CN106548259B - The method of geometry that rotating machinery health status is predicted under completely cut off data qualification - Google Patents
The method of geometry that rotating machinery health status is predicted under completely cut off data qualification Download PDFInfo
- Publication number
- CN106548259B CN106548259B CN201611131006.8A CN201611131006A CN106548259B CN 106548259 B CN106548259 B CN 106548259B CN 201611131006 A CN201611131006 A CN 201611131006A CN 106548259 B CN106548259 B CN 106548259B
- Authority
- CN
- China
- Prior art keywords
- truncated data
- data
- truncated
- rotating machinery
- life 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Entrepreneurship & Innovation (AREA)
- Biomedical Technology (AREA)
- Quality & Reliability (AREA)
- Tourism & Hospitality (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Operations Research (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses the methods of geometry that rotating machinery health status under a kind of completely cut off data qualification is predicted comprising：By being monitored in real time to rotating machinery vibrating, numerous truncation number strong points are obtained；Dimension-reduction treatment is carried out by numerous truncated data point features to numerous truncation number strong points, obtains the truncated data health manifold space with multiple truncated data point features based on Annual distribution；According to the time relationship between multiple truncated data point features in the truncated data health manifold space, link process is carried out to the multiple truncated data point feature, forms the health status evolutional path of time-based truncated data；It is fitted processing by the health status evolutional path to the truncated data, obtains the life value of the truncated data of the rotating machinery vibrating；Verification processing is carried out by the life value to the truncated data, rotating machinery health status is predicted.
Description
Technical field
The present invention relates to electric powder prediction, rotating machinery health status is pre- under especially a kind of completely cut off data qualification
The method of geometry of survey.
Background technology
With social progress and development, people do not only need to know the current state of monitored target, are also required to simultaneously
Understand the state in monitored target future, and then carry out follow-up work, effectively to control risk.However, the essence of prediction is
Seek the dynamic evolution rule in things degenerative process.In recent years, when a large amount of fail data is available, many data-driven sides
Method achieves preferable prediction effect, such as：Smart network, Hidden Markov Model.In addition, some methods attempt to solve it is dilute
Dredge the forecasting problem under the conditions of data or truncated data.However, these research work are often based upon the more considerable history of quantity
Sample data (sparse or block).
Under the premise of essential change does not occur for performance degradation rule, blocking for individual subject data of prediction does not interfere with it
The similitude of its individual manifold structure and sample totality manifold structure.However, when all samples totally be all truncated when (i.e. without times
The completely cut off data mode of what full longevity data), the healthy manifold space of research object is constructed, and ensure its knot as far as possible
The integrality of structure is in turn the difficult point and matter of utmost importance for having to solve for predicting.
Analytic explanation is carried out with the full longevity data instance being made of the 520 of rolling bearing characteristic value point.Utilize manifold
Learning method blocks degree structure healthy manifold space according to different.With being continuously increased for bearing high dimensional feature data volume, structure
The healthy manifold space structure built out is closer to full longevity data entirety manifold structure.Naturally, for being truncated to only have 200
The manifold structure of high dimensional feature data point and final architectural difference are very big.
The present invention using in the manifold space built neighborhood geodesic curve distance and accumulation geodesic curve distance it is strong as quantitative analysis
Health manifold space neutral can fail the Measure Indexes of prediction.Fig. 2 show part different data block under status condition (200,
300,400,500,510,515,520) geodesic curve distance and accumulation geodesic curve is apart from coordinate diagram.From manifold space health status
In the audio-visual picture of Measure Indexes, we can obtain conclusion as before, while provide neighborhood geodesic curve distance and accumulation
Quantitive error analysis between geodesic curve distance.Firstly the need of pointing out, quantitive error analysis here is all with full longevity data sample
The manifold of foundation neighborhood geodesic curve distance spatially and accumulation geodesic curve distance are that reference standard is calculated.It is calculated
Correlated results is as shown in table 1 (by taking neighborhood geodesic curve distance as an example, the result of calculation for accumulating geodesic curve distance is similar).N is in table 1
Truncated data amount M is the data volume that error calculation is carried out within the scope of N.(have light blue with the data instance corresponding to (500,400)
The data of color shade), it indicates when building manifold space with 500 truncated datas, the neighborhood geodesic curve of preceding 400 data points
Absolute error and relative error between distance and preceding 400 neighborhoods geodesic curve distance of full longevity data.
From the data in the table, for any one numerical result for showing meaning, whether relative error or exhausted
To error, error amount is all reducing from top to bottom, illustrates (to block changing for situation with being continuously increased for truncated data sample size
It is kind), the manifold space structure similarity that obtained manifold space structure is built with full longevity data is constantly promoted.
1 truncated data manifold structure error analysis of table
Wherein, AE is absolute error, and RE is relative error
As the above analysis, different data is blocked under state, and the structure of constructed manifold is different, obtains corresponding neighborhood
Geodesic curve distance and accumulation geodesic distance are different.It is then desired to be directed to truncated data state, design and development is a kind of effective strong
Health manifold construction method keeps the manifold space structure obtained through it and full longevity data manifold space structure difference as small as possible, i.e.,
How so that the manifold acquired under the conditions of truncated data is exactly the subset for obtaining manifold under life-cycle data qualification, for rear
Continuous truncated data prediction.
(such as different manifold construction methods：Manifold learning arithmetic), core concept is in global optimization condition
Under, as keep certain geometrical relationship or geometrical property constant as possible, such as：ISOMAP keeps geodesic distance between points not
Become, LLE holding local geometric linear relationships, LTSA keeps local tangent space projection coordinate relationship constant etc..In this way, when being cut
When the new sampled point having no progeny, the manifold that the addition of these new data points can construct former truncated data brings following problem：
1)' neighbour structure ' destroys：The addition of new data point set so that ' neighborhood ' of initial truncated data collection structure occurs
Essence variation, so causes, being the data point of ' neighborhood ' relationship in original truncated data may no longer keep original
' neighborhood ' relationship, this certainly will lead to the reallocation of local geometric relationship, and then change the structure that original blocks manifold；
2)' overall structure ' is destroyed：Block the addition of rear new data point set so that new change occurs for ' neighborhood ' geometrical relationship
Change (even if ' neighborhood ' relationship remains unchanged before), meanwhile, it is this new under the constraints of global optimizing
Geometrical relationship certainty adds global optimizing process, this optimization process necessarily makes the manifold obtained by former truncated data empty
Between in data point (embedded point) change, in turn, change the space structure of former truncated data manifold on the whole.
Invention content
The object of the present invention is to provide under a kind of completely cut off data qualification rotating machinery health status predict it is several where
Method solves above-mentioned ' neighbour according to the data point feature in the time relationship structure manifold space of each data point in truncated data
Domain structure ' destroy the technical issues of being destroyed with ' overall structure '.
The method of geometry that rotating machinery health status is predicted under a kind of completely cut off data qualification of the present invention, including：
By being monitored in real time to rotating machinery vibrating, numerous truncation number strong points are obtained；
By carrying out dimension-reduction treatment to numerous truncated data point features at numerous truncation number strong points, obtains having and be based on
The truncated data health manifold space of multiple truncated data point features of Annual distribution；
According to the time relationship between multiple truncated data point features in the truncated data health manifold space, to described
Multiple truncated data point features carry out link process, form the health status evolutional path of time-based truncated data；
It is fitted processing by the health status evolutional path to the truncated data, obtains the rotating machinery vibrating
Truncated data life value；
Verification processing is carried out by the life value to the truncated data, rotating machinery health status is predicted.
Preferably, by carrying out feature extraction processing to truncated data point, truncation number strong point high dimensional feature is obtained.
Preferably, described to include to numerous truncated data point features progress dimension-reduction treatment：Utilize MLLE algorithms pair
Numerous truncated data point features are calculated, and it is special that numerous truncated data point features are reduced to multiple truncation number strong points
Sign；According to time relationship, Annual distribution processing is carried out to the multiple truncated data point feature, forms the strong of the truncated data
Health manifold space；Wherein, the MLLE algorithms are to improve Local Liner Prediction.
Preferably, in the healthy manifold space according to the truncated data between multiple truncated data point features
Time relationship, carrying out link process to the multiple truncated data point feature includes：To in the healthy manifold space of truncated data
All two adjacent truncated data point features carry out neighborhood geodesic curve distance NGDs processing respectively, obtain in healthy manifold space
The NGDs of the adjacent truncated data point feature of all two based on time relationship；By two all adjacent truncation number strong points
NGDs carries out accumulated process, accumulation geodesic curve distance CGDs is formed in healthy manifold space, to obtain time-based section
The health status evolutional path of disconnected data.
Preferably, the health status evolutional path by the truncated data is fitted processing, obtains institute
The life value for stating the truncated data of rotating machinery vibrating includes：The CGDs is carried out based on future time using Gaussian function
Extension is handled, and obtains extension CGDs values for prediction and with future time.
Preferably, the health status evolutional path process of fitting treatment by the truncated data, obtains the rotation
The life value for turning the truncated data of mechanical oscillation further includes：The extension CGDs values are compared with preset CGDs threshold values；
According to comparison result, the life value of the truncated data of the rotating machinery vibrating is determined.
Preferably, described according to comparison result, determine that the life value of the truncated data of the rotating machinery vibrating includes：
If the extension CGDs values are more than the CGDs threshold values, future time of the CGDs values equal to the CGDs threshold values will be extended and determined
The life value of the truncated data of the rotating machinery vibrating.
Preferably, the life value by the truncated data carries out verification processing, to rotating machinery health shape
State carries out prediction：The life value of the truncated data is verified using trained feedforward neural network FFNN,
And it regard the life value of the truncated data by verification as predicted value.
Preferably, described that the life value of the truncated data is carried out using trained feedforward neural network FFNN
Verification includes：Using the object vector of the life value survival probability of the truncated data as input, trained feed forward neural
Network processes obtain the survival probability of the life value of the truncated data by verification, to utilize the cutting by verification
The survival probability of the life value of disconnected data predicts the service life of the rotating machinery.
' truncated data ' of the present invention is meant：All Condition Monitoring Datas when end-of-life/failure occurs are not arrived.
The method have the benefit that solve ' neighbour structure ' existing for classical manifold space construction process with
The technical issues of ' overall structure ' is destroyed, so as to preferably truncated data be utilized to predict rotating machinery health status.
Description of the drawings
Fig. 1 be the present invention completely cut off data qualification under rotating machinery health status predict method schematic diagram；
Fig. 2 is method shown in Fig. 1 according to the present invention, according to the known detailed process blocked historical data and handled
Figure；
The particular flow sheet of the method for the present invention；
Fig. 3 is the schematic diagram of neighborhood geodesic curve distance and accumulation geodesic curve distance that different data is blocked under state；
Fig. 4 is the schematic diagram of truncated data curve matching and prediction based on accumulation geodesic distance；
Fig. 5 is complete data sample service life and survival probability schematic diagram；
Fig. 6 is bearing data experiment device and schematic diagram；
Fig. 7 is the schematic diagram of the time domain index extracted by bear vibration data；
Fig. 8 is the schematic diagram of the energy indexes extracted by bear vibration data；
Fig. 9 be based on neighborhood geodesic curve distance under MLLEP method difference truncated data status conditions and accumulation geodesic curve away from
From schematic diagram；
Figure 10 be based on MLLEP methods obtain difference block in state health manifold space accumulate geodesic curve apart from affine pass
The schematic diagram of system；
Figure 11 is the schematic diagram that Gaussian 6 is fitted truncated data accumulation geodesic curve distance Curve；
Figure 12 is the schematic diagram for the input that healthy manifold spatially accumulates geodesic curve distance-FFNN networks；
Figure 13 is the contrast schematic diagram of method proposed by the present invention and existing method progress.
Specific implementation mode
Fig. 1 shows the method that rotating machinery health status is predicted under a kind of completely cut off data qualification of the invention, packet
It includes：
By being monitored in real time to rotating machinery vibrating, numerous truncation number strong points are obtained；
By carrying out dimension-reduction treatment to numerous truncated data point features at numerous truncation number strong points, obtains having and be based on
The truncated data health manifold space of multiple truncated data point features of Annual distribution；
According to the time relationship between multiple truncated data point features in the truncated data health manifold space, to described
Multiple truncated data point features carry out link process, form the health status evolutional path of time-based truncated data；
It is fitted processing by the health status evolutional path to the truncated data, obtains the rotating machinery vibrating
Truncated data life value；
Verification processing is carried out by the life value to the truncated data, rotating machinery health status is predicted.
Wherein, by carrying out feature extraction processing to truncated data point, truncation number strong point high dimensional feature is obtained.
Wherein, carrying out dimension-reduction treatment to numerous truncated data point features includes：Using MLLE algorithms to described numerous
Truncated data point feature is calculated, and numerous truncated data point features are reduced to multiple truncated data point features；According to
Time relationship carries out Annual distribution processing to the multiple truncated data point feature, forms the healthy manifold of the truncated data
Space；Wherein, the MLLE algorithms are to improve Local Liner Prediction.
Wherein, in the healthy manifold space according to the truncated data between multiple truncated data point features when
Between relationship, to the multiple truncated data point feature carry out link process include：To institute in the healthy manifold space of truncated data
There are two adjacent truncated data point features to carry out neighborhood geodesic curve distance NGDs processing respectively (that is, according to time relationship, respectively
Calculate the shortest distance between all two adjacent truncated data point features), obtain in healthy manifold space it is all based on
The NGDs of two adjacent truncated data point features of time relationship is (that is, between foring multiple two neighboring truncated data point features
The line of the shortest distance)；The NGDs at two all adjacent truncation number strong points is subjected to accumulated process (that is, by multiple adjacent two
The line of a truncated data shortest distance between points is attached processing), accumulation geodesic curve distance is formed in healthy manifold space
CGDs, to obtain the health status evolutional path of time-based truncated data (that is, the base in healthy manifold space
All truncated data point features are connected into the line segment of truncated data feature in time relationship).
Wherein, the health status evolutional path by the truncated data is fitted processing, obtains described
The life value of the truncated data of rotating machinery vibrating includes：The CGDs based on future time prolonged using Gaussian function
Long processing obtains extension CGDs values for prediction and with future time.
Wherein, the health status evolutional path process of fitting treatment by the truncated data, obtains the rotation
The life value of the truncated data of mechanical oscillation further includes：The extension CGDs values are compared with preset CGDs threshold values；Root
According to comparison result, the life value of the truncated data of the rotating machinery vibrating is determined.
Wherein, described according to comparison result, determine that the life value of the truncated data of the rotating machinery vibrating includes：If
The extension CGDs values are more than the CGDs threshold values, then will extend future time of the CGDs values equal to the CGDs threshold values and determine institute
State the life value of the truncated data of rotating machinery vibrating.
Wherein, the life value by the truncated data carries out verification processing, to rotating machinery health status
Carrying out prediction includes：The life value of the truncated data is verified using trained feedforward neural network FFNN, and
It regard the life value of the truncated data by verification as predicted value.
Wherein, described that the life value of the truncated data is tested using trained feedforward neural network FFNN
Card includes：Using the object vector of the life value survival probability of the truncated data as input, trained Feedforward Neural Networks
Network processing, obtains the survival probability of the life value of the truncated data by verification, to utilize the blocking by verification
The survival probability of the life value of data predicts the service life of the rotating machinery.
Above-mentioned estimates the survival probability of the life value of truncated data using product-limit estimator device PLE, is cut
The life value survival probability of disconnected data.
Above-mentioned being built according to the life value survival probability of truncated data is used as the trained feedforward neural network
The object vector of trained input.
The neighborhood geodesic curve distance NGDs of the present invention：Geodesic curve and geodesic distance have reacted healthy manifold spatially arbitrary two
The distance between point relationship is embedded in manifold space neutral energy forecasting problem for higher-dimension, and neighborhood geodesic curve distance indicates health stream
In shape space, any state point of the object in evolution degenerative process is with its neighborhood (by the order dependent radius of neighbourhood of predicted time
Determine) between state point apart from degree of a relation amount, to describe healthy manifold space local geometry.
The accumulation geodesic curve distance CGDs of the present invention：Accumulate geodesic curve distance indicate healthy manifold space any state point with
Geodesic curve distance between original state point (state point in prediction Object Evolution degenerative process), is appointed in manifold of higher dimension space
Meaning state deviates the measurement of original state.
The thinking and particular technique point of the solution technical problem of the present invention are carried out specifically with reference to Fig. 2-Fig. 8
It is bright.
The thinking solved the problems, such as
1) whether truncated data collection or full longevity data set, manifold space construction process of the same race are all to maintain data point
Between certain geometrical relationship or characteristic it is constant, this geometrical relationship or characteristic are present in naturally in former high-dimensional feature space.
Thus when ' neighborhood ' relationship is fixed, the geometrical property between this point kept under the conditions of truncated data and full longevity data qualification
Under geometrical property between obtained these points should differ an affine transformation；
2) existing manifold space construction process is mainly used in the fields such as pattern-recognition, classification, assessment and image procossing.
Manifold structure in these areas is all the integral transformation of all data point sets, few the problem of increasing point set newly occur.However
When there is new point set to add, these points are also considered the generalized point without any particularity, and are added to and are totally divided again
Class --- this process is invalid to predicting；
3) for the forecasting problem in health management system arranged, the structure of manifold has following property：
● it predicts to be permanently present inevitable contact with the time；
● from time series theory[111], prediction result only (office related to a certain range of partial data
Domain)；
● the introducing of new data point will not have an impact historical data and data structure；
It follows that the property of certainty and prediction theory itself based on healthy manifold space structure, it can be by drawing
Angle of incidence parameter and fixed manifold of higher dimension space ' neighborhood ' range, and then determine and cure its ' neighborhood ' structure, it is above-mentioned to solve
The problems such as ' neighbour structure ' is destroyed with ' overall structure ', important foundation is provided for performance degradation prognosis under truncation condition.
Fig. 2 shows the above method according to the present invention, according to the known flow chart for blocking historical data and being handled.
Healthy manifold developing algorithm towards completely cut off data
The present invention will combine the above-mentioned thinking solved the problems, such as, classical manifold learning arithmetic be improved, to meet truncated data item
Healthy manifold space neutral can fail the needs of prediction under part.Note：It is substantially same that manifold learning proposed by the present invention improves thought
When be also applied for all classical manifold learnings, such as：ISOMAP, LE, LTSA etc..The present invention carries out algorithm only by taking LLE as an example
Curve guide impeller, and work applied to subsequent prediction.
On the basis of the theory of LLE methods is with algorithm, LLE algorithms are improved, obtain improving being locally linear embedding into
(Modified local linear embedding for prognostics, MLLEP) algorithm description is as follows：
(1) neighborhood (improvement part content) is chosen
Calculate each sample point XiNeighborhood point, if sample points are k in neighborhood, k determines by the correlation of time series,
Then XiAny neighborhood point be represented by Xij, j=1 ..., k, and XijWith point XiThere is regular hour continuity；
(2) reconstruct power (former LLE) is calculated
For each sample point XiAnd its neighborhood Xij, j=1 ..., k, by optimal reconfiguration error function
In turn, all X are obtainediWeight vector and all sample points weight matrix；
(3)It calculatesdDimension is embedded (former LLE)
Calculating matrix M=(I-W)T(I-W) minimum d+1 feature vector u2,…,ud+1, then T=[u2,…,ud+1]TI.e.
To calculate the insertion result of gained.
Geodesic fitting and prediction are accumulated in healthy manifold space
It is illustrated in figure 4 the main process of truncated data object performance degradation prognosis method of geometry in healthy manifold space.
The healthy manifold obtained by MLLEP methods spatially, calculate accumulation geodesic curve distance Curve, and then carry out accumulation geodesic curve away from
Fitting from curve and prediction.And the process is the necessary ring of method of geometry be achieved one of truncated data forecasting problem
Section.
Associate cumulation geodesic curve of the present invention apart from the characteristics of, using in MATLAB analysis tools (CFTOOL)
' Gaussian 6 ' realizes the fitting to accumulating geodesic curve distance Curve, and then obtains the service life of truncated data object.
Performance degradation prognosis based on intelligent product-limit estimator device under completely cut off data qualification
Feedforward neural network
Feedforward neural network (Feed Forward Neural Network, FFNN) model is by input layer, hidden layer and defeated
Go out three layers of composition of layer.The form of FFNN input vectors isTotal niA input node, output vector form
For Tk={ Sk+Δ,Sk+2Δ,...,Sk+hΔTotal h output node.In the present invention, the input of FFNN is healthy manifold space neutral
The Measure Indexes data that can be failed export as the survival probability in research object following a period of time section.
According to identified network structure FFNN networks are built using structure FFNN training samples and test sample
Training and test sample input vector, and according to existing survival probability build training sample object vector.Completion pair
The training of FFNN networks, and eventually for following survival probability of estimation to realize performance degradation prognosis.
Survival probability estimation based on intelligent product-limit estimator device
Kaplan and Meier in 1958 is directed to a kind of Nonparametric Estimation that Random censorship problem proposes first[112], K-
M estimations, i.e. product-limit estimator device (Product limit estimator, PLE).PLE, which is effectively realized, estimates survival probability
Meter, is widely used in the field with fragmentary data problem.
For including sample the totality Γ, sample t of Right censored data and lifetime data simultaneously1,t2,…,tn, wherein n
For sample size.Work as tiWhen being Right censored data, δ is enabledi=0, work as tiIt is when dying of old age data, to enable δi=1；, this group data can be denoted as
(ti,δi) i=1,2 ..., n.
By these tiRearrangement (when lifetime data is equal with Right censored data, lifetime data is come and deletes mistake by size
Before data) then have
t(1)≤t(2)≤...≤t(n)
The product-limit estimator of S (t) is defined by the formula：
Training parameter collection structure under full longevity data qualification
The survival condition parameter set T of each data sample is collectively constituted by the survival probability of each time interval k.Data
The survival probability of sample is configured to can be used for FFNN training according to the output vector form (including output node number) of FFNN networks
Data.For full longevity data, the survival probability before equipment failure is 1, and the survival probability after failure is 0, as shown in Figure 5.Example
Such as, the data acquisition system of certain bearing failed (failure in the 3rd time interval) at the 22nd day, when FFNN network output node numbers are 5,
Then Tk={ 1,1,0,0,0 }, wherein k=0.
Training parameter collection structure under the conditions of truncated data
If equipment has just been withdrawn from reaching scheduled failure threshold not yet, data sample is truncated data.For
Truncated data, similarly, survival probability remains as 1 before data are truncated, the survival probability after blocking then need by
IPLE methods are estimated that calculation formula is as follows：
L (i) is the last observation period of i-th of sample.rk+n-1It is inefficacy ratio of the equipment in k+n-1 time intervals, counts
Calculation formula is rk+n-1=Fk+n-1/Rk+n-1, Fk+n-1For the failure number in the moment from k+n-2 to k+n-1, Rk+n-1For k+n-2 when
The normal operation number at quarter.
For example, totally 30 equipment are in operating status.Assuming that there is 1 bearing failure in+1 time interval of kth；Kth+
Separately there are 2 failures 1 to withdraw from 2 time intervals；There are 3 failures in+3 time interval of kth.For in+2 time interval of kth
The bearing inside withdrawn from, the training vector in k-th of time interval, the value before+2 time intervals of kth are all 1, i.e.,
Tk=1,1,,,(by taking FFNN network output node numbers are equal to 5 as an example).Training vector value in time interval later
It needs to be calculated with iPLE methods.First, it is understood that there is failure risk (only to consider in monitoring in the 3rd time interval
Equipment, do not consider to have failed and withdrawn from equipment) number of devices be Rk+3=30-1-2-1=26.If the equipment withdrawn from is not
It is withdrawn by, failure probability rk+3=Fk+3/Rk+3=3/26=0.15.Further, it is assumed that separately have 3 in the 4th time interval
Equipment failure, 2 are withdrawn from.Similarly, rk+4=Fk+4/Rk+4=3/23=0.1304.In the 5th time interval, there are 4 failures,
Then rk+5=Fk+5/Rk+5=4/18=0.222.There is S according to formula abovek+1=1.0；Sk+2=1.0；Sk+3=1.0 (1.0-
0.15)=0.85；Sk+4=0.85 (1.0-0.1304)=0.7391；Sk+5=0.7391 (1.0-0.222)=0.575, that is, exist
The training vector that equipment is withdrawn from k-th of time interval is Tk={ 1,1,0.850,0.7391,0.575 }.For each output
Node k+i, each training signal represent survival probability of the input sample in sample set.
The method of geometry of truncated data object performance degradation prognosis in healthy manifold space
The present invention is with rolling bearing data, the performance degradation of truncated data properties object in analytic explanation health manifold space
Prediction technique and process.
Experimental test device and data description
As shown in fig. 6,4 bearings are mounted on same main shaft, rotary speed 2000RPM keeps normal lubrication shape
State.6000lb radial loads are applied to bearing 2 and 3 simultaneously.Each two PCB353B33 high sensitivity ICP acceleration of Bearing configuration pass
Sensor is to acquire the vibration acceleration signal of X and Y both directions, and in experimentation, calendar time sampling in every 10 minutes is primary, adopts
Sample continues 1s sample frequencys 20kHz.In view of bearing 3 is by the direct effect of radial load, and event occurs when testing termination
Barrier has full longevity data qualification.The present invention is based on the original vibration signals of bearing 3, and 40 groups of full longevity data samples are obtained through emulation
This[113], and then simulate generation difference and block status data, provide necessary data basis for follow-up test and analysis of cases.
For 40 groups of full longevity data that emulation obtains, every group of data sample determines data sample length by the service life, and
(it is a time step between each two sample point, it is small to represent 2.4 by one time step by different sample time point t
When), every 10 sample time points indicate 1 time interval (one time interval are represented 1 day).Accordingly, for one group
The full longevity data for possessing 384 sample points, then have 39 time intervals (i.e.：39 days service life).
Table 2 is shown obtains 40 groups of data time step-lengths, time interval span and the letters such as moment, life value that fail through emulation
Breath.In order to simulate full truncated data situation, the present invention fails the later stage in bearing performance (during apparent decline occurs), to emulating axis
It holds data and carries out Random Truncation Data processing (being to block unit with day), specifically block situation referring to table 2.In view of needing to utilize simultaneously
These emulation data carry out test and validation to the performance degradation prognosis method under the conditions of truncated data proposed by the present invention, from volume
Number to randomly select 20 groups of data in 1~30 group of data to ' truncated data object performance degradation prognosis in healthy manifold space
Method of geometry ' carry out test verification；And using obtained 20 life prediction values as the life value of this 20 groups of data, and then shape
At 20 groups of full longevity data, with remaining 10 groups of sample datas composition part truncated data sample in the 1st~30 group；It is counted using iPLE
The survival probability of each sample is calculated, finally using the 31st~40 group of truncated data to ' performance based on intelligent product-limit estimator device declines
Move back prediction：The combination of geometry and non-geometric method ' carry out test and validation.As shown in table 2, it is the 1st to have the data of shaded background
20 groups randomly selected in~30 groups of data, which need to be fitted, to be predicted and obtains the sample of corresponding life prediction value, and is used directly to
It verifies ' method of geometry of truncated data object performance degradation prognosis in healthy manifold space '.
The training sample of 2 model of table
' the performance degradation prognosis based on intelligent product-limit estimator device：The combination of geometry and non-geometric method ' test sample
Performance degradation feature and data truncation status analysis under completely cut off data qualification
As previously mentioned, performance degradation feature extraction is premise and the basis of all prediction works.In view of truncated data sheet
For body there is the incompleteness of data information, failing the screening of information for this categorical data will be more stringent, so as to can be
It is as much as possible under the conditions of limited data resource to reflect performance degradation characteristic.
Fig. 7 show 4 time domain indexes extracted from the full longevity vibration data of bearing, including：RMS (root-mean-square value), peak value
The factor, kurtosis, nargin.Fig. 8 show 8 energy indexes that the full longevity vibration data of bearing is extracted through wavelet package transforms.
By Fig. 7 and Fig. 8 it is found that RMS and 8 energy value of bearing preferably reacts its performance degradation state, and conduct
The high dimensional feature (9) that the present invention is extracted.Meanwhile by diagram it is found that bearing performance decline course in, starting stage property
The speed that can be failed is slow, and kept stable whithin a period of time.Enter rapid decline after reaching certain recession level
Process, and rapid failure.Therefore, it in order to solve the problems, such as the performance degradation prognosis under completely cut off data qualification, still needs to a certain amount of
Performance degradation status data, to seek the essential Evolution of sustainable prediction.
Completely cut off data manifold of higher dimension space structure based on MLLEP
The content of present invention will extract the high dimensional feature (9 feature compositions) of bear vibration data in upper section, utilize the side MLLEP
Method builds the healthy manifold space of truncated data point set, with obtain under the conditions of different truncated datas obtained through MLLEP methods it is strong
Health manifold neighborhood geodesic curve distance spatially and accumulation geodesic curve distance.Based on this, completes MLLEP methods and be applied to truncation number
Compliance test result is built according to manifold space.
Test analysis is carried out with the original full longevity data instance being made of 520 data points.Neighborhood k=8, spatial context dimension
D=3 is spent, using MLLEP be successively 450,510,515 to the state of blocking and the complete healthy manifold of longevity data progress is built, and is calculated
The neighborhood geodesic curve distance and accumulation geodesic curve distance of each truncated data state, as shown in Figure 9.
By taking truncated data state is 450 and 510 as an example, calculates healthy manifold of the two through MLLEP structures and spatially accumulate
Affine scaling relationship between geodesic curve distance.It is computed, affine transformation coefficient k _ affine=0.8399.It is as shown in Figure 10
Accumulation geodesic distance curve that ' 510 ' a truncation number strong points directly build and are calculated through MLLEP methods and through affine transformation
The affine accumulation geodesic curve distance Curve obtained afterwards (establish imitative by the accumulation geodesic curve distance Curve established with ' 450 ' truncated datas
Penetrate relationship).
By Figure 10 and its data analysis it is found that MLLEP, which effectively maintains original difference, blocks status data and full longevity number
According to the geometric properties in high-dimensional feature space, this provides necessary precondition for the development of subsequent prediction work.
Completely cut off data object performance degradation prognosis based on MLLEP
Using the 20 groups of truncated datas and original full longevity data randomly selected in table 2, respective health is built using MLLEP
Manifold space, and neighborhood geodesic curve distance and accumulation geodesic curve distance are extracted in healthy manifold space, it is based on ' Gaussian
6 ' complete the fitting and prediction of accumulation geodesic curve distances, and then realize ' truncated data object performance degradation in healthy manifold space
The method of geometry of prediction ' test and validation.With original full longevity data instance to the completely cut off data object based on MLLEP
It can decline prediction progress analytic explanation (original full longevity data are through artificial truncation).
The manifold space of truncated data is built using MLLEP, and calculates accumulation geodesic curve distance, it is right using Gaussian 6
Truncated data accumulation geodesic curve distance Curve is fitted and predicts, it is as follows to obtain correlated results：
Fit found when optimization terminated:
General model Gauss6:
F (x)=
a1*exp(‐((x‐b1)/c1)^2)+a2*exp(‐((x‐b2)/c2)^2)+
a3*exp(‐((x‐b3)/c3)^2)+a4*exp(‐((x‐b4)/c4)^2)+
a5*exp(‐((x‐b5)/c5)^2)+a6*exp(‐((x‐b6)/c6)^2)
Coefficients (with 95%confidence bounds):
A1=-3.721e+005 (- 1.455e+007,1.381e+007)
B1=566.5 (186.5,946.5)
C1=30.49 (- 81.69,142.7)
A2=1.666e+009 (- 3.795e+011,3.828e+011)
B2=723.6 (- 3941,5388)
C2=62.36 (- 651.2,775.9)
A3=-347.6 (- 706.7,11.47)
B3=287 (259.8,314.1)
C3=94.16 (57.29,131)
A4=-266.5 (- 512.1, -20.89)
B4=427.1 (412.6,441.7)
C4=32.97 (14.97,50.96)
A5=30.92 (- 34.55,96.39)
B5=404.2 (393.1,415.4)
C5=9.422 (- 12.09,30.94)
A6=5297 (4348,6246)
B6=505.6 (432.1,579.1)
C6=247.2 (212.9,281.5)
Goodness of fit:
SSE:1.424e+006
R‐square:0.9993
Adjusted R‐square:0.9993
RMSE:53.25
By Figure 11 and fitting precision analysis it is found that 6 functions of Gaussian can be fitted truncated data health manifold well
Accumulation geodesic distance curve spatially.Further, to number in table 2 be the 1st~30 in 20 groups of differences randomly selecting cut
Disconnected status data carries out identical operating process, and is respectively blocked the life prediction value under state.
By above-mentioned curve matching precision and truncated data manifold of higher dimension space neutral energy metric parameter (accumulation geodesic curve away from
From) prediction it is found that 6 curves of Gaussian can effectively realize curve matching, and in certain fiducial interval range, can be obtained
Each corresponding life value of truncated data.And then on the basis of MLLEP builds manifold of higher dimension space, solves truncated data performance
Fail forecasting problem.
On the basis of the method for geometry that subsequent content of the present invention will be predicted in the truncated data proposed, in conjunction with intelligent product
Estimator method (data with existing can be efficiently used) is limited, the probability expression (survival probability) of performance degradation prognosis is provided, and is predicted
The service life of truncated data object.
Performance degradation prognosis based on intelligent product-limit estimator device：The geometry of health status is predicted
Present invention aims at：(1) iPLE is given full play to using can be highlighted in the validity of data and healthy manifold space
The advantage of Evolution, preferably to solve the forecasting problem of truncated data；(2) realize the method for geometry in healthy manifold space with
The combination of non-geometric method provides technical support for follow-up study.
Survival probability based on iPLE methods calculates
It is 1~30 data as the ' performance degradation prognosis based on intelligent product-limit estimator device using number in table 2：Geometry
And the combination of non-geometric method ' training sample (wherein dash area truncated data passed through 5.5 section contents --- ' health flow
The method of geometry of truncated data object performance degradation prognosis in shape space ' it is fitted and predicts to obtain life value, it regards in the present invention
For full longevity data), the data that number is 31~40 are as test sample.
According to the computational methods of survival probability described in 5.4 sections, based on data in table 2, complete data sample is built
It is as follows with survival probability under truncated data sample conditions.
(1) structure (being obtained with prediction through fitting) of full longevity data sample
It has been complete lifetime data to have 20 groups of data samples in table 2, according to the method in 5.4 sections, by taking No. 1 bearing as an example,
Assuming that 10 data points are a time interval, the bearing failure in t=363 is can see from table, i.e., in the 37th time
It fails in section.Then for the training sample of No. 1 bearing, survival probability Sk=1 (k=1 ..., 36), works as k>When 36, Sk=0.
(2) structure of truncated data sample
The survival probability method of 10 groups of test truncated datas in table is as described in 5.4.Based on 30 groups of truncated data samples
Out-of-service time (be fitted and predicted time) and time to chopping (Withdrawal Hour), rule, reference formula (5.3) are estimated according to K-M
It calculates 10 groups and blocks survival probability value of the bearing data in each time interval.
FFNN network structions
(1) structure of FFNN models
FFNN network models of the present invention include three layers, and input layer contains 7 nodes, input vector { xt,xt-1,xt-2,xt-3,
xt-4,xt-5,xt-6It is bearing truncated data in the current time t and before this accumulation geodesic curve distance at 6 moment, hidden layer respectively
10 nodes are set, output node number is 5, shaped like { Sk+1,Sk+2,Sk+3,Sk+4,Sk+5, represent the time zone residing for t moment
Between survival probability in 5 time intervals after k.
(2) training of structure FFNN prediction models and test sample
According to identified network structure, using the accumulation geodesic curve of 30 groups of training samples and 10 groups of test samples apart from number
According to (shaped like Figure 12), the training of FFNN networks and the input vector of test sample are built, and build according to existing survival probability
The object vector of training sample.Table 3 and table 4 give in FFNN model training input/output, and test input building process
Partial data and its structure type.
The input vector list of 3 FFNN network models of table
The object vector list of 4 FFNN network models of table
Training and test based on the prediction of intelligent product-limit estimator device
FFNN networks use error back propagation (Error back propagation, BP) algorithm, and transmission function is
Tansig and logsig, training function select the momentum gradient decreasing function traingdx of adaptive lr, frequency of training
Net.trainParam.epochs=3000, training objective precision are defaulted as 0.Utilize the instruction built in step in 5.6.2 (2)
Practice sample to be trained FFNN networks, and then completes the test for 10 groups of truncation test samples for being 31~40 to number.With 31
For number rolling bearing, obtain that test result is as follows shown in table.
The failure prediction result of 5 No. 31 bearings of table
As shown in table 5, in time interval k=27, the data of dash area are that first survival probability of each column is less than 0.5
Value, be based on this, the calculation formula in section of failing can be expressed as follows：
Wherein, n(k+i)To first appear first number less than 0.5 on kth+i prediction levels；Using the formula, calculate
The bimetry of No. 31 bearings is 29.25 days, i.e., fails in the 30th time interval.Similarly, to other bearing test samples, profit
Survival probability is predicted with intelligent product estimation device.Table 6 is the prediction for 10 bearing test samples that number is 31~40
As a result statistical form, it can be seen that the phase closing precision and error of prediction result and true lifetime are all more satisfactory, remove #32 and No. #36
Outside bearing, precision of prediction is all more than 95%.It follows that in the case where blocking training data sample entirely, based on MLLEP structure health streams
Shape space is fitted and predicts that the accumulation geodesic curve distance of manifold spatially obtains the life value of part truncated data, recycles intelligence
Energy product-limit estimator device predicts survival probability, more satisfactory prediction effect can be still obtained, to realize this hair
The bright effective combination for proposing method of geometry and iPLE methods lays the foundation for follow-up study work.
The survival probability forecast statistics result of 6 10 test samples of table
Although describing the invention in detail above, but the invention is not restricted to this, those skilled in the art of the present technique
It can be carry out various modifications with principle according to the present invention.Therefore, all to be changed according to made by the principle of the invention, all it should be understood as
Fall into protection scope of the present invention.
Claims (9)
1. the method for geometry that rotating machinery health status is predicted under a kind of completely cut off data qualification, including：
By being monitored in real time to rotating machinery vibrating, numerous truncation number strong points are obtained；
Dimension-reduction treatment is carried out by numerous truncated data point features to numerous truncation number strong points, obtains having based on the time
The truncated data health manifold space of multiple truncated data point features of distribution；
According to the time relationship between multiple truncated data point features in the truncated data health manifold space, to the multiple
Truncated data point feature carries out link process, forms the health status evolutional path of time-based truncated data；
It is fitted processing by the health status evolutional path to the truncated data, obtains cutting for the rotating machinery vibrating
The life value of disconnected data；
Verification processing is carried out by the life value to the truncated data, rotating machinery health status is predicted.
2. according to the method described in claim 1, wherein, by carrying out feature extraction processing to truncated data point, being blocked
Data point high dimensional feature.
3. described to carry out dimension-reduction treatment to numerous truncated data point features according to the method described in claim 2, wherein
Including：
Numerous truncated data point features are calculated using MLLE algorithms, numerous truncated data point features are reduced
To multiple truncated data point features；
According to time relationship, Annual distribution processing is carried out to the multiple truncated data point feature, forms the truncated data
Healthy manifold space；
Wherein, the MLLE algorithms are to improve Local Liner Prediction.
4. according to the method in claim 2 or 3, wherein in the healthy manifold space according to the truncated data
Time relationship between multiple truncated data point features, carrying out link process to the multiple truncated data point feature includes：
Neighborhood geodesic curve is carried out respectively to all two adjacent truncated data point features in the healthy manifold space of truncated data
Distance NGDs processing obtains all two adjacent truncated data point features based on time relationship in healthy manifold space
NGDs；
The NGDs at two all adjacent truncation number strong points is subjected to accumulated process, forms accumulation geodetic in healthy manifold space
Linear distance CGDs, to obtain the health status evolutional path of time-based truncated data.
5. according to the method described in claim 4, wherein, the health status evolutional path by the truncated data
It is fitted processing, the life value for obtaining the truncated data of the rotating machinery vibrating includes：
The extension based on future time is carried out to the CGDs to handle, obtain for prediction and carry future using Gaussian function
The extension CGDs values of time.
6. according to the method described in claim 5, wherein, the health status evolutional path by the truncated data
Process of fitting treatment, the life value for obtaining the truncated data of the rotating machinery vibrating further include：
The extension CGDs values are compared with preset CGDs threshold values；
According to comparison result, the life value of the truncated data of the rotating machinery vibrating is determined.
7. it is described according to comparison result according to the method described in claim 6, wherein, determine the rotating machinery vibrating
The life value of truncated data includes：
If the extension CGDs values are more than the CGDs threshold values, future time of the CGDs values equal to the CGDs threshold values will be extended
Determine the life value of the truncated data of the rotating machinery vibrating.
8. according to the method described in claim 1, wherein, the life value by the truncated data carries out at verification
Reason, carrying out prediction to rotating machinery health status includes：
The life value of the truncated data is verified using trained feedforward neural network FFNN, and verification will be passed through
Truncated data life value as predicted value.
9. according to the method described in claim 8, wherein, the trained feedforward neural network FFNN of the utilization is to described
The life value of truncated data carries out verification：
Using the object vector of the life value survival probability of the truncated data as input, at trained feedforward neural network
Reason, obtains the survival probability of the life value of the truncated data by verification, to utilize the truncated data by verification
The survival probability of life value predict service life of the rotating machinery.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611131006.8A CN106548259B (en) | 2016-12-09 | 2016-12-09 | The method of geometry that rotating machinery health status is predicted under completely cut off data qualification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611131006.8A CN106548259B (en) | 2016-12-09 | 2016-12-09 | The method of geometry that rotating machinery health status is predicted under completely cut off data qualification |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106548259A CN106548259A (en) | 2017-03-29 |
| CN106548259B true CN106548259B (en) | 2018-08-21 |
Family
ID=58397186
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611131006.8A Active CN106548259B (en) | 2016-12-09 | 2016-12-09 | The method of geometry that rotating machinery health status is predicted under completely cut off data qualification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106548259B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110991945B (en) * | 2018-09-30 | 2023-04-18 | 北京国双科技有限公司 | Information processing method and device for equipment spare part, storage medium and processor |
| CN111291986A (en) * | 2020-01-22 | 2020-06-16 | 中国人民解放军空军工程大学 | Air defense reverse guidance operation action accurate equipment guarantee scheme planning system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO1997044757A2 (en) * | 1996-05-21 | 1997-11-27 | Interval Research Corporation | Principal component analysis of image/control-point location coupling for the automatic location of control points |
| CN103324834A (en) * | 2013-04-07 | 2013-09-25 | 北京航空航天大学 | Service life prediction method for electromechanical system and critical components under completely truncated data condition |
| CN103439666A (en) * | 2013-09-02 | 2013-12-11 | 北京航空航天大学 | Geometric method for evaluating capacity recession of lithium ion battery |
| CN104463246A (en) * | 2014-12-08 | 2015-03-25 | 天津大学 | Manifold-based linear regression learning method |
-
2016
- 2016-12-09 CN CN201611131006.8A patent/CN106548259B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO1997044757A2 (en) * | 1996-05-21 | 1997-11-27 | Interval Research Corporation | Principal component analysis of image/control-point location coupling for the automatic location of control points |
| CN103324834A (en) * | 2013-04-07 | 2013-09-25 | 北京航空航天大学 | Service life prediction method for electromechanical system and critical components under completely truncated data condition |
| CN103439666A (en) * | 2013-09-02 | 2013-12-11 | 北京航空航天大学 | Geometric method for evaluating capacity recession of lithium ion battery |
| CN104463246A (en) * | 2014-12-08 | 2015-03-25 | 天津大学 | Manifold-based linear regression learning method |
Non-Patent Citations (2)
| Title |
|---|
| 基于非对称双边截断投影数据的CT重建方法;韩旭等;《全国射线数字成像与CT新技术研讨会论文集》;20160930;全文 * |
| 对数正态分布下基于截断数据的统计分析;李林元;《南京航空学院学报》;19911231;第23卷;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106548259A (en) | 2017-03-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110188397B (en) | Model and method for predicting icing of overhead transmission line | |
| Li et al. | A hybrid model based on synchronous optimisation for multi-step short-term wind speed forecasting | |
| Kapoor et al. | A grey wolf optimizer based automatic clustering algorithm for satellite image segmentation | |
| Okumus et al. | Current status of wind energy forecasting and a hybrid method for hourly predictions | |
| Amjady et al. | Short-term wind power forecasting using ridgelet neural network | |
| Liang et al. | A multi-variable stacked long-short term memory network for wind speed forecasting | |
| Liu et al. | A hybrid WA–CPSO-LSSVR model for dissolved oxygen content prediction in crab culture | |
| Vafaeipour et al. | Application of sliding window technique for prediction of wind velocity time series | |
| Shukla et al. | Clustering based unit commitment with wind power uncertainty | |
| Ghayekhloo et al. | A combination approach based on a novel data clustering method and Bayesian recurrent neural network for day-ahead price forecasting of electricity markets | |
| Wang et al. | A compound framework for wind speed forecasting based on comprehensive feature selection, quantile regression incorporated into convolutional simplified long short-term memory network and residual error correction | |
| Jalalkamali | Using of hybrid fuzzy models to predict spatiotemporal groundwater quality parameters | |
| Butler et al. | Exploiting SCADA system data for wind turbine performance monitoring | |
| Kabir et al. | Partial adversarial training for neural network-based uncertainty quantification | |
| Wang et al. | A hybrid optimization-based recurrent neural network for real-time data prediction | |
| CN112598248B (en) | Load prediction method, load prediction device, computer equipment and storage medium | |
| Carta et al. | Comparison of feature selection methods using ANNs in MCP-wind speed methods. A case study | |
| CN103324834B (en) | Mechatronic Systems under a kind of completely cut off data qualification and critical component life-span prediction method thereof | |
| CN112884008B (en) | Prediction evaluation method and device for running state of power information acquisition system | |
| Colak et al. | Multi-time series and-time scale modeling for wind speed and wind power forecasting part I: Statistical methods, very short-term and short-term applications | |
| Zhu et al. | Wind power forecasting based on new hybrid model with TCN residual modification | |
| Alamaniotis et al. | Monthly load forecasting using kernel based gaussian process regression | |
| CN111242351A (en) | Tropical cyclone track prediction method based on self-encoder and GRU neural network | |
| CN107909212A (en) | Climbing event prediction method based on mesoscale wind power evolution serializing | |
| CN106548259B (en) | The method of geometry that rotating machinery health status is predicted under completely cut off data qualification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20200617 Address after: No. 214, xiaoxinzhuang village, Beifang Town, Huairou District, Beijing Patentee after: Beijing Hengxing Yikang Technology Co., Ltd Address before: 100191 Beihang University, Haidian District, Xueyuan Road, 37, Beijing Patentee before: BEIHANG University |