CN101916394A - Knowledge fusion-based online soft measurement method - Google Patents
Knowledge fusion-based online soft measurement method Download PDFInfo
- Publication number
- CN101916394A CN101916394A CN2010102631005A CN201010263100A CN101916394A CN 101916394 A CN101916394 A CN 101916394A CN 2010102631005 A CN2010102631005 A CN 2010102631005A CN 201010263100 A CN201010263100 A CN 201010263100A CN 101916394 A CN101916394 A CN 101916394A
- Authority
- CN
- China
- Prior art keywords
- model
- module
- soft
- prediction
- 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.)
- Pending
Links
Images
Abstract
The invention relates to a knowledge fusion-based online soft measurement method, which comprises the following steps of: setting a measuring system; 2) selecting an empirical formula as a priori knowledge expression; 3) inputting the priori knowledge expression into a priori knowledge computing module for processing, and inputting the processed priori knowledge expression serving as an initial prediction model into a soft measurement model prediction module; 4) at the initial moment, directly predicting difficultly-measured key variables by the soft measurement model prediction module through the initial prediction model, and inputting a prediction result into a controller; 5) predicting difficultly-measured key variable values by the soft measurement model prediction module according to the acquired measured values of easily-measured variables, and inputting the difficultly-measured key variable values into the controller and a judgment module; 6) judging the accuracy of the predicted values by the judgment module according to the actual difficultly-measured variable values, and if a prediction error does not exceed a non-sensitive error zone, returning to step 5), otherwise, continuing the next step; and 7) maintaining and updating the conventional prediction model by an online learning model maintenance module by utilizing a new sample obtained by the last step, replacing a soft measurement model in the soft measurement model prediction module, and returning to the step 5).
Description
Technical field
The present invention relates to a kind of On-line Measuring Method, particularly about a kind of online soft sensor method based on knowledge fusion.
Background technology
Because the limitation of measuring technique, some significant variables in the Process Control System can't or be difficult to carry out on-line measurement.These variablees are as the key index of system state, the often off-line sample analysis of chamber by experiment, or use expensive in-line analyzer to obtain.The obvious delay (often being several hours) that the chamber measurement that experimentizes is produced, to cause these measuring-signals can not be as the feedback signal of control system, join in the whole control process, this all can produce serious influence to the performance of control system and the security of production run.
In recent years, researchers begin to utilize mass data measured in the production run to set up forecast model, to realize unknown difficult prediction of surveying variable.These models are called as soft sensor, and the notion of soft measurement also is suggested thus.Its basic thought is relevant with main difficult survey variable by measuring, and the secondary variable that is easy to measure (easily surveying variable) is estimated main difficult output of surveying variable.The basic step of soft sensor modeling mainly comprises starting stage, modelling phase and application stage.Wherein, in the application stage; because in fact process device be under the dynamic system environments, thereby regular meeting in operational process of system produces and changes slowly or be subjected to external interference to produce unexpected disturbance, the maintenance that causes the prediction effect of original model to obtain continuing.By way of example, wear and tear in machines, external causes such as surveying instrument deviation and environmental change all may cause the variation of system performance, promptly produce the data drift phenomenon, cause the predicated error of soft-sensing model constantly to increase.
At above-mentioned situation, we need carry out real-time maintenance to forecast model in the application stage of soft measurement.And the machine learning method that is widely used in modeling process can't be directly used in the model maintenance process, these machine learning algorithms generally can't carry out renewal initiatively, thereby can only be artificial regularly read new data, re-use all data that collected and carry out model maintenance.Obviously, this method wastes time and energy.Thereby, a kind of more efficiently model maintenance method is to introduce on-line learning algorithm, constantly the data of the up-to-date generation of reference are upgraded model and are adjusted in process of production, make model and real system be consistent to a certain extent, to solve the problem that often occurs in this reality production.
Can be used in the commonly used online learning method of handling soft problems of measurement and comprise the delta algorithm based on the SVM block algorithm, the on-line Algorithm that the empiric risk gradient descends, on-line Algorithm and the increment and the decrement algorithm of support vector machine etc. of least square method supporting vector machine.Yet these existing methods are easy to be absorbed among a kind of dilemma mostly, and promptly or be to keep all support vectors, but simultaneously along with the increase of data volume, the extensive performance and the counting yield of algorithm constantly reduce; Or in the model maintenance process, constantly abandon sample, with reduction model complexity, but also cause the consistance of algorithm and realistic model to lose simultaneously.This dilemma is the unavoidable bottleneck problems of present most on-line study methods, has also caused in long-term forecasting and maintenance process, and the precision of prediction of most of on-line study methods is relatively poor, and perhaps the training time is long.
Summary of the invention
At the problems referred to above, the purpose of this invention is to provide and a kind ofly can merge existing priori, not exclusively depend on sample, the reliability height, and can carry out the online soft sensor method of model real-time update according to actual sample information.
For achieving the above object, the present invention takes following technical scheme: a kind of online soft sensor method based on knowledge fusion, it may further comprise the steps: 1) measuring system that comprises data acquisition equipment, supervisory control comuter and controller is set, a default priori computing module, a soft-sensing model prediction module, a determination module and the on-line study model maintenance module of being equipped with in the described supervisory control comuter; 2) p (x) among the experimental formula y=p (x) is expressed as a kind of linear combination form q (x) of kernel function:
Wherein, K (s
i, x) for the kernel function of appointment; s
i∈ S
d, (i=1 .. d) are base vector, S
dBe the basal orientation quantity set; β
iIt is a undetermined parameter; 3) with formula (1) input priori computing module, the parameter s in the iterative computation formula (1)
i, d, β
i, flow to the soft-sensing model prediction module then; 4) in the soft-sensing model prediction module, establish: initial time t=0, the relevant variable measurements of easily surveying that whenever collects a new samples then makes: t=t+1; Initial time, the q as a result (x) that directly uses the step 3) input is as initial predicted model f
0(x)
Initial predicted model f
0(x) the relevant variable measurements of easily surveying that collects according to data acquisition equipment is predicted the unpredictable key variable value, and will predict the outcome and flow to controller; Wherein,
s
i∈ S
d, (i=1 .., d); 5) data acquisition equipment the relevant variable measurements x that easily surveys of new samples that t is collected constantly
tInput soft-sensing model prediction module, the soft-sensing model prediction module is according to current forecast model f
T-1(x) carry out the on-line prediction of new samples functional value, and the soft measurement that will the obtain f that predicts the outcome
T-1(xt) flow to controller and determination module; 6) determination module is preserved the anticipation function value f that t obtains constantly
T-1(xt), collect the t unpredictable key variable measured value y of sample constantly up to data acquisition equipment
t, and with y
tThe input determination module; Determination module is surveyed variate-value y according to the difficulty of reality
t, judge anticipation function value f
T-1(x
t) accuracy; If the absolute value of predicated error does not exceed insensitive error band ε, promptly | e
t|=| y
t-f
T-1(x
t) |≤ε, think that then the prediction effect of current model is better, need not carry out model modification, make f
t(x)=f
T-1(x), return step 5); Otherwise, | e
t|=| y
t-f
T-1(x
t) |>ε, then with sample (x
t, y
t) be transported to on-line study model maintenance module, proceed down-go on foot; 7) on-line study model maintenance module is utilized the new samples (x that obtains in the step 6)
t, y
t), to existing forecast model f
T-1(x) carry out maintenance update, and the forecast model f after will upgrading
t(x):
Be transported in the soft-sensing model prediction module, replace original soft-sensing model, return step 5).
In the described step 7), the forecast model after the renewal, promptly in the formula (3),
Be undetermined parameter, ask for
Concrete steps as follows: 1. at the data stream (x of sequential arrival
i, y
i)
I ∈ N, specify insensitive error band ε, learning rate η, coefficient of balance λ, the regularization coefficient gamma of control function complexity; 2. remember initialization prediction hypothesis constantly
Wherein
s
i∈ S
d, the output result who promptly directly adopts the priori computing module is as the initial predicted model; Calculate the d * d dimension kernel function matrix K of base vector
d, K
d(i, j) individual element is K (s
i, s
j); 3. read new samples (x
t, y
t), according to current forecast model expression formula f
T-1(x) calculate the anticipation function value of new samples:
And prediction error value e
t=y
t-f
T-1(x
t); 4. construct d dimensional vector K
t=[k (s
1, x
t) ..., k (s
d, x
t)]
T, the solving-optimizing problem
Wherein,
Be base vector s
iThe unit increment of pairing kernel function coefficient; The purpose of finding the solution of this minimum problems is: ask for when base vector the linear combination of corresponding kernel function and up-to-date sample point place kernel function distance the most near the time, promptly the 2-norm of the two is apart from approximate value δ
(t)Hour, base vector the coefficient increment θ of corresponding kernel function
(t)Calculating can get base vector the incremental vector θ of unit of corresponding kernel function coefficient
(t)=(K
d+ γ I
d)
-1K
t, wherein, I
dBe d dimension unit matrix; 5. upgrade base vector the coefficient of corresponding kernel function, order
I=1 ..., d; 6. obtaining t anticipation function constantly is
The model modification result is outputed to the soft-sensing model prediction module, and calculation process forwards the soft-sensing model prediction module to, continues to wait for the arrival of new samples.
In the described step 3), in the priori computing module, undetermined parameter d and β
i, and base vector s
iConcrete solution procedure as follows: 1. use the mode of experience directly perceived or circulation optimizing, the situation of formula p (x) rule of thumb, definite kernel function K (s
i, type x) and parameter are specified regularization coefficient μ, convergence threshold ε
f2. make iteration step number k=0, note expression function q (x) at this moment is q
0(x)=0, then error function is: r
0(x)=p (x)-q
0(x); 3. obtain the base vector s of k+1 iteration
K+1, make error function
If | r
k(s
K+1) |<ε
f, d=k then, q (x)=q
k(x), base vector s
iValue set be S
k,
I=1 ..., k, termination of iterations; 4. make S
K+1=S
k∪ { s
K+1, then construct one (k+1) * (k+1) dimension matrix K
K+1, K
K+1(i, j) individual element is K (s
i, s
j), and then construct (k+1) dimensional vector p (S
(k+1))=[p (s
1) ..., p (s
K+1)]
T5. compute vector matrix β
(k+1)=(K
K+1+ μ I
K+1)
-1P (S
(k+1)), I wherein
K+1Be (k+1) dimension unit matrix, order
Wherein
Be vector matrix β
(k+1)I element; 6. make r
K+1(x)=p (x)-q
K+1(x), 3. k=k+1 returns step.
The present invention is owing to take above technical scheme, it has the following advantages: 1, the present invention is owing to be dissolved into existing experimental formula in the industry in the soft-sensing model prediction module as priori, make the present invention's training sample that do not place one's entire reliance upon, therefore, compare with general on-line study method and to have higher precision of prediction, be a kind of applied widely, the soft-sensing model maintaining method that reliability is high.2, the present invention is owing to be provided with determination module, determination module the difficulty of reality can be surveyed variate-value and the anticipation function value compares, to judge the accuracy of forecast model, if predicated error has exceeded insensitive error band, then new samples is transported to on-line study model maintenance module, bring in constant renewal in forecast model by on-line study model maintenance module, to obtain the more accurate prediction effect, therefore, the present invention can carry out the active renewal and safeguard to have very strong adaptivity forecast model.3, on-line study model maintenance module of the present invention, constantly new forecast model is fed back in the soft-sensing model prediction module, the relevant variable measurements of easily surveying that the soft-sensing model prediction module can collect in real time according to data acquisition equipment, the crucial difficult variable measurements of surveying of on-line prediction, and difficulty is surveyed variable flow to controller, make its feedback signal, join in the whole control process, improve the performance of control system and the security of production run as control system.4, the present invention has made full use of the architectural feature that is comprised in the priori, can make forecast model in long-term model maintenance process, remain stronger generalization ability, guarantee that also precision of prediction of the present invention and training speed all are higher than general on-line study method.The present invention is skillfully constructed, and is accurate and practical, can be widely used in the actual measurement process.
Description of drawings
Fig. 1 is a structural representation of the present invention
Fig. 2 is a modular structure synoptic diagram of the present invention
Fig. 3 is the emulated data synoptic diagram of emulation example among the present invention
Fig. 4 is the measurement result synoptic diagram that the present invention is used for the emulation example
Embodiment
Below in conjunction with drawings and Examples the present invention is described in detail.
The present invention is based on following thought: by existing experimental formula in the industry spot is incorporated in the forecast model as priori, it is perfect to utilize the on-line study method that forecast model is carried out adaptive renewal on the one hand, on the other hand, the core information that utilizes priori and comprised guarantees that forecast model keeps stronger generalization ability and precision of prediction in long-term model maintenance process.The present invention has realized according to real time data forecast model being carried out the function of adaptive updates when unpredictable key variable being carried out the online soft sensor prediction.
The present invention includes following steps:
1) as shown in Figure 1 and Figure 2, according to the industry spot actual conditions, one measuring system that comprises data acquisition equipment 1, supervisory control comuter 2 and controller 3 is set, a default priori computing module 21, a soft-sensing model prediction module 22, a determination module 23 and the on-line study model maintenance modules 24 of being equipped with in the supervisory control comuter 2.
2) according to concrete applied environment, select deterministic experimental formula y=p (x), experimental formula p (x) is expressed as a kind of linear combination form q (x) of kernel function, its concrete form is:
Wherein, K (s
i, x) for the kernel function of appointment; s
i∈ S
d, (i=1 .. d) are base vector (also can be called support vector), S
dBe the basal orientation quantity set; β
iIt is a undetermined parameter.
3) with formula (1) as the priori expression formula, be input in the priori computing module 21, by priori computing module 21, the parameter s in the iterative computation formula (1)
i, d, β
iThereby, determine the formula that embodies of formula (1), and with formula (1) as priori to be merged, initial predicted model, flow to soft-sensing model prediction module 22.
4) suppose initial time t=0, data acquisition equipment 1 whenever collects the relevant variable measurements of easily surveying of a new samples, then makes t=t+1.In soft-sensing model prediction module 22, initial time, the output of directly using priori computing module 21 as a result q (x) as initial predicted model f
0(x), that is:
Adopt initial predicted model f
0(x) predict, and will predict the outcome and flow to controller 3.f
t(x) be defined as and utilize t full sample (x constantly
t, y
t) carry out the unpredictable key variable forecast model behind the model modification.
5) data acquisition equipment 1 t that constantly will newly the collect relevant variable measurements x that easily surveys of sample constantly
t, being transported in the soft-sensing model prediction module 22, soft-sensing model prediction module 22 is directly according to current forecast model f
T-1(x) carry out the on-line prediction of new samples functional value, and the soft measurement that will the obtain f that predicts the outcome
T-1(x
t) flow to controller 3.Simultaneously, soft-sensing model prediction module 22 also will be somebody's turn to do constantly and easily be surveyed variable measurements x according to being correlated with of sample
tThe anticipation function value f that obtains
T-1(x
t), be transported in the determination module 23.
6) determination module 23 is preserved the anticipation function value f that t obtains constantly
T-1(x
t), collect the t unpredictable key variable measured value y of sample constantly up to data acquisition equipment 1
t, and with y
tBe transported in the determination module 23.(for convenience of explanation, the present invention is only at the difficult variate-value y that surveys of current sample
tPicking rate always carry out the idiographic flow explanation faster than the situation that next sample is easily surveyed the arrival rate of variate-value.For the situation that does not meet above-mentioned hypothesis, determination module 23 will be with the relevant variable measurements of easily surveying of formation form stored samples, wait for that corresponding unpredictable key variable measured value arrives after, just can start at the judgement of this sample and upgrade link.At the waiting time of determination module 23, the forecast model in the soft-sensing model prediction module 22 remains unchanged.)
Otherwise, can think that the prediction effect of current model is better, need not carry out model modification, make f
t(x)=f
T-1(x), return step 5), calculation process forwards in the soft-sensing model prediction module 22, proceeds the difficult on-line prediction of surveying variable, and the anticipation function value is constantly flowed to controller 3, and waits for the arrival of new samples.
7) on-line study model maintenance module 24 is utilized the new samples (x that obtains in the step 6)
t, y
t), to existing forecast model f
T-1(x) carry out maintenance update, and the forecast model after will upgrading:
Be transported in the soft-sensing model prediction module 22, replace original soft-sensing model, return step 5).
Afterwards, along with on-line study model maintenance module 24 constantly feeds back to soft-sensing model prediction module 22 with new forecast model, the soft-sensing model of 22 foundations of soft-sensing model prediction module also will be brought in constant renewal in, to obtain the more accurate prediction effect.
Forecast model after the renewal, promptly in the formula (3),
Be undetermined parameter; Parameter d and base vector s
iBy step 2) in priori computation model 21 determine.
1. at the data stream (x of sequential arrival
i, y
i)
I ∈ N, specify insensitive error band ε, learning rate η, coefficient of balance λ, regularization coefficient gamma;
2. remember initialization prediction hypothesis constantly
Wherein
s
i∈ S
d, the output result who promptly directly adopts priori computing module 21 is as the initial predicted model.Calculate the d * d dimension kernel function matrix K of base vector
d, K
d(i, j) individual element is K (s
i, s
j);
3. read new samples (x
t, y
t), according to current forecast model expression formula f
T-1(x) calculate the anticipation function value of new samples:
And prediction error value e
t=y
t-f
T-1(x
t).
4. construct d dimensional vector K
t=[k (s
1, x
t) ..., k (s
d, x
t)]
T, the solving-optimizing problem
Wherein,
Be base vector s
iThe unit increment of pairing kernel function coefficient.The purpose of finding the solution of this minimum problems is: ask for when base vector the linear combination of corresponding kernel function and up-to-date sample point place kernel function distance the most near the time, promptly the 2-norm of the two is apart from approximate value δ
(t)Hour, base vector the coefficient increment θ of corresponding kernel function
(t)
Calculating can get base vector the incremental vector θ of unit of corresponding kernel function coefficient
(t)=(K
d+ γ I
d)
-1K
t, wherein, I
dBe d dimension unit matrix, γ is the regularization coefficient of control function complexity.
5. upgrade base vector the coefficient of corresponding kernel function, order
I=1 ..., d; Wherein, set learning rate η, can control the renewal speed of forecast model, prevent that the sharp-pointed noise in the sample from making model produce than large deviation.Simultaneously,, prevent that the model that forecast model is caused by error effect from departing from, set coefficient of balance λ again, with the result of balanced experimental formula and on-line prediction model in order to make full use of the high frequency core information that experimental formula comprises.
6. obtaining t anticipation function constantly is
The model modification result is outputed to soft-sensing model prediction module 22, and calculation process forwards soft-sensing model prediction module 22 to, continues to wait for the arrival of new samples.
Above-mentioned steps 3) in, in priori computing module 21, undetermined parameter d and β
i, and base vector s
iConcrete solution procedure as follows:
1. use the mode of experience directly perceived or circulation optimizing, the situation of formula p (x) rule of thumb, definite kernel function K (s
i, type x) and parameter are specified regularization coefficient μ, convergence threshold ε
f
2. make iteration step number k=0, note expression function q (x) at this moment is q
0(x)=0, then error function is: r
0(x)=p (x)-q
0(x);
3. obtain the base vector s of k+1 iteration
K+1, make error function
If | r
k(s
K+1) |<ε
f, d=k then, q (x)=q
k(x), base vector s
iValue set be S
k,
I=1 ..., k, termination of iterations;
4. make S
K+1=S
k∪ { s
K+1, then construct one (k+1) * (k+1) dimension matrix K
K+1, K
K+1(i, j) individual element is K (s
i, s
j), and then construct (k+1) dimensional vector p (S
(k+1))=[p (s
1) ..., p (s
K+1)]
T
5. compute vector matrix β
(k+1)=(K
K+1+ μ I
K+1)
-1P (S
(k+1)), I wherein
K+1Be (k+1) dimension unit matrix, order
Wherein
Be vector matrix β
(k+1)I element;
6. make r
K+1(x)=p (x)-q
K+1(x), 3. k=k+1 returns step.
Enumerating a concrete emulation example below is described in detail application of the present invention.
In this example, with [1,12] (y=sin (x)/x) is the basic change curve of system performance to the sinc function on the interval, add standard deviation on its basis and be 0.1 white noise, noise in the simulation practical problems, according to the actual industrial field condition, the mixed sequence of design time sequence and random series (abbreviation mixed sequence) produces emulated data.
The data production of mixed sequence as shown in Figure 3, produces 500 data on [1,12] interval, be 1~500 according to generation order number consecutively.The colour system of sample point by indigo plant to red representative data point sequence number variation tendency from small to large.Once equidistant data drift takes place every 100 chronomeres (after being 100 data of every generation) in the working point of supposing the system, and the direction of drift is the direction that independent variable x is increased.It is average that distribution of data points is obeyed with the working point, and standard deviation is 0.8 standardized normal distribution.Dependent variable among Fig. 3 (a) (treating measured value y) meets the sinc function fully, and the dependent variable among Fig. 3 (b) (treating measured value y) has then been added standard deviation on the basis of sinc function be 0.1 white noise.The easy survey variate-value x of data is corresponding, identical one by one among two figure.
In actual emulation, this paper uses the noise data that has among Fig. 3 (b) to carry out on-line study, uses the corresponding muting data among Fig. 3 (a) to carry out predicated error calculating simultaneously, to obtain more real estimation of error result.
Use above artificial data that the on-line learning algorithm that merges experimental formula is carried out emulation.Simulation process and parameter setting situation are as follows:
(1) adopts [1,12] interval sinc function as experimental formula, select the Gaussian kernel function
Specify regularization coefficient μ=0.01, insensitive error band ε=0.05, convergence threshold ε
f=0.05, learning rate η=0.5, coefficient of balance λ=0.7.
(2) adopt the sample sequence that contains noise to carry out emulation, sequential adding sample point is predicted and model modification, promptly whenever obtain a new samples, use original model that it is predicted, calculate prediction deviation according to actual value, utilize this sample to upgrade forecast model again, continue to repeat this process, all predict up to 500 sample points to finish.
(3) error is analyzed and compared, the simulation result under time series, random series, three kinds of situations of mixed sequence, as shown in Figure 4.
Relatively Chang Yong on-line study method and the prediction case of utilizing the present invention to measure, the predicted root mean square error value of the whole bag of tricks, as shown in the table:
The error ratio of five kinds of on-line study methods
As seen, adopt the present invention to carry out the online soft sensor prediction, its measuring accuracy is compared and will be improved a lot with the priori common method.
The various embodiments described above only are used to illustrate the present invention, and wherein the structure of each parts, connected mode etc. all can change to some extent, and every equivalents of carrying out on the basis of technical solution of the present invention and improvement all should not got rid of outside protection scope of the present invention.
Claims (3)
1. online soft sensor method based on knowledge fusion, it may further comprise the steps:
1) measuring system that comprises data acquisition equipment, supervisory control comuter and controller is set, a default priori computing module, a soft-sensing model prediction module, a determination module and the on-line study model maintenance module of being equipped with in the described supervisory control comuter;
2) p (x) among the experimental formula y=p (x) is expressed as a kind of linear combination form q (x) of kernel function:
Wherein, K (s
i, x) for the kernel function of appointment; s
i∈ S
d, (i=1 .. d) are base vector, S
dBe the basal orientation quantity set; β
iIt is a undetermined parameter;
3) with formula (1) input priori computing module, the parameter s in the iterative computation formula (1)
i, d, β
i, flow to the soft-sensing model prediction module then;
4) in the soft-sensing model prediction module, establish: initial time t=0, the relevant variable measurements of easily surveying that whenever collects a new samples then makes: t=t+1; Initial time, the q as a result (x) that directly uses the step 3) input is as initial predicted model f
0(x)
Initial predicted model f
0(x) the relevant variable measurements of easily surveying that collects according to data acquisition equipment is predicted the unpredictable key variable value, and will predict the outcome and flow to controller; Wherein,
s
i∈ S
d, (i=1 .., d);
5) data acquisition equipment the relevant variable measurements x that easily surveys of new samples that t is collected constantly
tInput soft-sensing model prediction module, the soft-sensing model prediction module is according to current forecast model f
T-1(x) carry out the on-line prediction of new samples functional value, and the soft measurement that will the obtain f that predicts the outcome
T-1(x
t) flow to controller and determination module;
6) determination module is preserved the anticipation function value f that t obtains constantly
T-1(x
t), collect the t unpredictable key variable measured value y of sample constantly up to data acquisition equipment
t, and with y
tThe input determination module; Determination module is surveyed variate-value y according to the difficulty of reality
t, judge anticipation function value f
T-1(x
t) accuracy;
If the absolute value of predicated error does not exceed insensitive error band ε, promptly | e
t|=| y
t-f
T-1(x
t) |≤ε, think that then the prediction effect of current model is better, need not carry out model modification, make f
t(x)=f
T-1(x), return step 5);
Otherwise, | e
t|=| y
t-f
T-1(x
t) |>ε, then with sample (x
t, y
t) be transported to on-line study model maintenance module, proceed next step;
7) on-line study model maintenance module is utilized the new samples (x that obtains in the step 6)
t, y
t), to existing forecast model f
T-1(x) carry out maintenance update, and the forecast model f after will upgrading
t(x):
Be transported in the soft-sensing model prediction module, replace original soft-sensing model, return step 5).
2. a kind of online soft sensor method based on knowledge fusion as claimed in claim 1 is characterized in that: in the described step 7), and the forecast model after the renewal, promptly in the formula (3),
Be undetermined parameter, ask for
Concrete steps as follows:
1. at the data stream (x of sequential arrival
i, y
i)
I ∈ N, specify insensitive error band ε, learning rate η, coefficient of balance λ, the regularization coefficient gamma of control function complexity;
2. remember initialization prediction hypothesis constantly
Wherein
s
i∈ S
d, the output result who promptly directly adopts the priori computing module is as the initial predicted model; Calculate the d * d dimension kernel function matrix K of base vector
d, K
d(i, j) individual element is K (s
i, s
j);
3. read new samples (x
t, y
t), according to current forecast model expression formula f
T-1(x) calculate the anticipation function value of new samples:
And prediction error value e
t=y
t-f
T-1(x
t);
4. construct d dimensional vector K
t=[k (s
1, x
t) ..., k (s
d, x
t)]
T, the solving-optimizing problem
Wherein,
Be base vector s
iThe unit increment of pairing kernel function coefficient; The purpose of finding the solution of this minimum problems is: ask for when base vector the linear combination of corresponding kernel function and up-to-date sample point place kernel function distance the most near the time, promptly the 2-norm of the two is apart from approximate value δ
(t)Hour, base vector the coefficient increment θ of corresponding kernel function
(t)
Calculating can get base vector the incremental vector θ of unit of corresponding kernel function coefficient
(t)=(K
d+ γ I
d)
-1K
t, wherein, I
dBe d dimension unit matrix;
3. a kind of online soft sensor method based on knowledge fusion as claimed in claim 1 or 2 is characterized in that: in the described step 3), and in the priori computing module, undetermined parameter d and β
i, and base vector s
iConcrete solution procedure as follows:
1. use the mode of experience directly perceived or circulation optimizing, the situation of formula p (x) rule of thumb, definite kernel function K (s
i, type x) and parameter are specified regularization coefficient μ, convergence threshold ε
f
2. make iteration step number k=0, note expression function q (x) at this moment is q
0(x)=0, then error function is: r
0(x)=p (x)-q
0(x);
3. obtain the base vector s of k+1 iteration
K+1, make error function
If | r
k(s
K+1) |<ε
f, d=k then, q (x)=q
k(x), base vector s
iValue set be S
k,
I=1 ..., k, termination of iterations;
4. make S
K+1=S
k∪ { s
K+1, then construct one (k+1) * (k+1) dimension matrix K
K+1, K
K+1(i, j) individual element is K (s
i, s
j), and then construct (k+1) dimensional vector p (S
(k+1))=[p (s
1) ..., p (s
K+1)]
T
5. compute vector matrix β
(k+1)=(K
K+1+ μ I
K+1)
-1P (S
(k+1)), I wherein
K+1Be (k+1) dimension unit matrix, order
Wherein
Be vector matrix β
(k+1)I element;
6. make r
K+1(x)=p (x)-q
K+1(x), 3. k=k+1 returns step.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2010102631005A CN101916394A (en) | 2010-08-25 | 2010-08-25 | Knowledge fusion-based online soft measurement method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2010102631005A CN101916394A (en) | 2010-08-25 | 2010-08-25 | Knowledge fusion-based online soft measurement method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN101916394A true CN101916394A (en) | 2010-12-15 |
Family
ID=43323903
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN2010102631005A Pending CN101916394A (en) | 2010-08-25 | 2010-08-25 | Knowledge fusion-based online soft measurement method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN101916394A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103020715A (en) * | 2012-11-29 | 2013-04-03 | 北京航空航天大学 | Fusion estimation method of poor-information multi-sensor data based on fuzzy self-service theory |
CN103714262A (en) * | 2014-01-20 | 2014-04-09 | 华北电力大学 | Method for updating thermotechnical soft measurement model based on buffer block sequence design |
CN104537260A (en) * | 2015-01-14 | 2015-04-22 | 清华大学 | Dynamic soft measurement method and system based on slow characteristic regression |
CN104123476B (en) * | 2014-08-12 | 2017-03-08 | 大连海事大学 | Forecasting of Gas Concentration method based on extreme learning machine and its device |
CN106909747A (en) * | 2017-03-06 | 2017-06-30 | 北京航空航天大学 | Fuzzy parameter membership function recognition methods in a kind of thermal convection current diffusion system |
CN109960146A (en) * | 2019-01-25 | 2019-07-02 | 石化盈科信息技术有限责任公司 | The method for improving soft measuring instrument model prediction accuracy |
CN115291505A (en) * | 2022-07-05 | 2022-11-04 | 北京可视化智能科技股份有限公司 | Method, device and system for improving automatic control loop application rate |
-
2010
- 2010-08-25 CN CN2010102631005A patent/CN101916394A/en active Pending
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103020715A (en) * | 2012-11-29 | 2013-04-03 | 北京航空航天大学 | Fusion estimation method of poor-information multi-sensor data based on fuzzy self-service theory |
CN103020715B (en) * | 2012-11-29 | 2016-06-29 | 北京航空航天大学 | A kind of weary information Fusion method of estimation based on fuzzy Self-booster opinion |
CN103714262A (en) * | 2014-01-20 | 2014-04-09 | 华北电力大学 | Method for updating thermotechnical soft measurement model based on buffer block sequence design |
CN103714262B (en) * | 2014-01-20 | 2016-08-10 | 华北电力大学 | A kind of thermal technology's soft-sensing model update method based on buffer stopper timing Design |
CN104123476B (en) * | 2014-08-12 | 2017-03-08 | 大连海事大学 | Forecasting of Gas Concentration method based on extreme learning machine and its device |
CN104537260A (en) * | 2015-01-14 | 2015-04-22 | 清华大学 | Dynamic soft measurement method and system based on slow characteristic regression |
CN104537260B (en) * | 2015-01-14 | 2018-02-09 | 清华大学 | The dynamic soft-measuring method and system returned based on slow feature |
CN106909747A (en) * | 2017-03-06 | 2017-06-30 | 北京航空航天大学 | Fuzzy parameter membership function recognition methods in a kind of thermal convection current diffusion system |
CN106909747B (en) * | 2017-03-06 | 2020-06-30 | 北京航空航天大学 | Fuzzy parameter membership function identification method in heat convection diffusion system |
CN109960146A (en) * | 2019-01-25 | 2019-07-02 | 石化盈科信息技术有限责任公司 | The method for improving soft measuring instrument model prediction accuracy |
CN115291505A (en) * | 2022-07-05 | 2022-11-04 | 北京可视化智能科技股份有限公司 | Method, device and system for improving automatic control loop application rate |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN101916394A (en) | Knowledge fusion-based online soft measurement method | |
CN101315557B (en) | Propylene polymerization production process optimal soft survey instrument and method based on genetic algorithm optimization BP neural network | |
Jiang et al. | A cuckoo search-support vector machine model for predicting dynamic measurement errors of sensors | |
CN109214708B (en) | Electric power system risk assessment method based on cross entropy theory optimization support vector machine | |
CN103268069B (en) | Based on the adaptive prediction control method of Hammerstein model | |
CN106843172B (en) | Complex industrial process On-line quality prediction method based on JY-KPLS | |
CN102968529B (en) | Method for quantifying computed result non-determinacy interval of water supply pipe network model | |
CN103177289B (en) | Modeling method for noise-uncertainty complicated nonlinear dynamic system | |
CN109615860A (en) | A kind of signalized intersections method for estimating state based on nonparametric Bayes frame | |
Tian et al. | Time-delay compensation method for networked control system based on time-delay prediction and implicit PIGPC | |
CN102004444A (en) | Multi-model predictive control method for component content in process of extracting rare earth | |
CN104048676B (en) | MEMS (Micro Electro Mechanical System) gyroscope random error compensating method based on improved particle filter | |
CN105046046B (en) | A kind of Ensemble Kalman Filter localization method | |
CN110210687A (en) | A kind of Nonlinear Dynamic production process product quality prediction technique returned based on local weighted slow feature | |
Grema et al. | Optimal feedback control of oil reservoir waterflooding processes | |
CN105511263A (en) | Distributed model predictive control method based on hierarchical decomposition | |
CN110296833A (en) | A kind of flexible measurement method and system of Hydraulic Cylinder combined test stand | |
CN113033920A (en) | Method and system for predicting marine drift trajectory | |
CN104950670A (en) | Integrated multi-model method for controlling CSTRs (continuous stirred tank reactors) | |
CN104794332B (en) | A kind of Uncertainty Analysis Method of skyscraper wind-excited responese analysis model | |
CN107025609A (en) | Electrical Power System Dynamic method for estimating state based on singular value decomposition CDKF | |
CN111859249A (en) | Ocean numerical forecasting method based on analytical four-dimensional set variation | |
CN103279030B (en) | Dynamic soft measuring modeling method and device based on Bayesian frame | |
CN103885867A (en) | Online evaluation method of performance of analog circuit | |
Pengfei et al. | Soft sensor model derived from wiener model structure: modeling and identification |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C12 | Rejection of a patent application after its publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20101215 |