Invention content
In order to overcome the shortcomings of that existing method, the present invention provide a kind of based on amendment type independent component analysis and Bayesian probability
The industrial process fault detection method of fusion.Technical solution is used by the present invention solves above-mentioned technical problem:One kind is based on
The industrial process fault detection method of amendment type independent component analysis and Bayesian probability fusion, includes the following steps:
(1) sampled data under production process normal operating condition is collected, the training dataset of modeling is formed:X=
[x1, x1..., xn]T, wherein X ∈ Rn×m, n is number of training, and m is process measurement variable number, and R is set of real numbers, Rn×mIndicate n
The real number matrix of × m dimensions, upper label T representing matrix transposition.
(2) data set X is standardized, it is 0 to make the mean value of each measurand, and variance 1 obtains new number
According to matrix
(3) utilize PCA methods to data matrixIt carries out whitening processing and obtains data matrix Z ∈ Rn×M.Wherein, M≤m is
Variable number in matrix Z.
(4) it is directed to different non-quadratic functions, calls MICA iterative algorithms to establish different MICA Fault Models, always
Meter establishes three MICA models;
(5) corresponding T is constructed to training data2And Q statistical magnitude, and respectively obtain T using with density estimation (KDE) method2With
Q statistical magnitude control limit corresponding under the conditions of confidence alpha=99%;
(6) new process sampled data x is collectedmew∈Rm×1, and be standardized to obtain
(7) different MICA models are respectively adopted, fault detect is carried out to it, that is, establish statistic T2And Q, such 3
MICA models total provides 3 groups and corresponds to new samples dataMonitoring result;
(8) by Bayesian inference, by obtain 3 groups of monitoring results be merged in the form of probability to get to point
It Dui Yingyu not T2With the probabilistic type monitoring index of Q statistical magnitudeWith BIQ, and make about new dataWhether normally determine
Plan.
Compared with the conventional method, the advantage of the invention is that:The present invention is established accordingly respectively using each non-quadratic function
MICA models.Then, it introduces Bayesian inference method and integrated and probability fusion is carried out to the monitoring result under different models, with
Obtain last comprehensive monitoring index.Compared to existing method, the present invention can solve to cause because of non-quadratic function diversity
Model uncertain problem, increase the reliability and stability of Fault Model.Therefore, the present invention can be largely
Upper reduction MICA methods enhance the reliability of Fault Model, to be more conducive to industry to the dependence of priori
The implementation of automation.
Specific implementation method
The present invention will be described in detail in the following with reference to the drawings and specific embodiments.
As shown in Figure 1, the invention discloses a kind of industry merged based on amendment type independent component analysis and Bayesian probability
Procedure failure testing method.For industrial process fault detection problem, production process is being collected just first with data collecting system
Data set under normal operating status, and it is standardized.Secondly, corresponding MICA is established for different non-quadratic functions
Fault Model, and model parameter is preserved with spare.Then, three groups of different monitorings are calculated to new sampled data to unite
Measure T2And Q, and different monitoring results are integrated into probability level using Bayesian probability fusion method.Finally, according to probability
Monitoring indexWith BIQWhether occurrence, which transfinites, is made the whether normal decision of new data.Specific implementation step of the present invention is such as
Under:
The first step:The sampled data under production process normal operating condition is collected, the training dataset of modeling is formed:X
=[x1, x1..., xn]T。
Second step:Data set X is standardized, it is 0 to make the mean value of each measurand, variance 1, is obtained new
Data matrix
Third walks:Using PCA methods to data matrixIt carries out whitening processing and obtains data matrix Z ∈ Rn×M, concrete operations
Shown in steps are as follows:
(1) it calculatesCovariance matrixWherein S ∈ Rm×m;
(2) all characteristic values and feature vector of calculating matrix S, and reject less than 0.0001 characteristic value and its corresponding
Feature vector obtains eigenvectors matrix P=[p1, p2..., pM]∈Rm×MAnd characteristic value diagonal matrix D=diag (λ1,
λ2..., λM)∈RM×M;
(3) rightWhitening processing is carried out, is obtained
4th step:K-th of non-quadratic function G is selected from three kinds of non-quadratic functionskAfterwards, MICA iterative algorithms are called to seek
With GkCorresponding MICA Fault Models repeat this operation until establishing three MICA models, wherein k=1,2,3 difference
For the label of three kinds of different non-quadratic functions.
The step of calling MICA iterative algorithms to establish Fault Model is as follows:
(1) setting needs the independent entry number d extracted and selected non-quadratic function Gk, k=1,2,3, wherein non-
Quadratic function has following three kinds of optional forms:
G1(u)=log cosh (u), G2(u)=exp (- u2/ 2), G3(u)=u4 (1)
Wherein, u is function GkIndependent variable.
(2) when extracting i-th (i=1,2 ..., d) a independent entry, choose M × M dimension unit matrix in i-th row as to
Measure ciInitial value;
(3) renewal vector c according to the following formulai:
ci←E{Zg(ci TZ)}-E{g′(ci TZ)}ci (2)
Wherein, g and g ' is the single order and second dervative of function G respectively, and desired value is sought in E { } expressions;
(4) to updated vector ciCarry out orthogonal standardization according to the following formula successively:
ci←ci/||ci|| (4)
(5) step (3)~(4) are repeated until vector ciConvergence, and preserve vectorial ci;
(6) i=i+1 is set, and (2)~(5) that repeat the above steps are until obtain all d vector C=[c1, c2..., cd]
∈RM×d。
(7) it calculates and corresponds to non-quadratic function GkMICA models hybrid matrix Ak∈Rm×dWith separation matrix Wk∈Rd ×m, as follows:
Ak=PD1/2C (5)
Wk=CTD-1/2PT (6)
5th step:Corresponding T is constructed to training data2And Q statistical magnitude, and obtained respectively using with density estimation (KDE) method
To T2Control corresponding under the conditions of confidence alpha=99% is 99% limit with Q statistical magnitudeAnd Qlim;
6th step:Collect new process sampled data xnew∈Rm×1, and be standardized to obtain
7th step:Different MICA models are respectively adopted, fault detect is carried out to it, i.e., establishes statistics respectively according to the following formula
AmountAnd Qk, such 3 MICA models, which amount to, provides 3 groups corresponding to new samples dataMonitoring result;
Wherein, | | | | it indicates to calculate the 2- norms of vector.
8th step:By Bayesian inference, by obtain 3 groups of monitoring results be merged in the form of probability to get
To corresponding respectively to T2With the probabilistic type monitoring index of Q statistical magnitudeWith BIQ, and make about new dataIt is whether normal
Decision.Specific implementation detail is further described as follows:
First, Bayesian probability fusion is carried out:
(A) to three groups of T2Statistic is merged:
(1) new data is calculated according to the following formulaBelong to the probability of failure:
Wherein, probabilityCalculation it is as follows:
Wherein, N and F indicates normal and fault condition, prior probability respectivelyWithValue α and 1- α respectively,
Conditional probabilityWithCalculation it is as follows:
(2) final probabilistic type index is calculated by following formula
(B) three groups of Q statistical magnitudes are merged
It is directed to Q statistical magnitude, calculates new data firstThen probability, the conditional probability etc. for belonging to failure obtain final
Probabilistic type index BIQ, T is merged with above-mentioned2The mode of statistic is identical.
Secondly, failure decision is carried out:
By what is be calculatedWith BIQThe concrete numerical value of index controls limit 1- α with probability and is compared.If any one
Index value is more than 1- α, then decision new dataFor fault sample;Conversely, the dataFor normal sample, and then to next
A new obtained data that sample continue fault detect.
With reference to the example of a specific industrial process come illustrate the present invention relative to existing method superiority with
Reliability.The process data comes from the experiment of the U.S. Tennessee-Yi Siman (TE) chemical process, and prototype is Yi Siman Chemical Manufactures
One actual process flow in workshop.Currently, complexity of the TE processes because of its flow, has been used as a standard test platform wide
It is general to be studied for fault detect.Entire TE processes include that 22 measurands, 12 performance variables and 19 composition measurements become
Amount.The data acquired are divided into 22 groups, including the data set and 21 groups of fault datas under 1 group of nominal situation.And at these
In fault data, 16 are known fault types, such as the changing of cooling water inlet temperature or feed constituents, valve viscous, anti-
Dynamics drift etc. is answered, also 5 fault types are unknown.In order to be monitored to the process, as shown in Table 1 33 are chosen
Next a process variable is explained in detail specific implementation step of the present invention in conjunction with the TE processes.
1. acquiring the process data under nominal situation, while fault data different in 21 is acquired, and chooses 960 normally
Data form matrix X ∈ R960×33, it is standardized.
2. using PCA methods to the data after standardizationIt carries out whitening processing and obtains Z ∈ R960×31。
3. the independent entry number d=9 retained is arranged, confidence alpha=99% selects different non-quadratic functions to carry out MICA moulds
Type is established:
1) non-quadratic function G is selected1MICA modelings are carried out, corresponding hybrid matrix A is obtained1∈R33×9With separation matrix W1
∈R9×33.Then T is constructed2And Q statistical magnitude, it is used in combination KDE methods to determine its corresponding control limitAnd Q1, lim;
2) non-quadratic function G is selected2MICA modelings are carried out, corresponding hybrid matrix A is obtained2∈R33×9With separation matrix W2
∈R9×33.Then T is constructed2And Q statistical magnitude, it is used in combination KDE methods to determine its corresponding control limitAnd Q2, lim;
3) non-quadratic function G is selected3MICA modelings are carried out, corresponding hybrid matrix A is obtained3∈R33×9With separation matrix W3
∈R9×33.Then T is constructed2And Q statistical magnitude, it is used in combination KDE methods to determine its corresponding control limitAnd Q3, lim。
4. obtaining current new sampled data, and it is standardized to obtain
In order to verify superiority of the present invention relative to existing method, chooses 21 kinds of failures and tested, equally to sampling
Data are standardized.
5. implementing on-line monitoring process
1) new data is monitored using three kinds of different MICA Fault Models, constructs the statistics under corresponding model
AmountAnd Qk;
2) probabilistic type monitoring index is obtained using Bayesian probability fusion calculationWith BIQ, and by occurrence and probability control
System limit 1- α are compared, and whether decision current data is normal.
Table 1:TE process monitoring variables.
Serial number |
Variable description |
Serial number |
Variable description |
Serial number |
Variable description |
1 |
Material A flow |
12 |
Separator liquid level |
23 |
D material inlet valves position |
2 |
Material D flows |
13 |
Separator pressure |
24 |
E material inlet valves position |
3 |
Material E flows |
14 |
Separator bottom of tower flow |
25 |
A material inlet valves position |
4 |
Combined feed flow |
15 |
Stripper grade |
26 |
A and C material inlet valves position |
5 |
Circular flow |
16 |
Pressure of stripping tower |
27 |
Compressor cycle valve location |
6 |
Reactor feed |
17 |
Stripper bottom rate |
28 |
Empty valve location |
7 |
Reactor pressure |
18 |
Stripper temperature |
29 |
Separator liquid phase valve location |
8 |
Reactor grade |
19 |
Stripper upper steam |
30 |
Stripper liquid phase valve location |
9 |
Temperature of reactor |
20 |
Compressor horsepower |
31 |
Stripper steam valve position |
10 |
Rate of evacuation |
21 |
Reactor cooling water outlet temperature |
32 |
Reactor condensate flow |
11 |
Separator temperature |
22 |
Separator cooling water outlet temperature |
33 |
Condenser cooling water flow |
It chooses 18 typical faults (removing failure 3, failure 9 and failure 15) to be monitored, by mean failure rate verification and measurement ratio pair
Than the results are shown in Table 2.As a result it shows that the method for the present invention has been obviously improved fault detect rate, improves fault detect effect.
Table 2:Mean failure rate verification and measurement ratio of the distinct methods to fault type in TE processes 18.
Above-mentioned case study on implementation only is used for illustrating the specific implementation of the present invention, rather than limits the invention.
In the protection domain of spirit and claims of the present invention, to any modification that the present invention makes, the protection of the present invention is both fallen within
Range.