A kind of real-time device performance monitoring method based on equipment working condition variationTechnical field
The invention belongs to nuclear power plant equipment belief system device performance monitoring technical fields, and in particular to one kind is based on settingThe real-time device performance monitoring method of standby operating condition variation.
Background technique
One of the important content of nuclear power plant equipment reliability management is exactly to carry out equipment to the important key equipment of nuclear power stationPerformance monitoring, and real-time device performance monitoring is a kind of most common mode of device performance monitoring, real-time device performance monitoringIt is exactly the fault mode and phenomenon of bonding apparatus, chooses the parameter for being actually able to reflect equipment state, such as: temperature, pressure, streamAmount, voltage, vibration, electric current, oil level, oil etc. are set by it being arranged corresponding alarm, early warning value (i.e. threshold value) to realizeStandby real-time monitoring, monitoring parameters meet or exceed threshold value, that is, alarm are generated, to remind engineer to need to pay close attention to, check that this setsStandby virtual condition, confirmation equipment whether there is failure, and take necessary measure according to actual needs, to prevent the equipment straightThe normal operation for influencing power station is connect, in order to avoid cause power station non-programmed halt shutdown or other operation consequences.
The threshold value of real-time device performance monitoring at present is arranged, usually by engineer according to operation alarming value, the relevant technologiesStandard, specification etc. are formulated, or are determined according to engineer experience, using fixed threshold set-up mode.With power station unitOperating status variation, monitoring parameters are larger in the difference of the various operating statuses of unit, are unable to the practical shape of real reaction equipmentState, the pre-/alarm sometimes, pre-/alarm cannot be generated in time, be not necessarily to generate pre-/alarm sometimes, produced pre-/alarm instead,Interference is caused to the work of engineer.For this purpose, the threshold value of dynamic change is to equipment real-time energy according to the operating status of unitMonitoring just becomes especially important.
Summary of the invention
The equipment real-time energy monitoring method based on dynamic threshold that it is an object of that present invention to provide a kind of, to large number of equipmentHistory data obtains the dynamic threshold for being able to reflect equipment status parameter by big data analysis method, makes up existing setThe deficiency of standby parameter fixed threshold setting enables Facilities Engineer to carry out accurately real-time monitoring to each parameter of equipmentAccurately confirm equipment state, and then accurately carries out equipment fault diagnosis.
The technical solution of the invention is as follows:
1. step 1: data acquisition includes the following steps:
Step 1: user is by the measuring point in measuring point coding, measuring point title, instrument keyword query real-time data base,Export measuring point data；
Step 2: the historic defects information of collecting device, the Historical Monitoring report of equipment；
Step 3: being saved, being deleted to the measuring point found, the routine operations such as transform sequence.
2. step 2: data prediction includes the following steps:
Step 1: carrying out exceptional value lookup, the scope of period, including ocean temperature, environment temperature, humidity；It adoptsWith the correlation analysis of linear single order, the correlation analysis of measuring point and the establishment of weight are realized, reject the ginseng of completely no correlationNumber；Rules Filtering, the error that eliminating system is generated by uncertain factor；Screened using trend, reject mutation, it is irregular fluctuation,The period of substantial deviation floor data；Fixed threshold range is defined, it is made to be bonded operating condition completely；Start and stop rule is set；
Step 2: model pre-warning effect detection screens defect information according to (2) step in the first step；RetainThe information of early warning judgement can be carried out by Parameters variation；History parameters are extracted, the analog detection of model is carried out；According to testing resultTo optimize and revise the algorithm of model；
Step 3: being determined whether to ignore the outlier according to field conditions and historical record by operator, simultaneouslyOperator carries out the adjustment of data time interval according to the situation of abnormal point；If judging, abnormal point is caused by measuring instrumentss failure,Then the data are deleted, a stepping row data missing values of going forward side by side are automatically replenished；If judging, abnormal point for equipment fault, is mentionedThe outlier is taken, further analysis processing is remained；
Step 4: shortage of data value caused by ignoring to abnormal data in above-mentioned third step is automatically replenished, before missingData measured value afterwards, setting take most dirty data automatically to fill or be averaged and fill function.
3. step 3: feature extraction includes the following steps:
Step 1: according to the design data of equipment, supervision document, drawing delimit parameter area, herein within the scope of carry outParameter is chosen, including equipment design parameter, equipment safety characterisitic parameter, equipment running status parameter；
Step 2: handled by the step (1) in data prediction by parameter normalization, take deviation percent orUnified two kinds of methods of percentage are completed, the determination of Lai Shixian weight；Operator can suitably be adjusted parameters weightingTo verify validity, the correctness of model；
4. step 4: module training includes the following steps:
Step 1: the equipment for there is historical failure data, using machine learning and data mining based on data-drivenAlgorithm is modeled, once actual monitoring data fit model alarming value, system make device failure alert automatically, and according to buildingWhen mould determine fault mode automatic prompt equipment fault type, while can also according to model automatic Prediction failure occur whenBetween, and make prompt；
Step 2: the equipment of not historical failure data is operated normally data using history and is built to carry out machine learningMould, the model are the sections of a normal operation, and beyond operating normally, then system makes unit exception warning automatically.
5. step 5: iteration optimization model includes the following steps:
Step 1: according to the defect information reported by mistake and failed to report, repeated data pretreatment is generated in model actual moving processIn step (1), step (3), execute optimize and revise model calculating dynamic threshold range, if model generate wrong report, sieve againThe time segment limit of the historical data under similar operating condition is selected, or adjusts the bound of dynamic threshold range according to the actual situation；
Step 2: adjusting the range, quantity, weight of modeling parameters according to the early warning effect in model actual moving processRealize the effect of simplified model；
6. step 6: model application includes the following steps:
Step 1: the dynamic threshold model that big data analysis is obtained puts into application；
Step 2: system becomes automatically according to the time model of inside configuration in conjunction with the variation of historical failure or abnormal dataGesture, the time that prediction failure occurs form anticipation trend curve, device failure alert or abnormal warning, including yellow or moreRank automatically forms event log and issues mail and prompting message automatically.
The present invention has the beneficial effect that compared with prior art
1. equipment early warning: finding that the recessive abnormal and unit performance of unit that traditional monitoring means can not be found is bad earlyThe slight change of change.It can decompose design parameter for unit exception, indicate the difference between abnormal parameters value and history optimal valueAway from.-- trend group
2. device performance monitoring is analyzed: by the dynamic threshold of real-time monitoring mechanical seal correlation measuring point, and to historyThe visualization of the comparison of optimum operating condition is presented, and makes equipment performance operation trend and Long-term change trend very clear, while providing equipmentThe development trend of following a period of time.-- dynamic threshold envelope
3. equipment trend prediction: pointing out equipment performance cracking inflection point, find that the state of measuring point trend deviates early, pass through and beStatistics calculates the time point that pre- measurement equipment concrete position will fail.-- alarm is estimated
4. device exception information tracks: abnormal development trend is tracked, by counting the association measuring point of unit exception and differentThe frequency often occurred forms systemic anomalous event.-- event analysis
Detailed description of the invention
Fig. 1 is a kind of real-time device performance monitoring method flow chart based on equipment working condition variation
Specific embodiment
In the following with reference to the drawings and specific embodiments to a kind of real-time device performance monitoring method based on equipment working condition variationIt is described in detail:
The present invention calculates to illustrate that the technical solution, target are by big with the dynamic threshold of three factory owner's pump mechanical sealing of the QinData mode excavates the personalized threshold value under equipment difference operating status, while carrying out intensity grade setting to dynamic threshold,For the subsequent requirement for meeting device performance monitoring dynamic management under different operating statuses.By mathematical modeling, establish to measuring pointThe judgement for whether meeting measuring point data variation moving law in the past established, is built for the prediction of variation tendency and future time instance numerical valueThe vertical judgement that meets situation of the mathematical model to measuring point current data.
One, data acquire:
1. user is surveyed by the measuring point in measuring point coding, measuring point title, instrument keyword query real-time data base, exportPoint data；
2. the historic defects information of collecting device, the Historical Monitoring report of equipment；
3. routine operations such as pair measuring point found is saved, deleted, transform sequences.
Two, data prediction:
1. carrying out exceptional value lookup, the scope of period, including ocean temperature, environment temperature, humidity；Using linearThe correlation analysis of single order realizes the correlation analysis of measuring point and the establishment of weight, rejects the parameter of completely no correlation；RuleScreening, the error that eliminating system is generated by uncertain factor；It is screened using trend, rejects mutation, irregular fluctuation, substantial deviationThe period of floor data；Fixed threshold range is defined, it is made to be bonded operating condition completely；Start and stop rule is set；
Wherein, the dependency analysis process of measuring point is as follows: pushing away at any time since covariance feature can embody two measuring pointsThe relative variability degree of shifting and direction, therefore using covariance numerical value as measuring point correlation judgment basis.In view of different surveysThe data dimension of point is different, and the variance product using the two is normalized covariance numerical value.So related coefficient has following spyPoint: 1. related coefficient numerical values reciteds represent the power of linearly related degree between two measuring points between [0,1]；2. relatedCoefficient symbols it is positive and negative, represent correlativity between measuring point and belong to positive correlation or negatively correlated.Related coefficient calculation formula are as follows:
2. model pre-warning effect detection screens defect information according to (2) step in the first step；Reservation can lead toCross the information that Parameters variation carries out early warning judgement；History parameters are extracted, the analog detection of model is carried out；Come according to testing result excellentChange the algorithm of adjustment model；
3. being determined whether to ignore the outlier, while operator according to field conditions and historical record by operatorThe adjustment of data time interval is carried out according to the situation of abnormal point；If judging, abnormal point is caused by measuring instrumentss failure, to thisData are deleted, and a stepping row data missing values of going forward side by side are automatically replenished；If judging, for equipment fault, it is different to extract this for abnormal pointConstant value point remains further analysis processing；
4. due in step (2) abnormal data ignore caused by shortage of data value be automatically replenished, according to missing front and back dataMeasured value, setting take most dirty data automatically to fill or be averaged and fill function.
Three, feature extraction:
1. according to the design data of equipment, supervision document, drawing delimit parameter area, herein within the scope of carry out parameter choosingIt takes, including equipment design parameter, equipment safety characterisitic parameter, equipment running status parameter；
2. handling by the step (1) in data prediction by parameter normalization, deviation percent or unified hundred are takenThan two kinds methods are divided to complete, the determination of Lai Shixian weight；Operator can suitably could be adjusted to verify to parameters weightingValidity, the correctness of model；
Four, module training:
1. for there is the equipment of historical failure data, using based on data-driven machine learning and data mining algorithm intoRow modeling, once actual monitoring data fit model alarming value, system makes device failure alert automatically, and according to true when modelingFixed fault mode automatic prompt equipment fault type, while the time that can be also occurred according to model automatic Prediction failure, and makeIt prompts out；
2. pair not equipment of historical failure data, data are operated normally using history to carry out machine learning modeling, it shouldModel is the section of a normal operation, and beyond operating normally, then system makes unit exception warning automatically.
Five, iteration optimization model:
1. the step according to the defect information reported by mistake and failed to report is generated in model actual moving process, in repeated data pretreatmentSuddenly (1), step (3) execute the dynamic threshold range for optimizing and revising model calculating, if model generates wrong report, screen again similarThe time segment limit of historical data under operating condition, or the bound of dynamic threshold range is adjusted according to the actual situation；
2. the range, quantity, weight for adjusting modeling parameters realize letter according to the early warning effect in model actual moving processChange the effect of model；
Six, model application:
The dynamic threshold model that big data analysis is obtained, in Chinese Nuclear Power equipment reliability management system (ERDB)It is monitored in three factory, the Qin, 8 main pump device performance monitorings, the time model that system is configured automatically according to inside, in conjunction with history eventThe variation tendency of barrier or abnormal data, the time that prediction failure may occur, and form anticipation trend curve, device failure alertOr abnormal warning, including yellow and the above rank, it automatically forms event log and issues mail and ERDB prompting message automatically.