CN105302848A - Evaluation value calibration method of equipment intelligent early warning system - Google Patents
Evaluation value calibration method of equipment intelligent early warning system Download PDFInfo
- Publication number
- CN105302848A CN105302848A CN201510105156.0A CN201510105156A CN105302848A CN 105302848 A CN105302848 A CN 105302848A CN 201510105156 A CN201510105156 A CN 201510105156A CN 105302848 A CN105302848 A CN 105302848A
- Authority
- CN
- China
- Prior art keywords
- measuring point
- data
- early warning
- abnormal
- assessed value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Testing And Monitoring For Control Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention belongs to the technical field of equipment state early warning, and particularly relates to an evaluation value calibration method of an equipment intelligent early warning system based on an SVM regression model. The evaluation value calibration method of the equipment intelligent early warning system can remove deviation of an abnormal detecting point with respect to the evaluation values of other normal measuring points in real time sequentially through establishment of an equipment intelligent early warning model, influence relation curve fitting, identification of abnormal measuring points and calibration of interfered evaluation values, has good universality and can be grafted to other regression algorithms for performance optimization. When the evaluation value of an interfered measuring point of equipment is adjusted on line, no excessive delay effect is generated, and a real-time early warning capability of the model for the equipment is ensured; discrimination and analysis for abnormal measuring points can be carried out in real time, so that the real-time performance of judgment is ensured, and online state analysis for equipment data can be carried out; and evaluation values without interference of the normal measuring points can be output steadily. According to the evaluation value calibration method of the equipment intelligent early warning system, the reliability of the early warning model is higher, and the life cycle is long.
Description
(1) technical field
The invention belongs to equipment state early warning technology field, particularly a kind of assessed value calibration steps of the device intelligence early warning system based on SVM regression model.
(2) background technology
As everyone knows, the running status of visual plant is huge for plant produced impact, the key equipments such as the boiler in such as Power Plant, generator.Feature under functional character, external performance and electric characteristic that equipment embodies in operational process in normal conditions and abnormality is discrepant.Device intelligence early warning refers to when part of appliance sprouts abnormal sign, sends early warning signal in advance to equipment management personnel, prevents from being extremely changed into uncontrollable fault, ensures the safety of producers, improves the serviceable life of equipment.
At present, two large class methods are mainly contained: one is adopt device fabrication manufacturer according to the threshold method of the different warning level of manufacturing process setting all parts at equipment state warning aspect, two is adopt the mathematics method for digging of non-linear multi-objective planning modeling to simulate complex redundancy relation between equipment various piece, carries out early warning by the gap analyzed between actual value and assessed value.Right latter two method all has respective feature and limitation.
For traditional fixed threshold method, the DCS system as set up in power plant can monitor the running status of key equipment in real time, and mostly these fixed thresholds are the Multi-stage alarming threshold value that sets according to the requirement of device fabrication producer.This classic method compares insurance, but the feature that equipment runnability is degenerated easily is ignored, and different parts have different degree of degenerations because of material, usage degree, the degree of wear, therefore this traditional method for early warning can often produce more false-alarm in the equipment use later stage, greatly reduces early warning efficiency.
So people more and more focus on the use of setting up nonlinear regression model (NLRM) method, its ultimate principle is that the model using the mass historical data of the abundant digging system equipment of existing Mathematical Modeling Methods to set up highly effective carries out the evaluation of equipment real-time status, and the assessment confidence level of model to equipment real-time status is the basic of equipment early warning success or not.But affect its to the evaluation confidence level of equipment early warning because have following 2 points: first, modal problem is the key parameter in different regression algorithm, such as the parameter such as activation function f type between hidden layer number of plies l, the node in hidden layer n in BP neural network, each layer; Further, the error penalty factor in support vector machine
, kernel function bandwidth
deng important parameter factor.These parameters all need to be optimized to reduce the deviation to the assessment of equipment instantaneous value on former method embedding parameter optimizing policy grounds.
In addition this another weak point setting up nonlinear regression model (NLRM) method is found in the recent period, due to close Function Mapping relation can be set up by between each for equipment parts during Nonlinear Modeling, to such an extent as to when certain parts generation pathology in equipment, the assessed value of miscellaneous part can produce departing from various degree, be difficult to the running status truly representing these parts of current device, bring certain false-alarm simultaneously, reduce the use value of this kind of intelligent early-warning method.For this problem, at present in Nonlinear Modeling warning algorithm for the affected problem of this normal component assessed value or be left in the basket or judge but do not have shaping calibration algorithm to carry out making up this defect problem from phenomenon, lack the thorough research to assessed value influenced problem mechanism of production.
Such as, paper " application of SVM regression estimation method in satellite failure diagnosis " (Electric Machines and Control, 12nd volume the 4th phase in 2008) monitoring and fault diagnosis of middle use Support vector regression method realization to satellite posture, reaching to adopting the semi definite programming method of Lancastrian to carry out core optimization in algorithm of support vector machine optimization the object reducing extensive error.Realizing adopting in multi output separately for each output parameter builds the method for regression model.The method simultaneously using segmentation training data to extract support vector carries out training data compression, and object reduces the training pattern time.But this section of article has just carried out the elaboration of phenomenon to the influenced defect problem of normal measuring point assessed value, do not explain phenomenon mechanism of production and corresponding solution, just presentation tells about the satellite component how differentiated and really break down.
The main stream approach of current raising Nonlinear Modeling algorithm early warning performance uses all kinds of parameter optimization strategy numerically to optimize, and obtains best recurrence Evaluated effect to improve equipment early warning performance.Such as paper " the SVM short-term load forecasting based on ARMA error correction and adaptive particle swarm optimization " (protecting electrical power system and control 14 phases in 2011) to the optimization limitation of Support vector regression model be ant group algorithm, particle cluster algorithm can only to kernel function bandwidth
, error penalty factor
two parameter carries out global optimizing, national patent " a kind of grain feelings prediction and warning method and system based on SVM " (application number 201410068731.X) is also had all to mention and set up Support vector regression forecast model by pretreated history health multi-parameter sample data, realize the real-time assessment to grain feelings safe class parameter, but key parameter optimizing is carried out to support vector machine and also can only adjust recurrence assessed value to certain limit, cannot fundamentally change the redundancy relationship between measuring point, the abnormal assessed value that still can have influence on other measuring points of one or more measuring point, so that reduce accuracy and the validity of equipment early warning system.Because abnormal measuring point to affect the assessment of its normal measuring point, so optimizing strategy inherently cannot eliminate the phenomenon that abnormal measuring point brings disturbing effect by the complex nonlinear mapping relations upsetting normal measuring point.
For above phenomenon, need the Function Mapping relation that new assessed value collimation technique reparation is disturbed by abnormal measuring point badly, effective compensation is carried out to the assessed for performance of device intelligence warning algorithm, realize effective matching of equipment normal condition data and highlighting of abnormality data, so just amplitude peak may improve the accuracy of device intelligence warning algorithm in online real-time assessment.
(3) summary of the invention
The present invention is in order to make up the deficiencies in the prior art, provide a kind of assessed value calibration steps of device intelligence early warning system, compensate for the present situation that in current various regression algorithm, accuracy is not high in equipment state assessment early warning, be intended to use Interpolation compensation method to eliminate abnormal measuring point departing from normal measuring point assessed value, make the assessed value of the normal measuring point of equipment be calibrated to the level meeting its virtual condition, thus achieve the performance optimization of regression algorithm; For the regression algorithms such as support vector machine give powerful guarantee in equipment state assessment early warning precision, from historical equipment data excavate all affect situation the non-linear effects relation of abnormal measuring point and normal measuring point, and be current real-time data correction assessed value devious with using interpolation method real-time online; Online assessment apparatus state aspect can be applied, and can effectively reduce because measuring point influences each other the false-alarm produced, simultaneously highly versatile, performance optimization can be carried out for other regression algorithms equally.
The present invention is achieved through the following technical solutions:
An assessed value calibration steps for device intelligence early warning system, is characterized in that: form by with lower part:
Part I: apparatus for establishing intelligent early-warning model, comprises the following steps:
1.1, import relevant device training data, training data be certain hour set up the normal history health data of multi-measuring point that is standby or equipment complex;
1.2, data prediction;
1.3, training data compression, adopts printenv compress mode to choose the quantitative data comprising all devices normal operating condition from training data;
1.4, build the Early-warning Model of multiple-input and multiple-output, n measuring point is successively as exporting the SVM regression model building n correspondence;
Part II: interact relation curve, comprises the following steps:
2.1, the number affecting type and abnormal measuring point is determined;
2.2, applicable interact relation curve is chosen;
2.3, the degree of fitting of calculated curve;
2.4, obtain parameter of curve, be saved in parameter database for subsequent use;
Part III: abnormal measuring point identification, comprises the following steps:
3.1, real time data is after recurrence assessment, uses healthy residual error bound Indexs measure, is screened by uninfluenced measuring point
3.2, real time data is after recurrence assessment, uses health data bound Indexs measure, is screened by the abnormal measuring point transfinited;
3.3, real time data is after recurrence assessment, uses wave characteristic Indexs measure, is screened by measuring point lower for intensity of anomaly;
3.4, real time data is after recurrence assessment, operating limit value degree of approximation Indexs measure, is therefrom elected by abnormal faint measuring point;
Part IV: the assessed value of calibrating disturbed measuring point, comprises the following steps:
4.1, each normal measuring point calls corresponding influence curve according to affecting type;
4.2, the mutation content of abnormal measuring point is calculated;
4.3, the departure degree of each normal measuring point assessed value is calculated according to influence curve;
4.4, method of interpolation is used to calibrate the assessed value of each normal measuring point successively.
Preferably, in step (1.2), erasing is taked to the Nan data in training data or after filling process means, is normalized operation.
Preferably, in step (2.2), the interact relation curve chosen is the S type curve of different gradient.
Preferably, in step (4.2), calculate abnormal measuring point and contrast with the situation before this measuring point no exceptions.
The invention has the beneficial effects as follows:
(1), abnormal measuring point departing from other normal measuring point assessed values can be eliminated in real time, assessed value through returning penalty method calibration more can the actual motion level of the normal measuring point of embodiment device, and the assessed value of abnormal measuring point can not be changed, greatly can improve algorithm of support vector machine and return the accuracy in assessment, there is good versatility simultaneously, performance optimization can be carried out in grafting to other regression algorithms;
(2), use the mode of multiple regression from history health data, excavate the abnormal measuring point of equipment and the non-linear effects relation normally between measuring point, because non-linear effects relation has the advantage of the influenced level trend extracting each measuring point in modeling process in advance, when the assessed value of the therefore influenced measuring point of on-line tuning equipment, too much delay effect can not be produced, ensure that model is to equipment real-time early warning ability;
(3), abnormal measuring point identification can be carried out in real time, although have multiple discrimination condition, mostly Rule of judgment is light and handy succinct and accurately applicable Logic judgment, does not relate to the computing of complicated and time consumption, the real-time judged is ensured, can carry out presence parsing for device data;
(4), method of interpolation is used to carry out the assessed value correction of influenced measuring point, and method of interpolation has various ways can select flexibly according to effect, and core Support vector regression model performance of the present invention is stablized, not only can fix but also can upgrade as required, coordinate method of interpolation can be that normal measuring point is stable and export glitch-free assessed value;
(5) data scale of the historical data, used is unrestricted, Early-warning Model is built by the redundancy relationship excavated between each measuring point in the device history data obtained, run Early-warning Model based on data-driven simultaneously, characteristic completely based on data makes the present invention not be subject to the restriction of expertise knowledge and physical arrangement, therefore Early-warning Model reliability of the present invention is higher, and life cycle is long.
(4) accompanying drawing explanation
Below in conjunction with accompanying drawing, the present invention is further illustrated.
Accompanying drawing 1 is process flow diagram of the present invention;
Accompanying drawing 2 is the condition judgment process flow diagram of abnormal measuring point identification division of the present invention;
Accompanying drawing 3 is the operational flow diagram of utilization Interpolation compensation method evaluation of alignment value of the present invention;
Accompanying drawing 4 is normal measuring point of the present invention and abnormal measuring point influence curve schematic diagram;
Accompanying drawing 5 is different S type function of the present invention and influence curve matching schematic diagram;
Accompanying drawing 6 is the equipment early warning situation comparison diagram in the present embodiment two before and after abnormal measuring point application the present invention;
Accompanying drawing 7 is the equipment early warning situation comparison diagram in the present embodiment two before and after normal measuring point application the present invention.
(5) embodiment
Accompanying drawing is a kind of specific embodiment of the present invention.
Embodiment 1
Based on an assessed value calibration steps for the device intelligence early warning system of SVM, as shown in Figure 1, the whole design of the present invention comprises following process:
Process 1 is equipment apparatus for establishing intelligent early-warning model, and this process mainly comprises four key steps.
Step 1.1 training data imports
Existing selected training data is one set up standby multi-measuring point history health data, and roughly operating process is: be as the criterion with selected measuring point from power plant's database, read the history run status data of this relevant device enough time; Then the screening rule of setting is utilized from all historical datas, to filter out the good healthy device data of running status as the training data building Support vector regression model.
The screening rule of training data is in order to whole historical data rejecting abnormalities data are obtained healthy service data, if there is the phenomenons such as numerical value transfinites, fluctuation violent, ripple disable in trend map in data trend, then this segment data can be considered abnormal data and should reject, and after screening, namely remaining data can be used as the training data of modeling.
If there is n the device databases observing measuring point from one to intercept the device data in m moment as training data, then can regard the vector of a n dimension at each observation measuring point data in j moment as, can be expressed as:
Then this training data file should be the matrix form of m × n.Concrete form is as follows:
。
Step 1.2 data prediction
Mainly filling process is carried out to Nan data, after process, carry out standardization again.
Inevitably there are some invalid empty data and Nan(Notanumber in the original training data obtained due to the reason such as network or software) data, if do not processed, the moment containing Nan data cannot Training Support Vector Machines regression model.For this kind of situation, the mode that the present invention adopts arest neighbors data interpolating to fill is to ask for the position that new data makes up Nan data.The formula adopted is:
（1）
Wherein, Value
boundary1, Value
boundary2for front and back data boundary, Time
boundary1, Time
boundary12for front and back boundary time, Value
nan, Time
nanfor Nan data and time thereof.
In order to eliminate the inaccurate impact of recurrence assessment result that different dimension between each measuring point causes, also need the operation carrying out standardization after training data process Nan data, each measuring point data of training data through standardization all meets average to be 0 variance be 1 normal distribution.The process of standardization will be carried out according to following formula:
（2）
Wherein, dataValue is training data, and meanValue is training data average, and stdValue is training data standard deviation.
Step 1.3 training data compresses
In order to save the model construction of SVM time and prevent model training from excessively occurring over-fitting, the present invention adopts a kind of parameterless training data compress mode to reach above object.The data mode that original training data comprises is a lot, but periodic data or adjacent data mode is very similar or even repeat, all data close with its state can be replaced by wherein data completely.
The present invention carries out data compression in the following manner, and first artificial subjectivity determines the state number d that packed data comprises, and then each extreme value state of observing under measuring point is extracted, comprises d altogether
1individual; If extreme value status number d
1still be less than the d of requirement, just the training data in m moment asked Euclid norm, obtain the vector of a m dimension, then the vector ascending order arrangement again that m is tieed up, according to certain step-length L equidistant state d extracting a given data from m dimensional vector
2; Merge d
1, d
2just packed data R can be obtained.The determination of step-length L is divided by by m and d and is rounded acquisition, and packed data R will comprise healthy running statuses all in training data, and it can replace training data to be used for building Support vector regression model, to reach the object of saving time with preventing over-fitting.
Step 1.4 builds the SVM regression model of multiple-input and multiple-output
SVM Regression is in the feature space of higher-dimension by multiple measuring point Parameter Mapping, then construct strongly-convex problem with target component for exporting, by the feature space Margin Vector (namely support vector machine) that catches constantly Tuning function system go to approach one can solve the problem while but also comprise the complete function expression of input parameter and output parameter complex nonlinear mapping relations.Therefore common SVM regression model may be only the pattern of multiple input single output.For realizing the output of multi-measuring point parameter, the present invention all builds a series of multiple-input and multiple-output SVM regression model as the mode exporting target component as these measuring point parameters while input parameter successively by using all measuring point parameters.
Be the packed data that upper step obtains the matrix form of m × n because the present invention uses modeling data, final SVM regression model can be many SVM regression model that n group is n-1 form.Modeling principle explanation is carried out below to build a single SVM regression model.
For the Nonlinear Mapping relation between input parameter and output parameter is showed, according to general SVM modeling principle, by introducing kernel function
nonlinear problem is converted into quadratic programming problem, and its objective function is
, for improving the Generalization Capability of SVM, special introducing slack variable
, error penalty factor and discriminant function
, thus ask for maximum spacing lineoid optimization problem and be converted into quadratic programming problem, as follows:
（3）
For solving above-mentioned optimization problem, need introduce Lagrange factor transition problem is dual problem, as follows:
（4）
Separate above-mentioned dual problem, obtain optimized Lagrange factor
, wherein
, l is the number of support vector machine, by Lagrange factor
, support vector machine l, slack variable
, the important parameter such as error penalty factor is saved in SVM model, thus solved by regression problem.
The structure of n SVM regression model carries out after structure completes, just to go out one relative to the assessed value vector in historical data to all measuring point parameter fittings of every bar test data vector all in the manner described above.
Process 2 is interact relation curve fitting process, and this process also belongs to model and sets up part, mainly comprises following four key steps.
Step 2.1, determines to affect type
The present invention needs the mode considering all abnormal other normal measuring points of measuring point combined effect, the determinative affecting type have anomaly source number x(and abnormal measuring point number) and the departure degree e size of normal measuring point.Then one is had to the data of n measuring point, then should have
(5) the individual excavation affecting type and need to carry out interact relation.
Suppose that every bar training data is a n-dimensional vector, then anomaly source number magnitude range exists
within, and the size of the departure degree e of normal measuring point should be
within, the limit departure degree f of certain measuring point refers to the increase gradually along with the mutation content of abnormal measuring point in regression algorithm as shown in Figure 4, the ultimate value that the assessed value of normal measuring point parameter can be approached gradually along specific track, its f value size is different because the numerical value dimension of this normal measuring point in training data is different.Below with measuring point
for example, asking for of its f value is illustrated:
First the extreme value of the original training data S data of measuring point i is extracted
with
;
And then the spacing of two extreme values is divided into point spacing such as 100, and add up in training data s the frequency y dropped in each decile spacing successively
iand average
,
Finally just can try to achieve ultimate value f according to following formula:
（6）
Step 2.2, chooses applicable match curve, and interact relation curve is the S type curve of multiparameter;
This step mainly finds the function be applicable to for expressing accompanying drawing 4 curve, and asking for this class function to what replace different abnormal measuring point combination on each normal measuring point affects nonlinear relationship.
In accompanying drawing 4, class of a curve is similar to animal population growth, Growth of Cells breeding isotype, consults related data document and can be used for expressing having of similar curved line relation
function,
function and
function, their mathematic(al) representation is as follows:
1、
（7）
Wherein important related coefficient is: A
1for initial value, A
2for ultimate value, x
0for curvilinear abscissa center, p is the gradient;
2、
（8）
Wherein important related coefficient is: y
0for offset, A
1for amplitude 1, A
2for amplitude 2, t
1for bandwidth 1, t
2for bandwidth 2;
3、
（9）
Wherein important related coefficient is: A
1for initial value, A
2for ultimate value, x
0for curvilinear abscissa center, dx is the gradient;
Each affects type all can carry out relation curve matching according to above three kinds of functions, and the related coefficient of setting fitting index adjustment often kind of function accurately can affect the mode of normal measuring point assessed value by abnormal expression measuring point.
Step 2.3, the degree of fitting of calculated curve, namely weighs the extent of deviation of curve prediction value and instantaneous value;
This step can represent the function expression of measuring point interact relation curve under particular condition for choosing, and solves the functional dependence coefficient with best-fit degree.
First, artificially subjectively specify the mode of the target measuring point that will study and affect anomaly source scale and the degree of its assessed value.For 63 numerical value of original training data measuring point i, measuring point j as artificial anomaly source, by abnormal measuring point
intensity of anomaly doubly progressively rise to 10 times with the step-length of 0.01 from original normal level 2, and the assessed value departure degree recording measuring point i is in following table 1, with intensity of anomaly e for transverse axis, normal measuring point assessed value departure degree f is the trend map that the longitudinal axis makes as accompanying drawing 4.
Then, to adjusting for the related coefficient in reference function with three kinds of S type curvilinear functions in step 2.3, and with curve degree index R
2carry out evaluating the performance representing measuring point interact relation curve, R
2the Function Fitting performance that larger expression is selected is better, as shown in Figure 5
the R of function
2value is greater than the R of other two functions
2value, should select
function, often kind of situation selects best reference function and related coefficient.Be evaluation function curve degree Index Formula below:
（10）
Wherein, y
ifor the normal departure degree in upper table,
for the average of the normal departure degree in upper table,
for the predicted value of selected function, n is that the intensity of anomaly of abnormal measuring point changes number of times.
Step 2.4, is saved in parameter database for subsequent use by parameter of curve;
Need step to ask in this step all affect the matched curve of type reference function and related coefficient be stored in the parameter database of specifying, as a part for regression model.
The present invention carries out three node layer storage rules, ensures the ageing of evaluation of alignment value.
1, type l is affected with abnormal measuring point
3as the memory node of third layer, l
3type is affected for what determine in step 2.1
（11）
2, with affected normal measuring point l
1for the memory node of ground floor because carry out assessed value calibration time be by influenced measuring point sequence number successively method of interpolation compensate;
3, with the numerical value l of affected normal measuring point
2for the memory node of the second layer, l
2determine in the following manner:
（12）
Wherein, S
maxfor the maximal value of this normal measuring point of training data, S
minfor the minimum value of this normal measuring point of training data, n is the degree of segmentation, and being commonly defined as 10, i is sequence number;
The content stored is the related coefficient adjusted in reference function in step 2.2 and step 2.3.
Process 3 is abnormal measuring point identifying, and this process belongs to model running category, mainly comprises four key steps.
This process is first explained in detail on the basis introducing model running.
The measure-point amount being input to every bar real time data of n SVM regression model must be identical with training data R, and need to carry out carrying out standardization under the standard of training data, to eliminate dimension impact.Assess in n the SVM regression model that the input that walked abreast by a n dimension real time data V is corresponding, the assessment vector of corresponding n dimension will be obtained
.By accompanying drawing 2 can find out this process will come with corresponding four steps right
numerical accuracy carry out the evaluation of four indexs, qualified words
value no longer corrects, underproof words
value needs to calibrate in process four.
Step 3.1, real time data, after recurrence assessment, uses healthy residual error bound Indexs measure, uninfluenced measuring point and doubtful abnormal measuring point can be screened respectively;
This step first illustratively, residual error is defined as the difference of instantaneous value and assessed value, and the acquisition of healthy residual error bound standard completes after SVM model training completes, inputted successively by original training data in SVM model and all carry out recurrence and assess and ask for residual values, then the most value of the residual error of each measuring point can be used as healthy residual error bound T.
Current real time data assess after residual error
if,
, then this measuring point belongs to uninfluenced measuring point, if
, then this measuring point belongs to doubtful abnormal measuring point, needs to proceed to lower step and detects more specifically.
Step 3.2, real time data, after recurrence assessment, uses health data bound Indexs measure, the abnormal measuring point transfinited can be screened;
Below first this step illustrates, this step only detects the real time data of the doubtful abnormal measuring point of step.The acquisition of health data bound standard completes after SVM model training completes, and the most value of each for original training data measuring point be can be used as health data bound D.
Use health data bound D to detect to the doubtful abnormal measuring point numerical value of real time data, the mode of detection is:
If
, then this measuring point is still doubtful abnormal measuring point, needs the detection carrying out lower step;
If
, then this measuring point is defined as abnormal measuring point.
Step 3.3, real time data, after recurrence assessment, uses wave characteristic Indexs measure, measuring point lower for intensity of anomaly can be screened;
First illustratively, this step only detects the real time data of the doubtful abnormal measuring point of step and corresponding assessed value to this step.Wave characteristic uses coefficient of variation feature CV to represent, its concrete expression formula is as follows:
（13）
Here, stdValue represents that size is the variance in the time window of 5, and stdValue represents that size is the average in the time window of 5, and the detection mode of this step is as follows:
First, intercepting respectively with this moment size that is terminal is real time data group and the assessment data group of the time window of 5;
Then, the coefficient of variation feature CV of difference real time data group and assessment data group;
Finally, successively upper step is compared, if
, this measuring point is defined as abnormal measuring point, if
, this measuring point still needs for doubtful abnormal measuring point the detection carrying out next step.
Step 3.4, real time data is after recurrence assessment, and operating limit value degree of approximation Indexs measure, therefrom can elect the measuring point at abnormal initial stage;
First illustratively, this step only detects the real time data of the doubtful abnormal measuring point of step and corresponding assessed value to this step.Measuring point ultimate value f has introduction in step 2.1, and its acquisition has also been after SVM model training completes, and introduces below and how to detect:
First, intercepting respectively with this moment size that is terminal is real time data group and the assessment data group of the time window of 5;
Then, each doubtful instantaneous value of abnormal measuring point and the difference of ultimate value is calculated respectively successively
, assessed value and ultimate value difference
;
Finally, calculate
with
average, if
, this measuring point is judged to be abnormal measuring point; If
, this measuring point is finally judged to be normal measuring point.
After above four steps complete, all measuring points of real time data can be divided into normal measuring point and abnormal measuring point two class.
Process 4 is adjust the assessed value process be disturbed, and this process still belongs to model running category, will according to accompanying drawing 3 with moment t
ifor example introduces four key steps of whole process.
Step 4.1, each normal measuring point can call corresponding influence curve according to affecting type;
This step is current time t
iaccording to discrimination n out in process 3 abnormal measuring point group
, be the type of the supplemental characteristic library call interact relation curve that each normal measuring point is kept from process 2, if
, the type of the interact relation curve that the ratio of slope should be selected larger.Mutually in the same time under owing to having common abnormal measuring point group p, so each normal measuring point now affects type belong to same classification; Not in the same time because abnormal measuring point group p can change, so normal measuring point does not in the same time affect type and belongs to a different category.
The particular procedures of this step is:
First, current time is determined
abnormal measuring point group
;
Then, according to abnormal measuring point group p, each normal measuring point is from parameter database determination interact relation curve type c
i.
Step 4.2, calculates the mutation content of each abnormal measuring point, and the situation mainly and before this measuring point no exceptions contrasts;
According to abnormal measuring point group p, determine each abnormal measuring point p successively
imutation content L, below introduce and how to calculate mutation content L:
First, with abnormal measuring point P
ifor example, from current time t
istart, inquire about abnormal measuring point p backward
ihistorical data in, distance current time t
ithe nearest normal moment t that do not report to the police
just;
Then, be abnormal measuring point p
icalculate mutation content L
i, p
justfor the normal moment t that do not report to the police
justreal time value, then:
, successively by the measuring point mutation content L of all abnormal measuring point groups.
Step 4.3, calculates the assessed value departure degree scope of each normal measuring point according to influence curve;
This step is used for locking an assessed value departure degree scope for each normal measuring point.Owing to the interact relation curve of the real time value in all normal ranges of each measuring point all cannot be determined in process 2, can only in four corner interval
ask for interact relation curve.What this step was done is determine current time t
inormal measuring point real time value belong to
in between any two spaced points, the interact relation curve lock accepted opinion valuation irrelevance scope that recycling is corresponding.
Now with affected normal measuring point p
ifor example, introduce how to determine assessed value irrelevance scope.
First, according to the p of normal measuring point
ireal time value
, determine its
in left and right close on spaced points
;
Then, according to interact relation curve type c
iand abnormal measuring point group p, determine
two corresponding interact relation curves
,
;
Finally, according to the measuring point mutation content L of abnormal measuring point group, bring into
,
in can obtain affected normal measuring point p
iassessed value irrelevance scope y
i, y
j.
Step 4.4, uses the assessed value of method of interpolation to normal measuring point to adjust according to above-mentioned steps;
This step is mainly introduced and how to be adopted method of interpolation that affected normal measuring point assessed value is adjusted to corresponding normal level.
Be below the assessed value departure degree y of each normal measuring point of method of interpolation accurate Calculation
x, computing formula is:
(14)
Normal measuring point p after adjustment
iassessed value
adjustment formula is:
(15).
Current time t
ia new assessed value is combined into not needing the assessed value of the abnormal measuring point group adjusted after the assessed value of all normal measuring point groups has all adjusted
, and
more can embody current time t
ithe status level of all measuring points.
Embodiment 2
In order to further illustrate implementation process of the present invention, the present invention chooses the important measuring point data from the A primary air fan equipment in No. 1 unit boiler subsidiary engine of certain thermal power plant, to verify the useful subsidy of the present invention to equipment state early warning.
The assessed value calibration steps key step based on support vector machine of the present invention to A primary air fan is as follows:
One, the history health data of A primary air fan is utilized to build the intelligent early-warning model process of multiple-input and multiple-output
First, selected to participate in the important parameter measuring point building SVM regression model, and from power plant's database extraction time sufficiently long data, obtain training data according to screening rule.
The measuring point that this example chooses A primary air fan has real power (MW), total primary air flow (t/h), A primary air fan top hole pressure (Kpa), A primary air fan electric current (A), the A primary air fan bearing X of sending out of unit to 10 and the closely-related parameter measuring points of this equipment running status such as vibrations (nm/s).
The time of the original historical data that this example is chosen is the time period being about half a year from June, 2013 in Dec, 2013, has 220000 data altogether.The data of first 4 months of data are used for building equipment Early-warning Model by we, and the data of then 2 months are for detecting the early warning effect of Early-warning Model.
For the modeling data of 4 months, we needed to utilize existing screening rule to reject abnormal data in each measuring point data successively, because this partial data does not represent the normal operating condition of A primary air fan equipment.Remaining data just can be used as the training data T building SVM regression model.
The data prediction of training data T.Need before extraction to carry out data prediction to training data T, utilize the position of formula (1) to the appearance Nan data of training data T to complete the process of sky data stuffing; Formula (2) is utilized to carry out standardization to eliminate dimension impact to training data T;
Training data T extracts packed data R.According to the mode that step 1.2 is told about, determine the data volume 2000 of packed data R artificially, first choose the extreme value status data of each measuring point, the rear normal state data utilizing norm rule to extract some with a fixed step size, extreme value status data and normal state data merge can obtain the packed data R that quantity is about 2000, and packed data R just contains the nearly all data mode of training data.
Then, packed data R is used to train 10 SVM regression models according to SVM Regression.A SVM is the form of multiple input single output, and use 10 measuring points of A primary air fan all as input here, 10 measuring points realize the function of multi output successively as the mode exported.The training process of model realizes Lagrange factor by formula (3), formula (4)
, support vector machine l, slack variable
, the important parameter information such as error penalty factor is stored in SVM regression model.
Two, for A primary air fan device data sets up interact relation parameter of curve database M
In order to save the model running time, this process completes by the present invention in model construction process.What interact relation curve embodied is the relation of the mutation content of abnormal measuring point and the assessed value irrelevance of normal measuring point.Because the A primary air fan equipment of research has 10 crucial measuring points, so should have by formula (5)
plant interact relation curve and need calculating parameter.
Be affected normal measuring point with measuring point 1, measuring point 2 is the situation of abnormal measuring point is example, introduces the correlation parameter how calculating interact relation curve.
First determine that the most value of the training data R of measuring point 1 is after 50 and 150, it is 110.7 that through type (6) weighted mean obtains assessment ultimate value f;
Then measuring point 1 is with minimum value 50 for initial value, and measuring point 2 mutation content from any regime values, progressively can suddenly change with the step-length of initial value 0.01 times, and records corresponding distortion assessed value, sets up the form as table 1 form.
Call three kinds of different S type functions successively
function,
function and
function carries out matching to the relation that above table contains, and uses fitting index R
2investigate the degree of fitting of curve, select R
2one group of best parameter value is saved in parameter database M.
Numerical value be 50 measuring point 1 obtain optimal function parameter after, take step-length as the increase degree interval value of 10, often obtaining a parameter all needs to preserve, to ensure the integrality of parameter database M.
All
the parameter of planting interact relation curve calculates all in the manner described above, and is saved in parameter database M.
Three, equipment Early-warning Model is run and abnormal measuring point identifying
Before operation Early-warning Model, for ensureing to implement carrying out smoothly of abnormal measuring point discrimination, we need to fulfil following work ahead of schedule:
First, from the history health data of 10 measuring points of A primary air fan, extract health data bound D, i.e. the most value of each measuring point history health data.
Then, we also need the healthy residual error bound T of all measuring points asking for A primary air fan, history health data is all put into successively in Early-warning Model and is carried out recurrence assessment, instantaneous value and assessed value are made difference and are obtained residual error data, being worth most namely as the healthy residual error bound T of this measuring point of each measuring point residual error data.
Make test data can be used as the real time data of equipment, all measuring point datas of every bar real time data as input, and are advanced in n SVM regression model the assessed value that can obtain n correspondence.
We are with the real time data V of A primary air fan
ifor example, carry out the identifying of specification exception measuring point.
Real time data V
iin SVM regression model, assessment obtains assessment data
after, successively use four indexs to carry out the discrimination of abnormal measuring point, first index is healthy residual error bound T, and the measuring point data be in outside residual error scope will be judged to make doubtful abnormal measuring point data, continue the judgement of next index.Second index is health data bound D, and the measuring point be in outside data area will be judged to make abnormal measuring point, and doubtful abnormal measuring point data is made in sentencing of other, continues the judgement of next index.3rd index is that undulatory property detects, and after asking for CV feature, compares CV according to formula (13)
realwith CV
commentsize, if CV
real>CV
comment, then sentence and make abnormal measuring point data, otherwise sentence and do the judgement that doubtful abnormal measuring point data continues next index.4th index is ultimate value degree of approximation index
, according to the decision rule of step 3.4, what satisfy condition is abnormal measuring point data.Through above screening layer by layer, real time data V
iall measuring points by point work two class: abnormal measuring point (measuring point 3,5,7) and normal measuring point (measuring point 1,2,4,6,8,9,10).
Four, the assessed value process that A primary air fan is disturbed normal measuring point is calibrated
This process is carried out on the basis of process three, still with moment t
ia primary air fan testing of equipment data V
ifor example, process three is by V
tmeasuring point be divided into abnormal measuring point and normal measuring point, our calibration here be the assessed value data of normal measuring point
.
What determine normal measuring point at process three discrimination abnormal measuring point out affects type, to calibrate the assessed value of measuring point 1
Call from primary air fan parameter database M normal measuring point be 1 and abnormal measuring point be 3,5,7 interact relation parameter of curve.
The respective mutation content L that can obtain measuring point 3,5,7 according to the computation rule of step 4.2 is
and the scope of measuring point 1 instantaneous value 8.53 is
between.According to the computation rule of step 4.3, according to interact relation curve can be regarded as and initial value
the assessed value irrelevance of correspondence be
.Finally calling the assessed value irrelevance of method of interpolation algorithm when instantaneous value is 8.53 is 2.67, and the assessed value of reality is 11.1, and the assessed value after calibration is 8.43.
Moment t
ia primary air fan testing of equipment data V
ithe new assessed value that all normal measuring point datas obtain after calibrating all in the manner described above
, the assessed value of other moment real time datas obtains corresponding evaluation of alignment value according to same way.As shown in Figure 5, what Fig. 6 showed is contrast situation before and after the calibration of abnormal measuring point 3, and Fig. 7 displaying is contrast situation before and after the calibration of affected normal measuring point 1, can find out between the moment 1300 ~ 1900, what exception occurred is measuring point 3, before calibration, the assessed value of affected normal measuring point 1 departs from, and after calibration, the assessed value of measuring point 1 realizes intact calibration, and numerical values recited more can the operation conditions level of this measuring point of embodiment device.
Claims (7)
1. an assessed value calibration steps for device intelligence early warning system, is characterized in that: form by with lower part:
Part I: apparatus for establishing intelligent early-warning model, comprises the following steps:
1.1, import relevant device training data, training data be certain hour set up the normal history health data of multi-measuring point that is standby or equipment complex;
1.2, data prediction;
1.3, training data compression, adopts printenv compress mode to choose the quantitative data comprising all devices normal operating condition from training data;
1.4, build the Early-warning Model of multiple-input and multiple-output, n measuring point is successively as exporting the SVM regression model building n correspondence;
Part II: interact relation curve, comprises the following steps:
2.1, the number affecting type and abnormal measuring point is determined;
2.2, applicable interact relation curve is chosen;
2.3, the degree of fitting of calculated curve;
2.4, obtain parameter of curve, be saved in parameter database for subsequent use;
Part III: abnormal measuring point identification, comprises the following steps:
3.1, real time data is after recurrence assessment, uses healthy residual error bound Indexs measure, is screened by uninfluenced measuring point
3.2, real time data is after recurrence assessment, uses health data bound Indexs measure, is screened by the abnormal measuring point transfinited;
3.3, real time data is after recurrence assessment, uses wave characteristic Indexs measure, is screened by measuring point lower for intensity of anomaly;
3.4, real time data is after recurrence assessment, operating limit value degree of approximation Indexs measure, is therefrom elected by abnormal faint measuring point;
Part IV: the assessed value of calibrating disturbed measuring point, comprises the following steps:
4.1, each normal measuring point calls corresponding influence curve according to affecting type;
4.2, the mutation content of abnormal measuring point is calculated;
4.3, the departure degree of each normal measuring point assessed value is calculated according to influence curve;
4.4, method of interpolation is used to calibrate the assessed value of each normal measuring point successively.
2. the assessed value calibration steps of a kind of device intelligence early warning system according to claim 1, is characterized in that: in step (1.2), takes erasing or after filling process means, be normalized operation to the Nan data in training data.
3. the assessed value calibration steps of a kind of device intelligence early warning system according to claim 1, is characterized in that: in step (2.2), the interact relation curve chosen is the S type curve of different gradient.
4. the assessed value calibration steps of a kind of device intelligence early warning system according to claim 1, is characterized in that: in step (4.2), calculates abnormal measuring point and contrasts with the situation before this measuring point no exceptions.
5. the assessed value calibration steps of a kind of device intelligence early warning system according to claim 1, it is characterized in that: in step (1.3), determine the state number d that packed data comprises, then each extreme value state of observing under measuring point is extracted, comprise d altogether
1individual; If extreme value status number d
1still be less than the d of requirement, just the training data in m moment asked Euclid norm, obtain the vector of a m dimension, then the vector ascending order arrangement again that m is tieed up, according to certain step-length L equidistant state d extracting a given data from m dimensional vector
2; Merge d
1, d
2just packed data R can be obtained.
6. the assessed value calibration steps of a kind of device intelligence early warning system according to claim 5, it is characterized in that: the determination of step-length L is divided by by m and d and is rounded acquisition, packed data R comprises healthy running statuses all in training data.
7. the assessed value calibration steps of a kind of device intelligence early warning system according to claim 1, is characterized in that: in step (4.3), first, according to the real time value of normal measuring point, close on spaced points about determining it; Then, according to interact relation curve type and abnormal measuring point group, determine two corresponding interact relation curves; Finally, according to the measuring point mutation content of abnormal measuring point group, bring the assessed value irrelevance scope that can obtain affected normal measuring point in two interact relation curves into.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510105156.0A CN105302848B (en) | 2014-10-11 | 2015-03-10 | A kind of assessed value calibration method of device intelligence early warning system |
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2014105328987 | 2014-10-11 | ||
| CN201410532898 | 2014-10-11 | ||
| CN201510105156.0A CN105302848B (en) | 2014-10-11 | 2015-03-10 | A kind of assessed value calibration method of device intelligence early warning system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105302848A true CN105302848A (en) | 2016-02-03 |
| CN105302848B CN105302848B (en) | 2018-11-13 |
Family
ID=55200121
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510105156.0A Active CN105302848B (en) | 2014-10-11 | 2015-03-10 | A kind of assessed value calibration method of device intelligence early warning system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105302848B (en) |
Cited By (27)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106127359A (en) * | 2016-08-30 | 2016-11-16 | 北京协同创新智能电网技术有限公司 | A kind of variable frequency pump rate of discharge method for early warning based on local weighted linear regression model (LRM) |
| CN106682781A (en) * | 2016-12-30 | 2017-05-17 | 山东鲁能软件技术有限公司 | Power equipment multi-index prediction method |
| CN106960554A (en) * | 2017-03-20 | 2017-07-18 | 山东科技大学 | Warning system evaluation method and device based on single simulation amount and presignal delay device |
| CN107067143A (en) * | 2016-12-30 | 2017-08-18 | 山东鲁能软件技术有限公司 | A kind of equipment safety grade separation method |
| CN107678930A (en) * | 2017-09-11 | 2018-02-09 | 华东理工大学 | A kind of bank's automatic terminal abnormal alarm method based on Smooth Support Vector Machines |
| CN107807961A (en) * | 2017-10-10 | 2018-03-16 | 国网浙江省电力公司丽水供电公司 | A kind of power distribution network big data multidomain treat-ment method based on Spark computing engines |
| CN108089962A (en) * | 2017-11-13 | 2018-05-29 | 北京奇艺世纪科技有限公司 | A kind of method for detecting abnormality, device and electronic equipment |
| CN108133326A (en) * | 2017-12-22 | 2018-06-08 | 华润电力（菏泽）有限公司 | A kind of status early warning method and system based on thermal power generating equipment |
| CN109492193A (en) * | 2018-12-28 | 2019-03-19 | 同济大学 | Abnormal network data based on depth machine learning model generate and prediction technique |
| CN109669849A (en) * | 2018-12-04 | 2019-04-23 | 哈尔滨工业大学 | It is a kind of based on the complication system health state evaluation method for not knowing depth theory |
| CN110321366A (en) * | 2019-06-27 | 2019-10-11 | 中国科学院自动化研究所 | Method and system are determined based on the statistic of on-line study |
| CN110391840A (en) * | 2019-09-17 | 2019-10-29 | 中国人民解放军国防科技大学 | Method and system for judging abnormality of telemetry parameters of sun synchronous orbit satellite |
| CN110469496A (en) * | 2019-08-27 | 2019-11-19 | 苏州热工研究院有限公司 | A kind of water pump intelligent early-warning method and system |
| CN111241208A (en) * | 2019-12-31 | 2020-06-05 | 安徽中科大国祯信息科技有限责任公司 | Method and device for monitoring abnormity of periodic time sequence data |
| CN111797250A (en) * | 2020-09-09 | 2020-10-20 | 腾讯科技（深圳）有限公司 | Media content push control method and device, computer equipment and storage medium |
| WO2021008142A1 (en) * | 2019-07-15 | 2021-01-21 | Huawei Technologies Co., Ltd. | Method and system for automatic anomaly detection in data |
| CN112633329A (en) * | 2020-12-05 | 2021-04-09 | 歌尔科技有限公司 | Method for detecting mechanical property defects of intelligent equipment |
| CN112650660A (en) * | 2020-12-28 | 2021-04-13 | 北京中大科慧科技发展有限公司 | Early warning method and device for power system of data center |
| CN112734977A (en) * | 2020-12-25 | 2021-04-30 | 安徽省安泰科技股份有限公司 | Equipment risk early warning system and algorithm based on Internet of things |
| CN113066549A (en) * | 2021-04-06 | 2021-07-02 | 青岛瑞斯凯尔生物科技有限公司 | Clinical effectiveness evaluation method and system of medical instrument based on artificial intelligence |
| CN113168171A (en) * | 2018-12-05 | 2021-07-23 | 三菱电机株式会社 | Abnormality detection device and abnormality detection method |
| CN113806615A (en) * | 2021-11-18 | 2021-12-17 | 成都索贝数码科技股份有限公司 | KPI (Key performance indicator) abnormity early warning method of intelligent IT operation and maintenance system |
| CN113822580A (en) * | 2021-09-24 | 2021-12-21 | 华润电力技术研究院有限公司 | Equipment working condition evaluation method and related equipment |
| CN114017136A (en) * | 2021-11-02 | 2022-02-08 | 浙江大唐国际绍兴江滨热电有限责任公司 | Soft light character plate alarm method and system for combined cycle generator set |
| CN114356998A (en) * | 2022-01-08 | 2022-04-15 | 浙江力石科技股份有限公司 | Big data tourism resource data compensation method |
| CN114742391A (en) * | 2022-04-02 | 2022-07-12 | 格莱杰(深圳)科技有限公司 | Artificial intelligence based exception handling method and device, computer equipment and medium |
| CN116302848A (en) * | 2023-05-19 | 2023-06-23 | 杭州安脉盛智能技术有限公司 | Detection method and device for bias of evaluation value, electronic equipment and medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CA2448663A1 (en) * | 2001-05-31 | 2002-12-05 | Goldman Sachs & Co. | Method and system for verifying the integrity of data in a data warehouse and applying warehoused data to a plurality of predefined analysis models |
| US20120123981A1 (en) * | 2010-08-04 | 2012-05-17 | Graves Spencer B | Software to facilitate design, data flow management, data analysis and decision support in structural health monitoring systems |
| CN102982383A (en) * | 2012-05-15 | 2013-03-20 | 红云红河烟草(集团)有限责任公司 | Energy supply and demand prediction method based on support vector machine |
| CN103345585A (en) * | 2013-07-12 | 2013-10-09 | 清华大学 | Wind power prediction correction method and system based on support vector machine |
| CN103473480A (en) * | 2013-10-08 | 2013-12-25 | 武汉大学 | Online monitoring data correction method based on improved universal gravitation support vector machine |
| CN104102773A (en) * | 2014-07-05 | 2014-10-15 | 山东鲁能软件技术有限公司 | Equipment fault warning and state monitoring method |
-
2015
- 2015-03-10 CN CN201510105156.0A patent/CN105302848B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CA2448663A1 (en) * | 2001-05-31 | 2002-12-05 | Goldman Sachs & Co. | Method and system for verifying the integrity of data in a data warehouse and applying warehoused data to a plurality of predefined analysis models |
| US20120123981A1 (en) * | 2010-08-04 | 2012-05-17 | Graves Spencer B | Software to facilitate design, data flow management, data analysis and decision support in structural health monitoring systems |
| CN102982383A (en) * | 2012-05-15 | 2013-03-20 | 红云红河烟草(集团)有限责任公司 | Energy supply and demand prediction method based on support vector machine |
| CN103345585A (en) * | 2013-07-12 | 2013-10-09 | 清华大学 | Wind power prediction correction method and system based on support vector machine |
| CN103473480A (en) * | 2013-10-08 | 2013-12-25 | 武汉大学 | Online monitoring data correction method based on improved universal gravitation support vector machine |
| CN104102773A (en) * | 2014-07-05 | 2014-10-15 | 山东鲁能软件技术有限公司 | Equipment fault warning and state monitoring method |
Non-Patent Citations (3)
| Title |
|---|
| 栗然等: "基于支持向量机回归的光伏发电出力预测", 《中国电力》 * |
| 王定成: "支持向量回归在线建模及应用", 《控制与决策》 * |
| 赵石磊等: "SVM回归估计方法在卫星故障诊断中的应用", 《电机与控制学报》 * |
Cited By (41)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106127359A (en) * | 2016-08-30 | 2016-11-16 | 北京协同创新智能电网技术有限公司 | A kind of variable frequency pump rate of discharge method for early warning based on local weighted linear regression model (LRM) |
| CN106682781A (en) * | 2016-12-30 | 2017-05-17 | 山东鲁能软件技术有限公司 | Power equipment multi-index prediction method |
| CN107067143A (en) * | 2016-12-30 | 2017-08-18 | 山东鲁能软件技术有限公司 | A kind of equipment safety grade separation method |
| CN106960554A (en) * | 2017-03-20 | 2017-07-18 | 山东科技大学 | Warning system evaluation method and device based on single simulation amount and presignal delay device |
| CN106960554B (en) * | 2017-03-20 | 2018-06-29 | 山东科技大学 | Alarm system evaluation method and device based on single simulation amount and presignal delay device |
| CN107678930A (en) * | 2017-09-11 | 2018-02-09 | 华东理工大学 | A kind of bank's automatic terminal abnormal alarm method based on Smooth Support Vector Machines |
| CN107807961B (en) * | 2017-10-10 | 2021-11-19 | 国网浙江省电力公司丽水供电公司 | Spark calculation engine-based power distribution network big data partition processing method |
| CN107807961A (en) * | 2017-10-10 | 2018-03-16 | 国网浙江省电力公司丽水供电公司 | A kind of power distribution network big data multidomain treat-ment method based on Spark computing engines |
| CN108089962A (en) * | 2017-11-13 | 2018-05-29 | 北京奇艺世纪科技有限公司 | A kind of method for detecting abnormality, device and electronic equipment |
| CN108133326A (en) * | 2017-12-22 | 2018-06-08 | 华润电力（菏泽）有限公司 | A kind of status early warning method and system based on thermal power generating equipment |
| CN109669849A (en) * | 2018-12-04 | 2019-04-23 | 哈尔滨工业大学 | It is a kind of based on the complication system health state evaluation method for not knowing depth theory |
| CN109669849B (en) * | 2018-12-04 | 2022-01-28 | 哈尔滨工业大学 | Complex system health state assessment method based on uncertain depth theory |
| CN113168171B (en) * | 2018-12-05 | 2023-09-19 | 三菱电机株式会社 | Abnormality detection device and abnormality detection method |
| CN113168171A (en) * | 2018-12-05 | 2021-07-23 | 三菱电机株式会社 | Abnormality detection device and abnormality detection method |
| CN109492193B (en) * | 2018-12-28 | 2020-11-27 | 同济大学 | Abnormal network data generation and prediction method based on deep machine learning model |
| CN109492193A (en) * | 2018-12-28 | 2019-03-19 | 同济大学 | Abnormal network data based on depth machine learning model generate and prediction technique |
| CN110321366A (en) * | 2019-06-27 | 2019-10-11 | 中国科学院自动化研究所 | Method and system are determined based on the statistic of on-line study |
| CN110321366B (en) * | 2019-06-27 | 2021-06-25 | 中国科学院自动化研究所 | Statistical quantity determining method and system based on online learning |
| US11561959B2 (en) | 2019-07-15 | 2023-01-24 | Huawei Cloud Computing Technologies Co., Ltd. | Method and system for automatic anomaly detection in data |
| WO2021008142A1 (en) * | 2019-07-15 | 2021-01-21 | Huawei Technologies Co., Ltd. | Method and system for automatic anomaly detection in data |
| CN110469496A (en) * | 2019-08-27 | 2019-11-19 | 苏州热工研究院有限公司 | A kind of water pump intelligent early-warning method and system |
| CN110391840A (en) * | 2019-09-17 | 2019-10-29 | 中国人民解放军国防科技大学 | Method and system for judging abnormality of telemetry parameters of sun synchronous orbit satellite |
| CN111241208A (en) * | 2019-12-31 | 2020-06-05 | 安徽中科大国祯信息科技有限责任公司 | Method and device for monitoring abnormity of periodic time sequence data |
| CN111241208B (en) * | 2019-12-31 | 2024-03-29 | 合肥城市云数据中心股份有限公司 | Abnormality monitoring method and device for periodic time sequence data |
| CN111797250B (en) * | 2020-09-09 | 2020-12-11 | 腾讯科技（深圳）有限公司 | Media content push control method and device, computer equipment and storage medium |
| CN111797250A (en) * | 2020-09-09 | 2020-10-20 | 腾讯科技（深圳）有限公司 | Media content push control method and device, computer equipment and storage medium |
| CN112633329A (en) * | 2020-12-05 | 2021-04-09 | 歌尔科技有限公司 | Method for detecting mechanical property defects of intelligent equipment |
| CN112633329B (en) * | 2020-12-05 | 2023-01-31 | 歌尔科技有限公司 | Method for detecting mechanical property defects of intelligent equipment |
| CN112734977A (en) * | 2020-12-25 | 2021-04-30 | 安徽省安泰科技股份有限公司 | Equipment risk early warning system and algorithm based on Internet of things |
| CN112734977B (en) * | 2020-12-25 | 2022-07-05 | 安徽省安泰科技股份有限公司 | Equipment risk early warning system and algorithm based on Internet of things |
| CN112650660A (en) * | 2020-12-28 | 2021-04-13 | 北京中大科慧科技发展有限公司 | Early warning method and device for power system of data center |
| CN112650660B (en) * | 2020-12-28 | 2024-05-03 | 北京中大科慧科技发展有限公司 | Early warning method and device for data center power system |
| CN113066549B (en) * | 2021-04-06 | 2022-07-26 | 青岛瑞斯凯尔生物科技有限公司 | Clinical effectiveness evaluation method and system of medical instrument based on artificial intelligence |
| CN113066549A (en) * | 2021-04-06 | 2021-07-02 | 青岛瑞斯凯尔生物科技有限公司 | Clinical effectiveness evaluation method and system of medical instrument based on artificial intelligence |
| CN113822580A (en) * | 2021-09-24 | 2021-12-21 | 华润电力技术研究院有限公司 | Equipment working condition evaluation method and related equipment |
| CN114017136A (en) * | 2021-11-02 | 2022-02-08 | 浙江大唐国际绍兴江滨热电有限责任公司 | Soft light character plate alarm method and system for combined cycle generator set |
| CN113806615A (en) * | 2021-11-18 | 2021-12-17 | 成都索贝数码科技股份有限公司 | KPI (Key performance indicator) abnormity early warning method of intelligent IT operation and maintenance system |
| CN114356998A (en) * | 2022-01-08 | 2022-04-15 | 浙江力石科技股份有限公司 | Big data tourism resource data compensation method |
| CN114742391A (en) * | 2022-04-02 | 2022-07-12 | 格莱杰(深圳)科技有限公司 | Artificial intelligence based exception handling method and device, computer equipment and medium |
| CN116302848A (en) * | 2023-05-19 | 2023-06-23 | 杭州安脉盛智能技术有限公司 | Detection method and device for bias of evaluation value, electronic equipment and medium |
| CN116302848B (en) * | 2023-05-19 | 2023-09-01 | 杭州安脉盛智能技术有限公司 | Detection method and device for bias of evaluation value, electronic equipment and medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105302848B (en) | 2018-11-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105302848A (en) | Evaluation value calibration method of equipment intelligent early warning system | |
| CN105117602B (en) | A kind of metering device running status method for early warning | |
| CN102289590B (en) | Method for estimating operating state of SF6 high-voltage circuit breaker and intelligent system | |
| CN109272146A (en) | A kind of Forecasting Flood method corrected based on deep learning model and BP neural network | |
| CN109659933A (en) | A kind of prediction technique of power quality containing distributed power distribution network based on deep learning model | |
| CN105825271B (en) | Satellite failure diagnosis and prediction method based on evidential reasoning | |
| CN108038300A (en) | Optical fiber state evaluating method based on improved membership function combination neutral net | |
| CN106483947A (en) | Distribution Running State assessment based on big data and method for early warning | |
| CN109583520B (en) | State evaluation method of cloud model and genetic algorithm optimization support vector machine | |
| CN114386312A (en) | Equipment fault diagnosis method | |
| Wu et al. | Applying of GA-BP neural network in the land ecological security evaluation | |
| CN111525587B (en) | Reactive load situation-based power grid reactive voltage control method and system | |
| CN104408562A (en) | Photovoltaic system generating efficiency comprehensive evaluation method based on BP (back propagation) neural network | |
| CN108921359A (en) | A kind of distribution gas density prediction technique and device | |
| CN104331773A (en) | Comprehensive assessment method for power network planning schemes | |
| CN108053110A (en) | A kind of transformer state inline diagnosis method based on PMU data | |
| CN106530082A (en) | Stock predication method and stock predication system based on multi-machine learning | |
| CN104239712A (en) | Real-time evaluation method for anti-interference performance of radar | |
| CN110837915A (en) | Low-voltage load point prediction and probability prediction method for power system based on hybrid integrated deep learning | |
| CN110032799A (en) | A kind of the angle similarity divided stages and monitoring method of microbiological pharmacy process | |
| CN115526258A (en) | Power system transient stability evaluation method based on Spearman correlation coefficient feature extraction | |
| CN114548494B (en) | Visual cost data prediction intelligent analysis system | |
| Sun et al. | Pruning Elman neural network and its application in bolt defects classification | |
| CN110533109A (en) | A kind of storage spraying production monitoring data and characteristic analysis method and its device | |
| CN113033898A (en) | Electrical load prediction method and system based on K-means clustering and BI-LSTM neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP01 | Change in the name or title of a patent holder |
Address after: 250101 5th floor, block B, Yinhe building, 2008 Xinluo street, high tech Zone, Jinan City, Shandong Province Patentee after: Shandong luruan Digital Technology Co.,Ltd. Address before: 250101 5th floor, block B, Yinhe building, 2008 Xinluo street, high tech Zone, Jinan City, Shandong Province Patentee before: SHANDONG LUNENG SOFTWARE TECHNOLOGY Co.,Ltd. |
|
| CP01 | Change in the name or title of a patent holder |