CN110298385A - One kind is based on * information and the online incipient fault detection method of increment SVDD - Google Patents
One kind is based on * information and the online incipient fault detection method of increment SVDD Download PDFInfo
- Publication number
- CN110298385A CN110298385A CN201910486042.3A CN201910486042A CN110298385A CN 110298385 A CN110298385 A CN 110298385A CN 201910486042 A CN201910486042 A CN 201910486042A CN 110298385 A CN110298385 A CN 110298385A
- Authority
- CN
- China
- Prior art keywords
- sample
- increment
- svdd
- model
- follows
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
- G06V30/194—References adjustable by an adaptive method, e.g. learning
Abstract
The invention proposes one kind to be based onThe online incipient fault detection method of information and increment SVDD.It is based onThe online incipient fault detection method of information and increment Support Vector data description SVDD, is first standardized the process measurement data under nominal situation, obtains process training sample set and correspondingThe value of information;Calculate each measurand withAssociation relationship between effect filters out characteristic variable relevant to efficiency according to accumulative mutual information contribution rate;SVDD initial model is established, adaptive updates association relationship and more new increment SVDD model are carried out for the newly-increased sample of process respectively, obtain the increment SVDD state model under nominal situation;The model under different conditions is obtained according to the measurement data of different conditions, and is used for online incipient fault detection.The advantages of present invention incorporates based on ENERGY METHOD and based on data-driven method, during not only can solve the problem of system parameter time-varying characteristics, energy feature variable during can also efficiently extracting, the rate and precision for improving incipient fault detection, effectively realize the on-line checking of initial failure.
Description
Technical field
The present invention relates to process initial failure on-line monitoring fields more particularly to one kind to be based onInformation and increment support to
Measure the online incipient fault detection method of data description (Support Vector Data Description, SVDD).
Background technique
SVDD is single class sorting algorithm based on support region, has the description of intuitive data and good popularization performance, also
It can solve the problems, such as that real process failure classes sample is difficult to obtain, have been widely used for the fields such as fault diagnosis.But SVDD
Need to solve quadratic programming problem, training complexity is higher, and it trains complexity and sample number exponentially grade relationship.?
In SVDD fault detection research, in order to reduce SVDD computation complexity, while the validity of sample data is improved, usually with statistics
Method carries out Feature Dimension Reduction and extracts feature samples, then establishes SVDD disaggregated model using this feature sample to realize failure
Detection.As the concept for combining the first and second law of thermodynamics, it can be used to more fully understand that process, quantization are inefficient
Direction and differentiation energy quality, it is the multi-field universal concept such as physics, chemistry, Mechanics of Machinery, is usedConcept also can
Reduce data dimension.It is based onModeling work amount can be effectively reduced in the method for information, increases computational efficiency, in model dimensionality reduction
While remain significantly similitude.Effect is whole in Chemical Processing SystemsThe specific manifestation form of information, and
And be also the evaluation index of process integral energy quality, so being based onThe method of information extraction is will in the processEffect is extracted
Out, to achieve the effect that dimensionality reduction.
However in actual industrial processes, systematic procedure parameter can change with the propulsion of time, therefore
Barrier feature can also change therewith, and Fault Model needs the newly-increased sample of training, and SVDD has to give up original instructed at this time
The model perfected, need to by new data together with historical data re -training, by its algorithm complexity it is found that with new data sample
It is continuously updated, the computation complexity of the algorithm exponentially increases.Once there is the case where mass data sample on-line training,
The training process of SVDD can waste a large amount of calculating time and memory space, and algorithm is caused to be unable to satisfy the need of system real-time update
It asks.Relative to traditional batch type SVDD, incremental learning technology can inherit the knowledge acquired, on the basis of original model
According to new data sample come more new model, this can not only make model knowledge have hereditability, additionally it is possible to cope with real process
Time-varying problem.Meanwhile history energy feature variable can not also be suitable for new process operating condition and new data, need to exist again
Energy feature variable is extracted in historical data and the set of new data.Usual newest measurand data can highlight active procedure
The characteristic of system, and the system performance that older historical data is reflected differs remoter with newest system mode.So in order to
It can reflect system current properties in time, need historical data to be added weight to be added to reduce the influence power of historical data
Forgetting factor.Furthermore, it is necessary to rationally adjust forgetting factor according to the situation of change of current system parameter.If can be from sample set
Effectively On-line testingInformation reuses fault detection algorithm after constructing energy feature sample set, it will is obviously improved early stage
The efficiency of fault detection.
Summary of the invention
The present invention is difficult to anti-aiming at the problem that system parameter time-varying characteristics in real process with the energy feature filtered out
Incremental Learning Algorithm and the power feature extraction based on variable forgetting factor has been respectively adopted in the problem of reflecting the dynamic characteristic of process
Method, in extracted in self-adaptive procedure faultWhile information, moreover it is possible to establish the incremental detection model of continuous renewal, Jin Erti
For a kind of incipient fault detection method.
One kind is based onInformation and the online incipient fault detection method of increment SVDD, comprising the following steps:
(1) collect process measurand withSample is imitated, and it is standardized, is obtainedAnd Y0;To sample
This progress data prediction is mainly pre-processed using data normalization, and specific implementation step is as follows:
For training sample data x1,x2,…xN, sample xiStandardization calculation formula are as follows:
Wherein, xi、Sample after respectively indicating i-th of original offline sample and standardization, θ are all sample datas
Arithmetic mean of instantaneous value, σ are the variance of all samples.By standardization, detection error caused by variable range can be eliminated;
(2) it calculates in step 1And Y0Between different measurands withThe association relationship of effectAccording to tired
Meter mutual information contribution rate principle filter out withImitate maximally related one group of energy feature sample set X0, establish energy feature sample
Detailed step are as follows:
In order to reduce influence of the sample size to Mutual Information Estimation, the mutual trust between two variables is calculated using k-nearest neighbor
Breath value:
Wherein, N is the number of total sample, and k is neighbour's number, sxAnd syIt is expressed as in the subspace of X and Y close to most
The number of samples of neighbour, φ () are digamma function.Simultaneously to ensure that mutual information can efficiently extract energy information, can make
The accumulation mutual information contribution degree of g measurand of selection are as follows:
G feature is included before CPMI is illustratedEffect information accounts for whole systemsThe ratio for imitating information, usually uses it
To determine the Characteristic Number filtered out.Since the mutual information of each feature is greater than 0, then CPMI is in the feature filtered out
Monotonic increase in number value range.In order to reach relatively good Feature Dimension Reduction effect, need so that CPMI is greater than defined control
System limit;
(3) the energy feature sample set X obtained by step 20, and construct initial SVDD model Γ0, establish the detailed of model
Step are as follows:
Give a training set [x1, x2..., xN], wherein N is sample number.A and R respectively indicates hyperspherical center and half
Diameter.By structural risk minimization it is found that this suprasphere radius minimization problem can be described as it is following with inequality
The quadratic programming problem of constraint, while introducing slack variable ξiAnd penalty factor:
For the quadratic programming problem of above-mentioned with constraint conditions, Lagrange multiplier can be introduced, by the quadratic programming problem
After being converted into its dual form, often it is more easier to solve, and kernel function can be introduced, luv space is projected into higher dimensional space
To solve nonlinear problem:
By asking for quadratic programming optimization algorithm SMO algorithm (Sequential Minimal Optimization, SMO)
Solution obtains optimal solution αi, to obtain the model parameter Γ of SVDD0;
(4) increment sample is introducedAnd Y1Afterwards, according to the power feature extraction rule based on variable forgetting factor come adaptive
Answer the more corresponding association relationship of new increment sampleThe energy feature of increment sample is filtered out according to Feature Dimension Reduction principle
SetThe step of adaptive energy feature extraction are as follows:
After the corresponding association relationship of newly-increased data sample all adds forgetting factor ρ, for the t+1 times newly-increased measurement sample
ThisWith corresponding bright effect data sample Yt+1, association relationship between the two are as follows:
Wherein ρ ∈ (0,1],Represent the column vector that j-th of measurand feature is constituted in t+1 newly-increased samples, m
Measured variable number is referred to,H(Yt+1)、Respectively indicate the comentropy and between the two of variable
Joint entropy.
The size of current forgetting factor can adaptively be adjusted according to the Parameters variation of time-varying system by becoming forgetting factor method, right
In the variable forgetting factor ρ of kth+1 time newly-increased sample association relationship are as follows:
CPMI in formulat+1Refer to the accumulative mutual information contribution rate of the t+1 times newly-increased data character pair sample, gt+1The choosing of value
Taking is determined according to the accumulative mutual information contribution rate principle of current sample, CPMIt+1Value is typically larger than 0.85.Finally, according to
The CPMI of current sample sett+1Filter out new energy feature set
(5) according to the initial SVDD model Γ of step 30With the energy feature sample of step 4)Update obtains increment
SVDD model Γ1, the specific steps of incremental update are as follows:
1) statistic and statistics limit;
Given sample set X=[x1,x2,…,xN], the dual form of SVDD suprasphere radius optimization problem are as follows:
Whereinδ is the Optimization Compensation factor, and kernel function can be denoted as Kij=K (xi,xj);By
KKT condition it is found that W to there is optimal solution, to αi, the first derivative of δ need to meet following condition:
Wherein Ω (xi) be SVDD sample discriminant function.
By formula (10) it is found that the KKT condition of SVDD is usually divided into three classes training sample set: non-supporting vector R, mark in ball
Quasi- supporting vector S and boundary supporting vector E.For the new samples x being just addedu, need to change model parameter to make the sample of extension
This collection reaches KKT condition again, so that new optimal solution is obtained, model coefficient variation are as follows:
Wherein αuIt is the new samples x being just addeduCorresponding Lagrange multiplier, Δ αuFor new samples Lagrange multiplier
Variable quantity, Δ αjOriginal sample set Lagrange multiplier variable quantity after being added for new samples.Standard supporting vector model coefficient
Variable quantity is writeable are as follows:
Enable T=K-1, for the sample in E, R set, κi≡ 0, from formula (13):
It enables
In formula,For model boundary sensitive factor, for the sample in S set,Carry it into formula (12)
In, then have:
2) sample attribute migrates;
Sample attribute migration is by increment sample pattern index variation amount Δ αuThe variation of original sample pattern coefficient is adjusted,
Extension sample set is set to meet KKT condition again, to tend to the process of a new equilibrium state.In increment sample attribute transition process
In, attribute is all satisfied Δ αu> 0, six kinds introduced below different sample attribute migration situations:
①κsAs the corresponding property value set of standard supporting vector collectionIn s-th of number, root
Δ α can be calculated according to formula (14)sIf it meets upper limit Δ αs≤C-αs, then Δ αuAnd κsJack per line, xsBecome by standard supporting vector
For boundary supporting vector:
In formula, λ is migration factor.
2. if Δ αsIn lower limit Δ αs≥-αs, then Δ αuAnd κsContrary sign, xsBecome non-supporting in ball from standard supporting vector
Vector:
3. for vector x non-supporting in ballr, then have dr> 0,R-th as vector sensitive factor collection non-supporting in ball
Δ d can be calculated according to formula (16) in elementr.If Δ dr< 0, non-supporting vector x in ballrIt can become standard supporting vector:
4. for boundary supporting vector xe, then have dr< 0, e-th of element of non-supporting vector sensitive factor collection can table in ball
It is shown asIt can then be calculatedIf Δ de> 0, boundary supporting vector xeIt can become standard supporting vector:
5. usually setting increment sample xuInitial model sample coefficient be zero, work as duSample x is then thought when >=0uFor ball
Interior target class sample, then its model sample coefficient is without updating.Work as du< 0, increment sample xuThen become standard supporting vector:
6. during increment SVDD model training, αuUpper dividing value is penalty factor, works as αuWhen≤C, increment sample αuIt will
Become boundary supporting vector, increment coefficient variable quantity are as follows:
λma=C- αu (22)
In increment SVDD, Δ α is enabledmax=min { λsp, λsm, λrs, λes, λc, λmaIt is used as increment sample αuModel coefficient
Variable quantity, formula (14) can obtain archetype sample coefficient variation delta αi;
3) T matrix update;
It also needs to update simultaneously during T matrix model increment iterative in formula (13), in order to reduce calculating finding the inverse matrix
Complexity, take following more new strategy.For an arbitrary sample xl∈R∪E∪{xuBecome standard supporting vector, t
The T of+1 renewal processt+1Are as follows:
As standard supporting vector xlWhen leaving set S, Tt+1Are as follows:
(6) increment sample is continually introduced, dynamically updates the increment SVDD model after t times out according to step 4), step 5)
Γt.And for different faults state sample, calculate the increment SVDD model Γ of different faults stateth;
(7) to on-line testing sample setData normalization and power feature extraction are carried out, it is calculated and corresponds to h failure shape
The relative distance γ of states modelh(xhi), it is based on relative distance minimum principle, detection and identification goes out the corresponding malfunction class of sample
Not, detailed step are as follows:
Diagnostic method based on relative distance can be by divided by the suprasphere radius in each failure state model, Lai Jiaqiang
The fault condition detection rate of decision rule raising algorithm.Relative distance not only can detect that failure, moreover it is possible to judge out of order tight
Weight degree.Use relative distance as criterion:
Compared with traditional technology, the medicine have the advantages that
The advantages of present invention incorporates based on ENERGY METHOD and based on data-driven method, system during not only can solve
The problem of parameter time varying characteristic, moreover it is possible to which the energy feature variable during efficiently extracting improves the rate of incipient fault detection
And precision, effectively realize the on-line checking of initial failure.
Detailed description of the invention
Fig. 1 is the flow chart of the method for the present invention;
Fig. 2 is the schematic diagram of industrial propylene rectification tower;
Fig. 3 is increment type and batch type model training time comparison diagram;
Fig. 4 a~Fig. 4 c is the existing online incipient fault detection of SVDD method as a result, wherein Fig. 4 a is sample in normal shape
Performance in states model, Fig. 4 b are performance of the sample on medium degenerate state model, and Fig. 4 c is sample in serious degenerate state
Performance on model;
Fig. 5 a~Fig. 5 c is the online incipient fault detection of increment type SVDD method as a result, wherein Fig. 5 a is sample in normal shape
Performance in states model, Fig. 5 b are performance of the sample on medium degenerate state model, and Fig. 5 c is sample in serious degenerate state
Performance on model;
Fig. 6 a~Fig. 6 c is the online incipient fault detection of the method for the present invention as a result, wherein Fig. 6 a is sample in normal condition mould
Performance in type, Fig. 6 b are performance of the sample on medium degenerate state model, and Fig. 6 c is sample in serious degenerate state model
On performance.
Specific embodiment
Come that the present invention is described in detail below with reference to factory's embodiment and attached drawing, this is fully understood with case study
The process of invention application implementation and the validity of method.
One kind is based onInformation and the online incipient fault detection method of increment SVDD, specific embodiment are as follows:
Using certain industrial propylene rectification tower as research object, the charging of tower mainly based on propane and propylene, the two it is opposite
Volatility, in order to reach preferable separating effect, is operated close to 1 using cascade towers.Certain industrial propylene rectification tower process is such as
Shown in Fig. 2.1# propylene rectification tower is equivalent to the stripping section of distillation process, which is equipped with 77 layers of column plate, in the condition of flow control
Lower its tower reactor discharging is used as recycled propane.The kettle material of 2# propylene rectification tower squeezes into 1# propylene rectification tower tower by reflux pump
Top, as being 1# propylene rectification tower overhead reflux.Propylene product is sent to propylene product guard bed after tower top cooler, thus
Remove some oxide impurities.The industry rectifying case is built upon on the ASPEN Math of certain factory's propylene rectification tower,
The propylene distillation process is simulated using 8.4 software of Aspen Tech.In order to simulate fouling of heat exchangers event in distillation process
Barrier, process are calculated using the double steam heat exchanger modules calculated in detail by the film coefficient and tube wall resistance of geometry calculation
Overall heat-transfer coefficient chooses the Fouling in Condenser factor and Spline smoothing occurs to simulate distillation process initial failure.If condenser
Fouling factor increases, and will lead to heat transfer efficiency reduction, and exchanger heat flux is reduced, and process is effectiveIt is wasted,Effect will
It reduces.The index as measurement process available energy utilization efficiency is imitated, can significantly reflect distillation process initial failure situation, institute
Whether to illustrate initial failure using the Spline smoothing of the Fouling in Condenser factor.In simulation process, by by the condensation of No. 2 towers
The fouling factor parameter of device changes to 8.6 × 10- from 03(sqmK)/Watt, to simulate the initial failure in distillation process.
(1) collect process measurand withSample is imitated, and it is standardized, is obtainedAnd Y0;To sample
This progress data prediction is mainly pre-processed using data normalization, and specific implementation step is as follows:
For training sample data x1,x2,…xN, sample xiStandardization calculation formula are as follows:
Wherein, xi、Sample after respectively indicating i-th of original offline sample and standardization, θ are all sample datas
Arithmetic mean of instantaneous value, σ are the variance of all samples.By standardization, detection error caused by variable range can be eliminated;
(2) it calculates in step 1And Y0Between different measurands withThe association relationship of effectAccording to tired
Meter mutual information contribution rate principle filter out withImitate maximally related one group of energy feature sample set X0, establish energy feature sample
Detailed step are as follows:
In order to reduce influence of the sample size to Mutual Information Estimation, the mutual trust between two variables is calculated using k-nearest neighbor
Breath value:
Wherein, N is the number of total sample, and k is neighbour's number, sxAnd syIt is expressed as in the subspace of X and Y close to most
The number of samples of neighbour, φ () are digamma function.Simultaneously to ensure that mutual information can efficiently extract energy information, can make
The accumulation mutual information contribution degree of g measurand of selection are as follows:
G feature is included before CPMI is illustratedEffect information accounts for whole systemsThe ratio for imitating information, usually uses it
To determine the Characteristic Number filtered out.Since the mutual information of each feature is greater than 0, then CPMI is in the feature filtered out
Monotonic increase in number value range.In order to reach relatively good Feature Dimension Reduction effect, need so that CPMI is greater than defined control
System limit, usually control limit take 85%;
(3) the energy feature sample set X obtained by step 20, and construct initial SVDD model Γ0, establish the detailed of model
Step are as follows:
Give a training set [x1, x2..., xN], wherein N is sample number.A and R respectively indicates hyperspherical center and half
Diameter.By structural risk minimization it is found that this suprasphere radius minimization problem can be described as it is following with inequality
The quadratic programming problem of constraint, while introducing slack variable ξiAnd penalty factor:
For the quadratic programming problem of above-mentioned with constraint conditions, Lagrange multiplier can be introduced, by the quadratic programming problem
After being converted into its dual form, often it is more easier to solve, and kernel function can be introduced, luv space is projected into higher dimensional space
To solve nonlinear problem:
By asking for quadratic programming optimization algorithm SMO algorithm (Sequential Minimal Optimization, SMO)
Solution obtains optimal solution αi, to obtain the model parameter Γ of SVDD0;
(4) increment sample is introducedAnd Y1Afterwards, according to the power feature extraction rule based on variable forgetting factor come adaptive
Answer the more corresponding association relationship of new increment sampleThe energy for filtering out increment sample according to Feature Dimension Reduction principle is special
Collection is closedThe step of adaptive energy feature extraction are as follows:
After the corresponding association relationship of newly-increased data sample all adds forgetting factor ρ, for the t+1 times newly-increased measurement sample
ThisWith corresponding bright effect data sample Yt+1, association relationship between the two are as follows:
Wherein ρ ∈ (0,1],Represent the column vector that j-th of measurand feature is constituted in t+1 newly-increased samples, m
Measured variable number is referred to,H(Yt+1)、Respectively indicate the comentropy and between the two of variable
Joint entropy.
The size of current forgetting factor can adaptively be adjusted according to the Parameters variation of time-varying system by becoming forgetting factor method, right
In the variable forgetting factor ρ of kth+1 time newly-increased sample association relationship are as follows:
CPMI in formulat+1Refer to the accumulative mutual information contribution rate of the t+1 times newly-increased data character pair sample, gt+1The choosing of value
Taking is determined according to the accumulative mutual information contribution rate principle of current sample, CPMIt+1Value is typically larger than 0.85.Finally, according to
The CPMI of current sample sett+1Filter out new energy feature set
(5) according to the initial SVDD model Γ of step 30With the energy feature sample of step 4)Update obtains increment
SVDD model Γ1, the specific steps of incremental update are as follows:
1) statistic and statistics limit;
Given sample set X=[x1, x2..., xN], the dual form of SVDD suprasphere radius optimization problem are as follows:
Whereinδ is the Optimization Compensation factor, and kernel function can be denoted as Kij=K (xi, xj);By
KKT condition it is found that W to there is optimal solution, to αi, the first derivative of δ need to meet following condition:
Wherein Ω (xi) be SVDD sample discriminant function.
By formula (10) it is found that the KKT condition of SVDD is usually divided into three classes training sample set: non-supporting vector R, mark in ball
Quasi- supporting vector S and boundary supporting vector E.For the new samples x being just addedu, need to change model parameter to make the sample of extension
This collection reaches KKT condition again, so that new optimal solution is obtained, model coefficient variation are as follows:
Wherein αuIt is the new samples x being just addeduCorresponding Lagrange multiplier, Δ αuFor new samples Lagrange multiplier
Variable quantity, Δ αjOriginal sample set Lagrange multiplier variable quantity after being added for new samples.Standard supporting vector model coefficient
Variable quantity is writeable are as follows:
Enable T=K-1, for the sample in E, R set, κi≡ 0, from formula (13):
It enables
In formula,For model boundary sensitive factor, for the sample in S set,It carries it into formula (12),
Then have:
2) sample attribute migrates;
Sample attribute migration is by increment sample pattern index variation amount Δ αuThe variation of original sample pattern coefficient is adjusted,
Extension sample set is set to meet KKT condition again, to tend to the process of a new equilibrium state.In increment sample attribute transition process
In, attribute is all satisfied Δ αu> 0, six kinds introduced below different sample attribute migration situations:
①κsAs the corresponding property value set of standard supporting vector collectionIn s-th of number, root
Δ α can be calculated according to formula (14)sIf it meets upper limit Δ αs≤C-αs, then Δ αuAnd κsJack per line, xsBecome by standard supporting vector
For boundary supporting vector:
In formula, λ is migration factor.
2. if Δ αsIn lower limit Δ αs≥-αs, then Δ αuAnd κsContrary sign, xsBecome non-supporting in ball from standard supporting vector
Vector:
3. for vector x non-supporting in ballr, then have dr> 0,R-th as vector sensitive factor collection non-supporting in ball
Δ d can be calculated according to formula (16) in elementr.If Δ dr< 0, non-supporting vector x in ballrIt can become standard supporting vector:
4. for boundary supporting vector xe, then have dr< 0, e-th of element of non-supporting vector sensitive factor collection can table in ball
It is shown asIt can then be calculatedIf Δ de> 0, boundary supporting vector xeIt can become standard supporting vector:
5. usually setting increment sample xuInitial model sample coefficient be zero, work as duSample x is then thought when >=0uFor ball
Interior target class sample, then its model sample coefficient is without updating.Work as du< 0, increment sample xuThen become standard supporting vector:
6. during increment SVDD model training, αuUpper dividing value is penalty factor, works as αuWhen≤C, increment sample αuIt will
Become boundary supporting vector, increment coefficient variable quantity are as follows:
λma=C- αu (22)
In increment SVDD, Δ α is enabledmax=min { λsp,λsm,λrs,λes,λc,λmaIt is used as increment sample αuModel coefficient
Variable quantity, formula (14) can obtain archetype sample coefficient variation delta αi;
3) T matrix update;
It also needs to update simultaneously during T matrix model increment iterative in formula (13), in order to reduce calculating finding the inverse matrix
Complexity, take following more new strategy.For an arbitrary sample xl∈R∪E∪{xuBecome standard supporting vector, t
The T of+1 renewal processt+1Are as follows:
As standard supporting vector xlWhen leaving set S, Tt+1Are as follows:
(6) increment sample is continually introduced, dynamically updates the increment SVDD model after t times out according to step 4), step 5)
Γt.And for different faults state sample, calculate the increment SVDD model Γ of different faults stateth;
(7) to on-line testing sample setData normalization and power feature extraction are carried out, it is calculated and corresponds to h failure shape
The relative distance γ of states modelh(xhi), it is based on relative distance minimum principle, detection and identification goes out the corresponding malfunction class of sample
Not, detailed step are as follows:
Diagnostic method based on relative distance can be by divided by the suprasphere radius in each failure state model, Lai Jiaqiang
The fault condition detection rate of decision rule raising algorithm.Relative distance not only can detect that failure, moreover it is possible to judge out of order tight
Weight degree.Use relative distance as criterion:
Increment SVDD and SVDD most essential difference are structure inheritance, so the training time of increment SVDD can be than passing
The training time of system SVDD is short, as shown in Figure 3.As can be seen from the results, as sample size is constantly increasing, SVDD model training
Time and the gap of increment SVDD are increasing.This is because tradition SVDD algorithm must be by last training sample and newly-increased
Sample put together re -training solve double optimization problem, cause historical sample repeatedly to be trained, while to the sample being just added
This is without rationally being screened;Unlike batch type SVDD, the inheritability of structure creates increment SVDD and can use
The advantages of last training result carrys out more new model, and increment SVDD selectively carries out study instruction to increment sample
Practice, it is possible to substantially reduce the training time.
The method of the present invention is compared with SVDD method, the online early detection rate of increment SVDD, testing result such as 1 institute of table
Show.SVDD method is compared with the testing result of increment SVDD, it is known that quantitation algorithm and non incrementalalgorithm are examined in failure
Difference is unobvious in survey rate.Increment type SVDD is smaller than the suprasphere space of non-increment type SVDD, and training data dimension and
Number of samples is more, it is possible thereby to illustrate that increment SVDD can selectively retain the sample that those most possibly become supporting vector
This, to shorten the training time while not reducing fault detection rate.That is, delta algorithm does not sacrifice classification
Ability carrys out the training speed of lift scheme, illustrates that delta algorithm is more adaptive to system dynamic change problem.Table 1 and Fig. 4, Fig. 5,
Result in Fig. 6 also shows the algorithm fault detection rate without power feature extraction than there is the event of the algorithm of power feature extraction
Barrier verification and measurement ratio wants low, it was demonstrated that is based onThe Feature Dimension Reduction technology of information extraction can effectively reject unrelated noise information and
Failure variation characteristic is extracted, increases the classification performance of model, to instruct actual production process to a certain extent.
Verification and measurement ratio of 1 distinct methods of table in initial failure state
Content described in this specification embodiment is only enumerating to the way of realization of inventive concept, protection of the invention
Range should not be construed as being limited to the specific forms stated in the embodiments, and protection scope of the present invention is also and in art technology
Personnel conceive according to the present invention it is conceivable that equivalent technologies mean.
Claims (1)
1. one kind is based onInformation and the online incipient fault detection method of increment SVDD, comprising the following steps:
1) collect process measurand withSample is imitated, and it is standardized, is obtainedAnd Y0;Sample is carried out
Data prediction is pre-processed using data normalization, and specific implementation step is as follows:
For training sample data x1,x2,…xN, sample xiStandardization calculation formula are as follows:
Wherein, xi、Sample after respectively indicating i-th of original offline sample and standardization, θ are the arithmetic of all sample datas
Average value, σ are the variance of all samples;
2) it calculates in step 1)And Y0Between different measurands withThe association relationship of effectAccording to accumulative mutual
Information contribution rate principle filter out withImitate maximally related one group of energy feature sample set X0, establish the detailed of energy feature sample
Step are as follows:
In order to reduce influence of the sample size to Mutual Information Estimation, the association relationship between two variables is calculated using k-nearest neighbor:
Wherein, N is the number of total sample, and k is neighbour's number, sxAnd syIt is expressed as in the subspace of X and Y close to arest neighbors
Number of samples, φ () be digamma function.Simultaneously to ensure that mutual information can efficiently extract energy information, can make to choose
G measurand accumulation mutual information contribution degree are as follows:
G feature is included before CPMI is illustratedEffect information accounts for whole systemsThe ratio for imitating information, is determined to sieve with it
The Characteristic Number selected;Since the mutual information of each feature is greater than 0, then CPMI is in the Characteristic Number value model filtered out
Enclose interior monotonic increase;In order to reach relatively good Feature Dimension Reduction effect, need so that CPMI is limited greater than defined control;
3) the energy feature sample set X obtained by step 2)0, and construct initial SVDD model Γ0, establish the detailed step of model
Are as follows:
Give a training set [x1,x2,…,xN], wherein N is sample number;A and R respectively indicates hyperspherical center and radius.
By structural risk minimization it is found that this suprasphere radius minimization problem can be described as it is following with inequality constraints
Quadratic programming problem, while introducing slack variable ξiAnd penalty factor:
For the quadratic programming problem of above-mentioned with constraint conditions, Lagrange multiplier can be introduced, which is converted
After its dual form, be often more easier to solve, and kernel function can be introduced, by luv space project to higher dimensional space to
Solve nonlinear problem:
By solving for quadratic programming optimization algorithm SMO algorithm (Sequential Minimal Optimization, SMO)
To optimal solution αi, to obtain the model parameter Γ of SVDD0;
4) increment sample is introducedAnd Y1Afterwards, according to the power feature extraction rule based on variable forgetting factor come adaptive updates
The corresponding association relationship of increment sampleThe energy feature set of increment sample is filtered out according to Feature Dimension Reduction principleThe step of adaptive energy feature extraction are as follows:
After the corresponding association relationship of newly-increased data sample all adds forgetting factor ρ, for the t+1 times newly-increased measurement sampleWith it is correspondingImitate data sample Yt+1, association relationship between the two are as follows:
Wherein ρ ∈ (0,1],The column vector that j-th of measurand feature is constituted in t+1 newly-increased samples is represented, m is referred to
Measured variable number,H(Yt+1)、Respectively indicate the comentropy and connection between the two of variable
Close entropy;
The size of current forgetting factor can adaptively be adjusted according to the Parameters variation of time-varying system by becoming forgetting factor method, for the
The variable forgetting factor ρ of k+1 newly-increased sample association relationship are as follows:
CPMI in formulat+1Refer to the accumulative mutual information contribution rate of the t+1 times newly-increased data character pair sample, gt+1The selection of value is
It is determined according to the accumulative mutual information contribution rate principle of current sample;Finally, according to the CPMI of current sample sett+1Screening
New energy feature set out
5) according to the initial SVDD model Γ of step 3)0With the energy feature sample of step 4)Update obtains increment SVDD mould
Type Γ1, the specific steps of incremental update are as follows:
5-1) statistic and statistics limit;
Given sample set X=[x1,x2,…,xN], the dual form of SVDD suprasphere radius optimization problem are as follows:
Whereinδ is the Optimization Compensation factor, and kernel function can be denoted as Kij=K (xi,xj);By KKT
Condition it is found that W to there is optimal solution, to αi, the first derivative of δ need to meet following condition:
Wherein Ω (xi) be SVDD sample discriminant function;
By formula (10) it is found that the KKT condition of SVDD is usually divided into three classes training sample set: non-supporting vector R, standard branch in ball
Hold vector S and boundary supporting vector E;For the new samples x being just addedu, need to change model parameter to make the sample set of extension
Reach KKT condition again, so that new optimal solution is obtained, model coefficient variation are as follows:
Wherein αuIt is the new samples x being just addeduCorresponding Lagrange multiplier, Δ αuFor the change of new samples Lagrange multiplier
Change amount, Δ αjOriginal sample set Lagrange multiplier variable quantity after being added for new samples.The variation of standard supporting vector model coefficient
It measures writeable are as follows:
Enable T=K-1, for the sample in E, R set, κi≡ 0, from formula (13):
It enablesIn formula,For model boundary
Sensitive factor, for S set in sample,It carries it into formula (12), then:
5-2) sample attribute migrates;
Sample attribute migration is by increment sample pattern index variation amount Δ αuThe variation for adjusting original sample pattern coefficient, makes to expand
Exhibition sample set meets KKT condition again, to tend to the process of a new equilibrium state;In increment sample attribute transition process,
Its attribute is all satisfied Δ αu> 0, sample attribute migration situation are divided into following six kinds:
①κsAs the corresponding property value set of standard supporting vector collectionIn s-th of number, according to formula
(14) Δ α can be calculatedsIf it meets upper limit Δ αs≤C-αs, then Δ αuAnd κsJack per line, xsSide is become from standard supporting vector
Boundary's supporting vector:
In formula, λ is migration factor;
2. if Δ αsIn lower limit Δ αs≥-αs, then Δ αuAnd κsContrary sign, xsFrom standard supporting vector become in ball it is non-supporting to
Amount:
3. for vector x non-supporting in ballr, then have dr> 0,R-th yuan as vector sensitive factor collection non-supporting in ball
Δ d can be calculated according to formula (16) in elementr;If Δ dr< 0, non-supporting vector x in ballrIt can become standard supporting vector:
4. for boundary supporting vector xe, then have dr< 0, e-th of element of non-supporting vector sensitive factor collection is represented by ballIt can then be calculatedIf Δ de> 0, boundary supporting vector xeIt can become standard supporting vector:
5. usually setting increment sample xuInitial model sample coefficient be zero, work as duSample x is then thought when >=0uFor mesh in ball
Class sample is marked, then its model sample coefficient is without updating;Work as du< 0, increment sample xuThen become standard supporting vector:
6. during increment SVDD model training, αuUpper dividing value is penalty factor, works as αuWhen≤C, increment sample αuIt will become
Boundary supporting vector, increment coefficient variable quantity are as follows:
λma=C- αu (22)
In increment SVDD, Δ α is enabledmax=min { λsp,λsm,λrs,λes,λc,λmaIt is used as increment sample αuModel coefficient variation
Amount, formula (14) can obtain archetype sample coefficient variation delta αi;
5-3) T matrix update;
It also needs to update simultaneously during T matrix model increment iterative in formula (13), calculates answering for finding the inverse matrix to reduce
Miscellaneous degree takes following more new strategy;For an arbitrary sample xl∈R∪E∪{xuBecome standard supporting vector, t+1 times
The T of renewal processt+1Are as follows:
As standard supporting vector xlWhen leaving set S, Tt+1Are as follows:
6) increment sample is continually introduced, t times out increment SVDD model Γ is dynamically updated according to step 4), step 5)t;And it is right
In different faults state sample, the increment SVDD model Γ of different faults state is calculatedth;
7) to on-line testing sample setData normalization and power feature extraction are carried out, it is calculated and corresponds to h failure state model
Relative distance γh(xhi), it is based on relative distance minimum principle, detection and identification goes out the corresponding malfunction classification of sample, in detail
Thin step are as follows:
Diagnostic method based on relative distance can be by divided by the suprasphere radius in each failure state model, to reinforce determining
Rule improves the fault condition detection rate of algorithm;Relative distance not only can detect that failure, moreover it is possible to judge out of order serious journey
Degree.Use relative distance as criterion:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910486042.3A CN110298385B (en) | 2019-06-05 | 2019-06-05 | exergy information and incremental SVDD (singular value decomposition) based online early fault detection method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910486042.3A CN110298385B (en) | 2019-06-05 | 2019-06-05 | exergy information and incremental SVDD (singular value decomposition) based online early fault detection method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110298385A true CN110298385A (en) | 2019-10-01 |
CN110298385B CN110298385B (en) | 2021-07-27 |
Family
ID=68027725
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910486042.3A Active CN110298385B (en) | 2019-06-05 | 2019-06-05 | exergy information and incremental SVDD (singular value decomposition) based online early fault detection method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110298385B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112184037A (en) * | 2020-09-30 | 2021-01-05 | 华中科技大学 | Multi-modal process fault detection method based on weighted SVDD |
CN112733872A (en) * | 2020-08-26 | 2021-04-30 | 南京航空航天大学 | Aeroengine fault detection method based on dynamic radius support vector data description |
CN113237920A (en) * | 2021-05-17 | 2021-08-10 | 西南交通大学 | Method for detecting fault heat source of valve-side sleeve of extra-high voltage converter transformer |
CN112733872B (en) * | 2020-08-26 | 2024-05-03 | 南京航空航天大学 | Aeroengine fault detection method based on dynamic radius support vector data description |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105868783A (en) * | 2016-03-31 | 2016-08-17 | 华东理工大学 | Reduction type support vector data description method based on information entropy |
CN106124988A (en) * | 2016-06-28 | 2016-11-16 | 江苏科技大学 | A kind of motor multi-state fault detection method based on RBF, multilamellar FDA and SVDD |
US20190042977A1 (en) * | 2017-08-07 | 2019-02-07 | Sas Institute Inc. | Bandwidth selection in support vector data description for outlier identification |
US20190095400A1 (en) * | 2017-09-28 | 2019-03-28 | Sas Institute Inc. | Analytic system to incrementally update a support vector data description for outlier identification |
-
2019
- 2019-06-05 CN CN201910486042.3A patent/CN110298385B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105868783A (en) * | 2016-03-31 | 2016-08-17 | 华东理工大学 | Reduction type support vector data description method based on information entropy |
CN106124988A (en) * | 2016-06-28 | 2016-11-16 | 江苏科技大学 | A kind of motor multi-state fault detection method based on RBF, multilamellar FDA and SVDD |
US20190042977A1 (en) * | 2017-08-07 | 2019-02-07 | Sas Institute Inc. | Bandwidth selection in support vector data description for outlier identification |
US20190095400A1 (en) * | 2017-09-28 | 2019-03-28 | Sas Institute Inc. | Analytic system to incrementally update a support vector data description for outlier identification |
Non-Patent Citations (1)
Title |
---|
MADAKYARU M 等: "Improved data-based fault detection strategy and application to distillation columns", 《PROCESS SAFETY AND ENVIRONMENTAL PROTECTION》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112733872A (en) * | 2020-08-26 | 2021-04-30 | 南京航空航天大学 | Aeroengine fault detection method based on dynamic radius support vector data description |
CN112733872B (en) * | 2020-08-26 | 2024-05-03 | 南京航空航天大学 | Aeroengine fault detection method based on dynamic radius support vector data description |
CN112184037A (en) * | 2020-09-30 | 2021-01-05 | 华中科技大学 | Multi-modal process fault detection method based on weighted SVDD |
CN112184037B (en) * | 2020-09-30 | 2022-11-11 | 华中科技大学 | Multi-modal process fault detection method based on weighted SVDD |
CN113237920A (en) * | 2021-05-17 | 2021-08-10 | 西南交通大学 | Method for detecting fault heat source of valve-side sleeve of extra-high voltage converter transformer |
CN113237920B (en) * | 2021-05-17 | 2022-04-22 | 西南交通大学 | Method for detecting fault heat source of valve-side sleeve of extra-high voltage converter transformer |
Also Published As
Publication number | Publication date |
---|---|
CN110298385B (en) | 2021-07-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109840362B (en) | Multi-objective optimization-based integrated just-in-time learning industrial process soft measurement modeling method | |
CN109213127A (en) | A kind of HVAC system gradual failure diagnostic method based on deep learning | |
CN103440368A (en) | Multi-model dynamic soft measuring modeling method | |
CN108376264A (en) | A kind of handpiece Water Chilling Units method for diagnosing faults based on support vector machines incremental learning | |
CN104657586B (en) | Technology Modeling optimization method is purified based on the high sulfur-containing natural gas that unusual service condition is detected | |
CN109144028A (en) | A kind of rectifying column efficiency deterioration detecting | |
CN110298385A (en) | One kind is based on * information and the online incipient fault detection method of increment SVDD | |
CN110986407A (en) | Fault diagnosis method for centrifugal water chilling unit | |
CN108919755A (en) | A kind of distributed fault detection method based on muti-piece Nonlinear and crossing relational model | |
CN113325721A (en) | Model-free adaptive control method and system for industrial system | |
Liu et al. | Model fusion and multiscale feature learning for fault diagnosis of industrial processes | |
CN106842948A (en) | The method for optimally controlling of the HDP molecular distillation systems based on BP networks | |
Dong et al. | Quality monitoring and root cause diagnosis for industrial processes based on Lasso-SAE-CCA | |
CN113049259B (en) | Fuzzy control method of rack control system, storage medium and equipment | |
CN108596364B (en) | Dynamic early warning method for major hazard source in chemical industry park | |
CN108204997A (en) | Normal line oil flash point online soft sensor method | |
CN109086887A (en) | Method for early warning of the depth RBF neural in conjunction with the AHP based on entropy weight | |
CN102289718A (en) | Method for softly measuring height of mud layer in settlement process of red mud | |
CN116662925A (en) | Industrial process soft measurement method based on weighted sparse neural network | |
CN116842358A (en) | Soft measurement modeling method based on multi-scale convolution and self-adaptive feature fusion | |
CN116430726A (en) | Method for controlling turboset based on subtractive clustering and fuzzy neural network | |
CN115936061A (en) | Thermal power plant flue gas oxygen content soft measurement method and system based on data driving | |
Hu et al. | Research on the fault identification method of oil pumping unit based on residual network | |
CN112182854B (en) | Intelligent monitoring method and system for abnormal furnace conditions of blast furnace | |
He et al. | Temperature intelligent prediction model of coke oven flue based on CBR and RBFNN |
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 |