of Log Time and Curve Time: Match the corresponding DCS candidate curve set according to the auxiliary machine object, and use the feature points such as state flip and parameter change in the curve to correct the trip time recorded in the log, and accurately locate the actual trip trigger position. S301. Match the corresponding DCS candidate curve set according to the auxiliary machine object; Specifically, the DCS candidate curve set is not arbitrarily selected, but is determined by the auxiliary machine objects, auxiliary machine types, interlock response descriptions, and affected main parameters in the structured trip event semantic framework. Based on the auxiliary machine objects and event actions obtained from the large model analysis, combined with the equipment alias mapping table, measurement point affiliation relationship, and interlock logic relationship established in S100, the system indexes and filters the DCS measurement point group that matches the current trip event from the historical DCS curve data to form the corresponding DCS candidate curve set. The DCS candidate curve set configuration is as follows for different auxiliary machine objects: For pump-type equipment, taking a feedwater pump as an example; when the auxiliary machine is a feedwater pump, the DCS candidate curve set includes the feedwater pump operating status curve, feedwater pump current curve, feedwater pump outlet pressure curve, feedwater flow curve, standby feedwater pump operating status curve, standby feedwater pump current curve, feedwater regulating valve opening curve, steam drum water level curve, and main steam pressure curve. Among them, the feedwater pump operating status curve and feedwater pump current curve are used to reflect whether the feedwater pump body has tripped or abnormally stopped; the feedwater pump outlet pressure curve and feedwater flow curve are used to reflect the changes in the medium parameters of the feedwater system; the standby feedwater pump operating status curve and standby feedwater pump current curve are used to reflect whether the standby equipment has been interlocked; the feedwater regulating valve opening curve is used to reflect the action of the regulating object; and the steam drum water level curve and main steam pressure curve are used to reflect the disturbance caused by the feedwater pump tripping to the main parameters of the unit, thus providing curve basis for the subsequent division of the main parameter disturbance section and the recovery and stabilization section. For example, in the event of "feed pump A tripping", if the large model identifies semantic information such as "feed pump A tripping", "standby feed pump B starting", and "water level fluctuation" from the log, the system will use the operating status curve and current curve of feed pump A as direct tripping evidence, the operating status curve and current curve of standby feed pump B as delayed response evidence, and the steam drum water level curve, main steam pressure curve and feed water regulating valve opening curve as evidence of main parameter disturbance and manual handling. For fan-type equipment, taking the induced draft fan as an example; when the auxiliary machine is the induced draft fan, the DCS candidate curve set includes the induced draft fan operating status curve, the induced draft fan current curve, the outlet air pressure curve, the inlet air volume curve, the inlet guide vane opening curve, the standby induced draft fan operating status curve, the standby induced draft fan current curve, the furnace negative pressure curve, and the oxygen content curve. Among them, the induced draft fan operation status curve and induced draft fan current curve are used to reflect whether the induced draft fan body has tripped or abnormally stopped; the outlet air pressure curve and inlet air volume curve are used to reflect the change in the flow state of the flue gas system; the inlet guide vane opening curve is used to reflect the action of the regulating mechanism; the standby induced draft fan operation status curve and standby induced draft fan current curve are used to reflect whether the standby fan has been interlocked and started; and the furnace negative pressure curve and oxygen curve are used to reflect the impact of the induced draft fan tripping on the furnace negative pressure control and combustion state, thus providing curve basis for the subsequent division of the interlock response section, the main parameter disturbance section and the recovery and stabilization section. For example, in the event of “induced draft fan A tripping”, the system uses the operating status curve and current curve of induced draft fan A as direct evidence of tripping, the operating status curve and current curve of standby induced draft fan B or the action curve of related dampers and inlet guide vanes as evidence of interlocking response, and the relevant curves of outlet air pressure, inlet air volume, furnace negative pressure, oxygen content and combustion status as evidence of system impact. Other auxiliary power generation equipment shall be determined according to general rules; when the auxiliary equipment is a coal mill, condensate pump, circulating water pump, or other auxiliary power generation equipment not listed separately, the DCS candidate curve set shall include curves in at least the following dimensions: The main body monitoring dimension curves include the auxiliary machine operation status curve and the drive current curve, which are used to determine whether the auxiliary machine body has tripped, stopped or abnormally acted. Process parameter dimension curves, including outlet medium parameter curves and associated flow parameter curves, are used to determine whether the corresponding medium pressure, temperature, flow rate or process status changes after the auxiliary equipment trips. Interlock response dimension curves, including standby equipment operating status curves, standby equipment current curves, and related regulating valve, guide vane, or baffle opening curves, are used to determine whether standby equipment is activated after auxiliary equipment trips and whether the regulating object has undergone interlocking action. The system impact dimension curves include the main parameter curves affected by the auxiliary machine tripping, used to determine the subsequent impact of the auxiliary machine tripping on the unit's operating status; for example, when the coal mill trips, the relevant curves of boiler load, coal feed rate, combustion state or main steam pressure can be selected; when the condensate pump trips, the relevant curves of condenser water level, condensate flow rate or deaerator water level can be selected; and when the circulating water pump trips, the relevant curves of circulating water flow rate, condenser vacuum or exhaust steam temperature can be selected. S302. Based on the feature points in the DCS candidate curve set that reflect equipment operating status reversal, current drop, outlet pressure change, related flow change and standby equipment response change, the trip time recorded in the log is corrected by curve alignment. The time of the tripping event recorded in the log is used as the initial location basis, but not directly as the final annotation starting point. The system searches for the curve change combination that matches the auxiliary machine tripping in the DCS candidate curve set, and uses the curve change position that can reflect the actual operating status change of the auxiliary machine as the alignment correction basis. Furthermore, to avoid the inability to directly integrate measurement points such as current, pressure, flow rate, water level, valve opening, and operating status due to different physical units, the curve changes of each DCS measurement point are converted into dimensionless normalized disturbance quantities.

[0027] For any DCS measurement point and candidate time points First at the candidate time point Previously, continuous sampling data from the measurement point under the same operating condition without significant operational switching was selected as a local reference segment; the local stability benchmark was determined based on the sampled values ​​in this local reference segment. The local fluctuation amplitude is determined based on the degree of dispersion of the sampled values ​​in the local reference segment relative to the local stability reference. Then, determine the minimum resolvable change based on the data acquisition accuracy, storage accuracy, or minimum state change step size of the measuring point. Finally, based on the actual sampled values ​​at the candidate time points... With local stability benchmark The deviation, combined with the local fluctuation amplitude and the smallest distinguishable change The normalized perturbation at the candidate time point is obtained. The normalized perturbation can be determined as follows: ; in, Indicates the first Each DCS measurement point at the candidate time point The actual sampled value at that location; Indicates the first Each DCS measurement point at the candidate time point Previous local stability benchmark; Indicates the first Each DCS measurement point at the candidate time point The previous local fluctuation range; Indicates the first The minimum resolvable change at each DCS measurement point; Indicates the first Each DCS measurement point at the candidate time point The normalized perturbation at the location.

[0028] For discrete state quantities such as operating status and interlocking status, their state change values ​​can be determined based on the original state code or preset state code in the DCS system, and the minimum state change step size between adjacent states can be taken as the minimum resolvable change of the measuring point. For continuous analog quantities such as current, pressure, flow rate, water level, and valve opening, the accuracy can be determined based on the acquisition accuracy, storage accuracy, or minimum display resolution configured in the DCS measurement point setup. For delayed response evidence measurement points Its response time lag Normalized perturbation at the location It is determined in the same way.

[0029] S303. Determine the actual trip trigger location; When determining the actual trip trigger location in S303, the trip event time recorded in the log is used as the initial positioning basis. The curve change combination that matches the auxiliary machine trip is searched in the corresponding DCS candidate curve set. The curve change combination includes one or more of the following: the auxiliary machine operating status changes from running to stopping, the auxiliary machine current suddenly drops from the working value to no load or zero value, the outlet pressure or related flow abnormally decreases, the standby equipment changes from stopping to running, and the associated main parameters are subsequently disturbed. When the log recording time and the curve change position are inconsistent, the curve change position that can simultaneously reflect the auxiliary machine status change and the process parameter response is taken as the true trip trigger position. When determining the true trip trigger position, the auxiliary machine object and trip action identified by the large model are used as semantic gating conditions, and the true trip trigger position is determined by combining direct trip evidence measurement points, delayed response evidence measurement points, timing consistency, and conflict penalty. The specific formula is as follows: ; in, This indicates the actual trip trigger position determined after curve alignment correction, in units of time; This refers to auxiliary machine objects identified by the large model from operation logs, alarm records, or operation records; Indicates the candidate trigger time, in units of time; Represents auxiliary machine objects The corresponding set of candidate trigger times consists of the DCS curve changes around the log recording time; The semantic gating factor of the auxiliary machine object is a dimensionless quantity used to determine candidate trigger times. Do the corresponding curve measurement points match the auxiliary machine objects identified by the large model? Consistent, if candidate time For DCS measurement points where state reversal or parameter abrupt change occurs, their equipment codes and auxiliary machine objects identified from the text are compared. If the standardized codes are completely consistent, then ,otherwise ; The tripping action semantic gating factor is a dimensionless quantity used to determine candidate trigger times. Check whether the curve changes in the vicinity are consistent with semantic actions such as tripping, stopping, power failure, or interlocking. Place If the direction of change of the measuring points in the set is consistent with the expected physical change of the "trip" action (such as the operating state changing from 1 to 0, the current suddenly dropping from a high value to near zero, and the pressure / flow rate starting to decrease from a stable value), then... ,otherwise ; Represents auxiliary machine objects The set of directly corresponding trip evidence measurement points includes measurement points that can directly reflect changes in auxiliary machine tripping, such as operating status, drive current, outlet pressure, and related flow. Represents the set of measurement points for direct tripping evidence. Any measuring point in the range; Indicates the measurement point for direct tripping evidence The evidence weight is a dimensionless quantity, and the evidence weight is determined by the attribution relationship between the measuring point and the auxiliary machine object, the integrity of the curve, and the consistency of historical verification. Indicates the measurement point for direct tripping evidence At the candidate trigger time The normalized perturbation at point is a dimensionless quantity; Represents auxiliary machine objects The indirectly corresponding set of delay response evidence measurement points includes measurement points related to the operating status of standby equipment, standby equipment current, valve action, main parameter disturbance and recovery process; Represents the set of evidence measurement points for delayed response. Any measuring point in the range; Indicates the evidence measurement points for delayed response The evidence weight is a dimensionless quantity, and the evidence weight is determined by the interlocking relationship between the delayed response measurement point and the auxiliary machine tripping event, the response stability, and the consistency of historical verification. Indicates the evidence measurement points for delayed response The response lag relative to the candidate trigger time, in units of time, is determined by interlocking logic, similar historical events, or the sequential relationship of curves within the same event. Indicates the evidence measurement points for delayed response The normalized perturbation at the response lag time is a dimensionless quantity; Indicates the candidate trigger time The timing consistency factor, a dimensionless quantity, is used to determine whether the auxiliary equipment status reversal, drive current change, outlet pressure or related flow change, standby equipment response, and main parameter disturbance conform to the propagation sequence after the auxiliary equipment trip. Its calculation method is as follows: A time check sequence is set, for example, requiring changes in direct trip evidence. Prior to this, there was a delayed response to changes in evidence. Following this, and with the main parameter perturbation evidence changing even later; check whether the changes in all evidence measurement points conform to this preset sequence, and the proportion of sequences that conform is [the percentage of such sequences]. The value of is in the range [0,1]. Indicates the candidate trigger time The conflict penalty factor at the candidate time point is a dimensionless quantity used to reflect situations such as inconsistency between objects, opposite directions of change, no response from backup equipment, missing master parameter disturbances, or contradictory curve evidence. Its calculation method is as follows: Statistical analysis is performed at the candidate time point... The number of contradictory pieces of evidence found nearby, such as: a measuring point showing the auxiliary machine status as "running" but the current has dropped to zero; standby equipment with interlocking logic requirements failing to respond within the expected time; direct evidence of changes contradicting physical principles (e.g., pressure rising after tripping). For each type of contradiction, Add a fixed penalty value, which is 0 if there is no contradiction; This represents the trigger positioning stability factor, which is a dimensionless quantity used to keep the denominator stable based on the data integrity, time synchronization reliability, and curve availability near the current candidate trigger time. Furthermore, Indicates the measurement point for direct tripping evidence At the candidate trigger time The normalized perturbation at the location. Indicates the evidence measurement points for delayed response The normalized disturbance at the response lag time is determined according to the aforementioned calculation method of the normalized disturbance of the local reference at the measurement point. That is, it is formed by the conversion of the actual sampled value, local stability reference, local fluctuation amplitude and minimum resolvable change within the same measurement point. Therefore, it is a dimensionless quantity. The formula for determining the actual trip trigger location in step S303 is illustrated below with an example of the specific scenario of "#1 induced draft fan A tripping": A thermal power unit experienced a "#1 induced draft fan A trip" event. The large model identified the auxiliary machine object as ID_FAN_1A from the operation log, and the event action was tripping or abnormal shutdown. The log recorded the tripping time as follows: The system constructs a set of candidate trigger times: ,in, , , ,and The above candidate times are only used to illustrate the process of formula substitution and do not limit the range of candidate times for different units, different auxiliary machines or different DCS sampling periods. To facilitate the explanation of the formula substitution process, the weighted term for direct trip evidence is defined as follows: ; The weighted term for delayed response evidence is defined as follows: ; The comprehensive evaluation result of the candidate triggering time is defined as: ; The normalized perturbation at each measuring point is calculated according to the following formula: ; For candidate time In the DCS curve, the operating state of #1 induced draft fan A remains running, without any reversal from running to stopped. The current and outlet air pressure also show only slight fluctuations. Therefore, although the measurement point corresponding to this candidate moment belongs to ID_FAN_1A, it meets the semantic gating requirements for auxiliary machine objects. ; However, it does not meet the action change requirements of "tripping, stopping, or abnormal shutdown," therefore, the semantic gating of tripping actions does not provide support. ; Substitute the comprehensive evaluation results: ; For candidate time and #1 induced draft fan A is already in a stopped state, but the reversal from running to stopped does not occur at this candidate moment; rather, it is a continuation of the state after the trip. Although some curves may still have significant disturbances at this time, these disturbances correspond to the subsequent response after the trip, not the instant of the trip trigger. Therefore, the trip action semantic gating does not support this candidate moment or provides weaker support than the actual trigger moment, denoted as: , ; therefore: ; ; thus, , ,and Since there is no corresponding trip trigger moment, there is no need to continue calculating its delay response evidence items; the system focuses on candidate moments that satisfy the action gating. Perform a complete substitution calculation.

[0030] Calculate candidate time : The set of test points for direct tripping evidence is as follows: ; , ,and ; For the operating status measurement point of #1 induced draft fan A: , , ,and ; therefore: ; For the current measuring point A of #1 induced draft fan: , , ,and ; therefore: ; For the outlet air pressure measurement point: , , ,and ; therefore: ; Substitute the weighted terms of the direct trip evidence: ; Calculate candidate time Evidence of delayed response: The set of evidence measurement points for delayed response is as follows: ; The instance weight is: , The response lag time is: ; For the standby #1 induced draft fan B operating status measurement point, the sampled value at the response lag time is 1, the local stability benchmark before the response is 0, the local fluctuation amplitude is 0, and the minimum resolvable change is 1, then: ; For the furnace negative pressure measuring point, the sampled value at the response lag time is -200 Pa, the local stability benchmark before the response is -50 Pa, the local fluctuation amplitude is 10 Pa, and the minimum resolvable change is 1 Pa. Then: ; Substitute the delayed response evidence weighting: ; Calculate candidate time Overall evaluation results: At that time, the curve measurement point object is consistent with the ID_FAN_1A identified by the large model, therefore: ; The reversal of operating status, sudden drop in current, and change in outlet air pressure all conform to the tripping action, therefore: ; The sequence of curve changes conforms to the propagation logic of "#1 induced draft fan A trips, furnace negative pressure disturbance, standby #1 induced draft fan B interlocks and starts, furnace negative pressure gradually recovers," therefore: ; No inconsistencies in objects, opposite directions of change, or unresponsive backup equipment were found. Therefore: ; To illustrate the formula substitution process, an example stable quantity is taken as the trigger location stabilization factor: ; Substitute into the comprehensive evaluation formula: ; Determining the actual trip trigger location: The comprehensive evaluation result of each candidate trigger moment is as follows: ; ; ; ; therefore: ;Right now: ; Therefore, although the trip time recorded in the operation log is However, the DCS curve is in Evidence has emerged including the reversal of the operating status of #1 induced draft fan A, a sudden drop in current, changes in outlet air pressure, and the subsequent interlock response of the backup #1 induced draft fan B. Therefore, the system will... The actual tripping trigger location has been corrected, instead of directly using the log recording time as the final annotation starting point.

[0031] In practical implementation, the key parameters are determined as follows: Weight of evidence , Based on a historically labeled sample database, feature importance analysis algorithms (such as feature importance ranking based on tree models or coefficient analysis of linear models) are used for learning. In the initial stage, the weighting can be set according to domain knowledge (such as direct state evidence having a higher weight than indirect parameter evidence), and dynamically adjusted based on subsequent review results.

[0032] Response hysteresis The theoretical lag time can be determined by analyzing the interlocking logic diagram recorded by the DCS system, or by statistically determining the average time delay of the standby equipment response curve relative to the trip trigger point in multiple historical trip events of the same auxiliary equipment type.

[0033] S400, Operating Condition Stage Division and Interlocking Response Chain Construction: Based on the event semantic framework and the curve change sequence, an interlocking response chain is constructed from trip triggering to recovery, and the continuous curve segment is dynamically divided into six time-progressive operating condition stages. S401. Based on the structured trip event semantic framework and the curve change order in the DCS candidate curve set, construct a full-process interlocking response chain from trip triggering to recovery; The interlocking response chain constructed in S401 is formed based on the interlocking description in the structured trip event semantic framework and the sequence of curve changes in the DCS candidate curve set. The interlocking response chain includes: auxiliary machine tripping node, standby equipment start-up node, related valve action node, medium pressure or flow recovery node, main parameter disturbance node, manual handling node, and stabilization recovery node; the interlocking response chain determines the propagation path of the auxiliary machine tripping event from triggering to recovery, and the propagation path is used as the basis for segmenting and labeling curve segments; The auxiliary equipment tripping node is determined by the actual tripping trigger position determined by S300; the standby equipment start-up node is determined by the response changes in the standby equipment operating status curve or standby equipment current curve; the relevant valve action node is determined by the state changes in the valve action curve; the medium pressure or flow recovery node is determined by the recovery trend in the pressure curve or flow curve; the main parameter disturbance node is determined by the continuous response of steam pressure, water level, air volume, furnace negative pressure, load, or combustion parameters; the manual handling node is determined by the operation record and the corresponding curve changes; the stabilization recovery node is determined by the position where the affected main parameter returns to a stable operating state.

[0034] S402. Divide the continuous curve segment corresponding to the auxiliary machine trip along the time axis into the pre-trip symptom segment, trip trigger segment, interlock response segment, main parameter disturbance segment, manual handling segment, and recovery stability segment; The division of curve segments in S402 is not based on fixed time windows, but rather dynamically determines the marking boundaries according to the actual curve changes of the auxiliary equipment tripping event. Specifically, the pre-trip symptom segment is determined by the location where the auxiliary equipment state has not yet reversed but abnormal changes have occurred in current, pressure, flow, or related process parameters; the trip trigger segment is determined by the location where the auxiliary equipment state reverses or current suddenly drops; the interlock response segment is determined by the location where standby equipment starts, valves operate, or interlock signals change; the main parameter disturbance segment is determined by the location where load, steam pressure, water level, air volume, furnace negative pressure, or combustion parameters produce continuous responses; the manual handling segment is determined by the curve fluctuation characteristics near the corresponding time in the operation record; and the recovery to stability segment is determined by the location where the affected main parameters return to a stable operating state. Furthermore, when dividing continuous curve segments into stages, the stage boundaries are determined based on the degree of deviation of evidence between adjacent stages, the degree of alignment of nodes in the interlocking response chain, and the degree of stage confusion, specifically using the following formula: ; in, Indicates the operating condition stage and subsequent working conditions The dynamic boundary markers between them, in units of time; This indicates the current working condition stage to be divided; Indicates the current operating condition stage Adjacent and subsequent operating phases; Indicates the boundary time point of the candidate stage, in units of time; Indicates the operating condition stage and subsequent working conditions The candidate boundary search interval is defined by the adjacent stage transition nodes in the interlocking response chain, rather than being arbitrarily searched on the complete event timeline. Indicates the operating condition stage At candidate time points The strength of the dominant evidence at a given stage is a dimensionless quantity. The strength of the dominant evidence at a given stage is determined by the stage of operation. The normalized perturbation quantities in the corresponding feature measurement point set are weighted and formed. The determination process is as follows: First, determine the corresponding feature measurement point set according to the physical meaning of the current working condition stage. Then, integrate the normalized perturbation quantities of each feature measurement point in the set at the candidate time point according to the degree of correlation between the measurement point and the current stage to obtain the dominant evidence strength of the stage at the candidate time point. Indicates the subsequent operating phase At candidate time points The strength of the dominant evidence at a given stage is a dimensionless quantity, and its determination method is similar to... They are consistent, all formed by the combined normalized perturbations of characteristic measurement points at the corresponding stages. Because... and Both are formed by dimensionless normalized perturbations, so the two can be directly multiplied, differed, and compared; The minimum distinguishable factor at the stage boundary is a dimensionless quantity. This minimum distinguishable factor is determined by a dimensionless stable quantity formed by the sampling period, time synchronization accuracy, and curve noise level, and is used to avoid instability in the boundary evaluation results due to excessively small differences in the evidence strength between adjacent stages. The boundary calculation stability factor is a dimensionless quantity used in... or When the denominator is weak, keep the denominator stable to avoid distortion of calculation results due to insufficient stage evidence or low degree of fit of interlocking nodes; Indicates candidate time points The degree of alignment with the corresponding stage transition node in the interlocking response chain is a dimensionless quantity. This degree of alignment is used to determine whether the candidate time point is close to the stage transition node determined by the interlocking response chain, such as the transition from the trip trigger node to the interlocking response node, the transition from the interlocking response node to the main parameter disturbance node, the transition from the manual handling node to the recovery and stabilization node, etc. Indicates candidate time points The nearby stage confusion is a dimensionless quantity used to reflect whether there are other stage features interfering with the evidence of the current stage and the evidence of the subsequent stages, or whether there are multiple overlapping stage features near the current candidate boundary that make the boundary difficult to distinguish. It should be noted that, The smaller the value, the closer the evidence for the current stage and subsequent stages are to converging. The larger the value, the more obvious the stage evidence is around the candidate time point, both in the current stage and in subsequent stages. The larger the value, the closer the candidate boundary is to the stage transition node in the interlocking response chain; The larger the value, the more pronounced the stage confusion near the candidate boundary. By minimizing the above deviation term, we can achieve better results in the candidate boundary search interval. The location where adjacent stage evidence converges significantly, interlocking nodes are closely aligned, and stage confusion is low is selected as the dynamic stage boundary. The characteristic measurement point set for the pre-trip symptom stage includes current, pressure, flow, or related process parameters that have shown abnormal changes before the auxiliary equipment status has reversed; the characteristic measurement point set for the trip trigger stage includes measurement points for operating status reversal, sudden current drop, or sudden outlet pressure change; the characteristic measurement point set for the interlock response stage includes measurement points for standby equipment startup, standby equipment current change, valve action, and interlock signal change; the characteristic measurement point set for the main parameter disturbance stage includes measurement points for steam pressure, water level, load, furnace negative pressure, or combustion parameters; the characteristic measurement point set for the manual handling stage includes the control object and its curve response measurement points in the operation record; and the characteristic measurement point set for the recovery stabilization stage includes relevant measurement points during the process of the affected main parameters returning to stability.

[0035] Regarding the division of operating conditions and construction of the interlocking response chain in step S400, the following example will be used to illustrate the "#1 induced draft fan A trip" event in S303: In this example, S303 has determined the actual trip trigger location based on semantic gating, direct trip evidence, delayed response evidence, timing consistency, and conflict penalty: Although the trip time recorded in the operation log is: However, the DCS curve is in Since the #1 induced draft fan A has experienced a reversal in its operating status, a sudden drop in current, and a decrease in outlet air pressure, the log recording time is no longer used as the starting point of the interlock response chain in S400. Instead, the actual trip trigger position determined by S303 is used as the auxiliary machine trip node. Constructing the interlocking response chain: Based on the structured trip event semantic framework and the change sequence of the DCS curve, the system sequentially identifies the following nodes: To facilitate the description of the temporal relationship of the interlocking response chain, the auxiliary machine trip node, standby equipment start node, regulating mechanism action node, main parameter disturbance node, and recovery stability node in the interlocking response chain are sequentially denoted as follows: , , , ,and ; The auxiliary equipment tripping point is: This node is determined by the actual trip trigger position determined by S303, corresponding to the change of the operating status of #1 induced draft fan A from running to stopping, the sudden drop in the current of #1 induced draft fan A, and the start of a decrease in the outlet air pressure. The backup device startup node is: This node is determined by the change in the operating status curve of #1 induced draft fan B from stopped to running and the rise of the current curve of #1 induced draft fan B from a low state to the start-up running state. It is used to characterize the interlocking start of the standby #1 induced draft fan B after the trip of #1 induced draft fan A. The adjustment mechanism's action nodes are: This node is determined by the "adjusting the induced draft fan blade opening" in the operation log and the corresponding change in the induced draft fan blade opening curve. For example, if the DCS curve shows that the induced draft fan blade opening changes from the original operating opening to the new adjusted opening, then this node is identified as a manual handling or adjustment mechanism action node. The main parameter perturbation node is: This node is determined by the continuous changes in furnace negative pressure, inlet air volume, outlet air pressure, or oxygen content curves. It should be noted that furnace negative pressure may begin to fluctuate immediately after a trip, but in the stage division, the position where the evidence of the main parameter disturbance becomes dominant relative to the evidence of the interlock response is taken as the main parameter disturbance node or the starting point of the main parameter disturbance segment. The stable node to be restored is: This node is determined by the main parameters such as furnace negative pressure, inlet air volume, outlet air pressure, or oxygen content returning to a stable state. For example, since... As the furnace negative pressure gradually returns to a stable operating range and subsequent fluctuations decrease, this position is taken as the starting reference node for the recovery to stability; the resulting interlocking response chain is as follows: to to to to That is: from the tripping of #1 induced draft fan A to the interlocking start of #1 induced draft fan B to the adjustment of the induced draft fan blades until the furnace negative pressure disturbance becomes dominant until the parameters return to stability; Based on the above interlocking response chain, the system divides the continuous curve segment corresponding to the "#1 induced draft fan A trip" event into multiple time-progressive operating condition stages: the pre-trip symptom stage can be determined as: ; It should be noted that the starting point of the pre-trip symptom segment is not based on a fixed time length but rather on the actual trip trigger position determined by S303. Looking back, assuming that the operating state of #1 induced draft fan A has not yet reversed, the determination is based on the continuous deviation of precursor measurement points such as the current of #1 induced draft fan A, outlet air pressure, inlet air volume, or furnace negative pressure relative to the local stability benchmark. For example, in this example, the system is triggered from the actual trip location: ; Tracing back the DCS curve, if in and Nearby, the current and outlet air pressure of #1 induced draft fan A remain within a locally stable fluctuation range; while from Initially, the current and outlet air pressure of #1 induced draft fan A began to deviate continuously from their respective local stability references, and this deviation continued until... Near the state reversal time, then It was determined to be the starting point of the pre-trip symptom segment.

[0036] The characteristic measurement points in this stage include the current of #1 induced draft fan A, the outlet air pressure, the inlet air volume, and the furnace negative pressure. If, during this stage, the current of #1 induced draft fan A and the outlet air pressure deviate only slightly from the local stability reference, but the operating state of #1 induced draft fan A has not yet changed from running to stopped, then this stage is marked as the pre-trip symptom stage, rather than the trip trigger stage. For example, near the candidate point: , ; The above disturbance only indicates a slight deviation in the curve. Since no state reversal occurs, the trip action gating does not provide support, so this stage is not considered the actual trip trigger position. The trip trigger segment can be determined as follows: The characteristic measurement points in this stage include the operating status of #1 induced draft fan A, the current of #1 induced draft fan A, and the outlet air pressure. During this stage, the operating status of #1 induced draft fan A changes from running to stopping, the current suddenly drops from the current level corresponding to the operating range to the current level corresponding to the stopping range, and the outlet air pressure shows a significant decrease. For example, at the actual trip trigger location Location: , , ; The above characteristics collectively indicate that this stage is the trip trigger stage; The interlocking response segment can be determined as follows: The characteristic measurement points in this stage include the operating status of #1 induced draft fan B, the current of #1 induced draft fan B, and the operation curves of relevant dampers or moving blades. During this stage, #1 induced draft fan B changes from being stopped to running, the current of the standby fan increases, and the relevant regulating mechanisms begin to respond.

[0037] For example, the disturbance in the operating state of the standby #1 induced draft fan B is: ; The standby fan current and blade operation curves are also calculated according to the local stability benchmark, local fluctuation amplitude and minimum resolvable change of their respective measuring points to determine the corresponding normalized disturbance quantity, which is used to characterize the interlocking response intensity. The main parameter disturbance segment can be determined as follows: The characteristic measurement points in this stage include furnace negative pressure, inlet air volume, outlet air pressure, and oxygen content. Although the furnace negative pressure can begin to change after a trip, in the dynamic stage division, the starting point of this stage is defined as the position where the evidence of the main parameter disturbance becomes dominant relative to the evidence of the interlock response.

[0038] For example, the negative pressure measuring point in the furnace at a location with significant disturbance can be calculated as follows: ; This disturbance indicates that the furnace negative pressure has deviated significantly from the local stable operating level before the trip, and the disturbance lags behind the trip of #1 induced draft fan A and the start-up of standby #1 induced draft fan B, which conforms to the propagation sequence of "trip trigger - interlock response - main parameter disturbance"; The manual treatment section can be defined as follows: The characteristic measurement points in this stage include the opening degree of the induced draft fan blades, the opening degree of the inlet damper, or the set value of the furnace negative pressure control. This stage may overlap with the main parameter disturbance recovery process in time, but when the tag is output, manual handling can be recorded as a stage tag or auxiliary tag.

[0039] For example, the operation log shows that the operators were... When the opening of the induced draft fan blades is adjusted, the DCS curve shows that the blade opening changes from the original opening to the new adjusted opening, and the furnace negative pressure gradually rises afterward. The system then marks the curve segment near this time as the manual handling segment. The recovery and stabilization phase can be identified as follows: The characteristic measurement points at this stage include furnace negative pressure, inlet air volume, outlet air pressure, oxygen content, and the current of #1 induced draft fan B, etc. If from... Once the aforementioned main parameters and standby equipment operating parameters enter a stable state and no further significant disturbances related to the tripping event occur, this stage is marked as the recovery and stabilization phase. The following example, using the boundary determination from the "interlock response segment to the main parameter disturbance segment," illustrates the calculation process of the dynamic boundary formula in S402: In the above interlock response chain, the start-up node of the standby #1 induced draft fan B is: ; The dominant position for evidence of principal parameter perturbation is after this node; therefore, the system determines the candidate boundary search interval within the range of curve variations between and near this node. ; Within the candidate boundary search interval, the system generates a set of candidate boundary time points based on the DCS curve sampling period, the trend of evidence strength changes in adjacent stages, and the curve change position near the stage transition node of the interlocking response chain: ; in: ; ; ; Among them, the current stage This is the interlocking response section, the subsequent stage. Main parameter disturbance segment; Strength of dominant evidence in the interlocking response phase The strength of the dominant evidence for the stage of the main parameter disturbance segment is formed by comprehensively considering the operating status of the standby #1 induced draft fan B, the current of the standby #1 induced draft fan B, and the movement curve of the blades. It is formed by combining the furnace negative pressure, inlet air volume, outlet air pressure, and oxygen content curves; In this example, the strength of the dominant evidence for the three candidate boundary time points is as follows: , ; , ; , ; The degree of fit between the candidate time points and the corresponding stage transition nodes in the interlocking response chain is as follows: , ,and ; The stage confusion near the candidate boundary is: , ,and ; To facilitate the explanation of the formula substitution process, the example uses: , ; To facilitate the explanation of the calculation process for the stage boundary, the boundary deviation evaluation value, which is jointly formed by the degree of deviation of evidence between adjacent stages, the degree of fit between nodes in the interlocking response chain, and the degree of stage confusion, is defined as: ; That is: when hour, ; when hour, ; when hour, ; so ; The system will It is determined to be the dynamic boundary between the interlocking response segment and the main parameter disturbance segment; Therefore, we can conclude that: Interlocking response segment = ; Main parameter disturbance segment = ; As can be seen from the above example, S400 does not redetermine the trip trigger time, but rather follows the actual trip trigger position determined by S303: Using this location as the auxiliary machine tripping node, the standby equipment start-up node, the regulating mechanism action node, the main parameter disturbance node, and the recovery and stabilization node are constructed in sequence to form a complete interlocking response chain.

[0040] Meanwhile, the stage boundaries in S402 are not selected based on fixed time windows, nor are they simply taken at the start-up time of standby equipment or the lowest point of furnace negative pressure. Instead, they are determined within the candidate boundary search interval by jointly considering the degree of deviation between adjacent stage evidence, the degree of fit between interlocking response chain nodes, and the degree of stage confusion. Therefore, the "#1 induced draft fan A trip" event can be dynamically divided into the pre-trip symptom stage, the trip trigger stage, the interlocking response stage, the main parameter disturbance stage, the manual handling stage, and the recovery and stabilization stage, ensuring that each stage has clear textual, graphical, and computational basis.

[0041] S500, multi-level annotation result output: Generate event main label, stage label and impact label for each working condition stage, and output structured annotation results containing label and time interval information; S501. For each completed working condition stage, generate the corresponding event main label, stage label, and impact label; The main event label in S501 is used to characterize the type of auxiliary machine tripping event; the stage label is used to characterize the development stage of the curve segment in the entire process of auxiliary machine tripping, including at least one of "symptom period", "trigger time", "interlocking period", "disturbance peak period", "artificial intervention period", and "restoration stabilization period"; the impact label is used to characterize the objects affected by the auxiliary machine tripping on the generator set operating parameters, including one or more of the following: feedwater pressure fluctuation, feedwater flow disturbance, steam drum water level disturbance, furnace negative pressure disturbance, air volume disturbance, coal volume disturbance, main steam pressure disturbance, load fluctuation, and combustion state disturbance, so that the same curve segment can simultaneously carry event category, stage position, and operation impact information; The main event label corresponds to the auxiliary machine tripping event itself, indicating which auxiliary machine tripping condition the labeled sample belongs to; the stage label corresponds to each operating condition stage divided in S402, indicating the position of the curve segment in the entire auxiliary machine tripping process; the impact label corresponds to the operating parameters of the affected generator set, indicating the specific impact of the auxiliary machine tripping event on the unit's operating status.

[0042] S502. Output the annotation results containing the above labels and their corresponding time interval information.

[0043] The annotation results output in S502 are in the form of structured samples. The structured samples include: event number, auxiliary machine object, actual trip trigger position, curve segment start and end position, event main label, stage label, impact label, label source and review status. The label sources include log parsing source, alarm record source, operation record source, curve alignment source and interlock response source. The generated annotation data is used to build an abnormal operating condition sample library, train a fault diagnosis model, perform operation review analysis and train an operating condition identification model.

[0044] The embodiments disclosed in this invention are preferred embodiments, but are not limited thereto. Those skilled in the art can easily understand the spirit of this invention based on the above embodiments and make different extensions and variations, but as long as they do not depart from the spirit of this invention, they are all within the protection scope of this invention.

Claims

1. A method for intelligent annotation of power generation operating condition data assisted by a large model, characterized in that, Includes the following steps: Based on the auxiliary machine objects in the structured trip event semantic framework generated by the large model, match the DCS candidate curve set corresponding to the auxiliary machine objects; Using the trip event time recorded in the log as the initial location basis, extract evidence of at least one curve change from the DCS candidate curve set, including equipment operating status reversal, drive current sudden drop, outlet pressure change, related flow change, standby equipment response change, and main parameter disturbance. The auxiliary machine objects and tripping actions identified by the large model are used as semantic gating conditions. Combined with direct tripping evidence, delayed response evidence, timing consistency and conflict penalty, the actual tripping trigger position after curve alignment correction is determined. Using the actual trip trigger location as the auxiliary machine trip node, and based on the semantic framework of the structured trip event and the relationship between the curve changes in the DCS candidate curve set, an interlocking response chain from trip trigger to recovery of stability is constructed. In determining the actual trip trigger position after curve alignment correction, the matching result of the auxiliary machine object identified by the large model is used as the semantic gating factor of the auxiliary machine object, and the matching result of the curve change direction and the trip action is used as the semantic gating factor of the trip action. The normalized disturbance weighted value of the direct trip evidence measurement point set and the normalized disturbance weighted value of the delayed response evidence measurement point set are gating and verified respectively. Combining the timing consistency factor, the trigger positioning stability factor and the conflict penalty factor, the candidate moment with the highest comprehensive score is selected as the actual trip trigger position. The actual trip trigger position is determined according to the following formula: ； in, This indicates the actual trip trigger position determined after curve alignment correction; This refers to auxiliary machine objects identified by the large model from operation logs, alarm records, or operation records; Indicates the candidate trigger time; Indicates the auxiliary object The corresponding set of candidate trigger times; This represents the semantic gating factor of the auxiliary machine object; Indicates the semantic gating factor of the tripping action; Indicates the auxiliary object The set of measurement points directly corresponding to the tripping evidence; Represents the set of measurement points for direct tripping evidence. Any measuring point in the range; Indicates the measurement point for direct tripping evidence Weight of evidence; Indicates the measurement point for direct tripping evidence At the candidate trigger time The normalized perturbation at the location; Indicates the auxiliary object The set of indirect corresponding delay response evidence measurement points; Represents the set of evidence measurement points for delayed response. Any measuring point in the range; Indicates the evidence measurement points for delayed response Weight of evidence; Indicates the evidence measurement points for delayed response The response lag relative to the candidate trigger moment; Indicates the evidence measurement points for delayed response The normalized perturbation at the response lag time; Indicates the candidate trigger time Temporal consistency factor at the location; Indicates the candidate trigger time Conflict punishment factor at the location; This indicates that the location stabilization factor has been triggered. Based on the degree of deviation of evidence between adjacent stages, the degree of fit of interlocking response chain nodes, and the degree of stage confusion, the continuous curve segment corresponding to the auxiliary machine tripping is dynamically divided into multiple time-progressive operating condition stages.

2. The intelligent annotation method for power generation condition data assisted by a large model as described in claim 1, characterized in that, The DCS candidate curve set is determined by the auxiliary machine object, auxiliary machine type, interlock response description, and affected operating parameters in the structured trip event semantic framework. For different types of auxiliary equipment, the DCS candidate curve set should at least include the body monitoring dimension curves that reflect the operating status or changes in drive parameters of the auxiliary equipment, the process parameter dimension curves that reflect changes in process medium pressure, flow rate, air volume, temperature or water level, the interlock response dimension curves that reflect the activation of standby equipment and the action of valves, guide vanes or baffles, and the system impact dimension curves that reflect the disturbance of the unit's operating status.

3. The intelligent annotation method for power generation operating condition data assisted by a large model as described in claim 1, characterized in that, In the process of dynamically dividing the operating condition into multiple time-sequential progressive stages, the operating condition stages include at least the pre-trip symptom stage, the trip trigger stage, the interlock response stage, the main parameter disturbance stage, the manual handling stage, and the recovery and stabilization stage. The dynamic annotation boundary between adjacent working condition stages is determined by the difference in the strength of dominant evidence between adjacent stages, the minimum distinguishable factor of the stage boundary, the stability factor of the boundary calculation, the stage confusion degree, and the degree of fit of the interlocking response chain nodes. The candidate time point with the smallest boundary evaluation value is selected as the dynamic annotation boundary.

4. The intelligent annotation method for power generation condition data assisted by a large model as described in claim 1, characterized in that, Before determining the actual trip trigger location, the curve changes of each measuring point in the DCS candidate curve set are first converted into dimensionless normalized disturbance quantities. The normalized disturbance is calculated based on the actual sampled value of the measuring point at the candidate time point, the local stability benchmark before the candidate time point, the local fluctuation amplitude, and the minimum resolvable change of the measuring point.

5. The intelligent annotation method for power generation condition data assisted by a large model according to claim 4, characterized in that, The normalized perturbation is determined according to the following formula: ； in, Indicates the first Each DCS measurement point at the candidate time point The actual sampled value at that location; Indicates the first Each DCS measurement point at the candidate time point Previous local stability benchmark; Indicates the first Each DCS measurement point at the candidate time point The previous local fluctuation range; Indicates the first The minimum resolvable change at each DCS measurement point; Indicates the first Each DCS measurement point at the candidate time point The normalized perturbation at the location.

6. The intelligent annotation method for power generation operating condition data assisted by a large model according to claim 1, characterized in that, The interlock response chain is formed based on the interlock description in the structured trip event semantic framework and the sequence of curve changes in the DCS candidate curve set. It includes at least the auxiliary machine trip node, the standby equipment start node, the relevant valve action node, the medium pressure or flow recovery node, the main parameter disturbance node, the manual handling node, and the recovery and stabilization node. Among them, the auxiliary equipment tripping node is determined by the actual tripping trigger position; the standby equipment start-up node is determined by the response change of the standby equipment operating status curve or standby equipment current curve; the relevant valve action node is determined by the state change of the valve action curve; the main parameter disturbance node is determined by the continuous response of steam pressure, water level, air volume, furnace negative pressure, load or combustion parameters; the manual handling node is determined by the operation record and the corresponding curve change; and the stabilization recovery node is determined by the position where the affected main parameter returns to a stable operating state.

7. The intelligent annotation method for power generation condition data assisted by a large model according to claim 1, characterized in that, Before matching the DCS candidate curve set, the following is also included: The system collects operation logs, alarm records, operation records, interlock action records, and DCS historical curve data generated during the operation of the generator set. It also standardizes and unifies the equipment name, measurement point name, event time, and curve time in the collected data to form unified operating condition data to be labeled. The large language model, fine-tuned by text in the power generation field, is invoked to perform semantic parsing on the operation logs, alarm records, and operation records. From the parsing results, auxiliary equipment objects, tripping actions, interlocking descriptions, standby equipment activation descriptions, manual handling descriptions, and recovery descriptions are identified, and the structured tripping event semantic framework is generated based on the identification results.

8. The intelligent annotation method for power generation operating condition data assisted by a large model according to claim 5, characterized in that, After completing the division of working conditions into stages, corresponding event main labels, stage labels, and impact labels are generated for each working condition stage, and structured annotation results containing the labels and their corresponding time interval information are output. Among them, the main event label is used to characterize the type of auxiliary machine tripping event, the stage label is used to characterize the development stage of the curve segment in the entire process of auxiliary machine tripping, and the impact label is used to characterize the object of the impact of auxiliary machine tripping on the generator set operating parameters. The structured annotation results should include at least the event number, auxiliary machine object, actual trip trigger location, start and end positions of the curve segment, main event label, stage label, impact label, label source, and review status.

Images