CN104950873A - Method for detecting intermittent oscillation of industrial control circuits in online manner - Google Patents

Abstract

The invention discloses a method for detecting intermittent oscillation of industrial control circuits in an online manner. The method includes steps of (1), preliminarily acquiring a big window historical datum from each to-be-detected control circuit, processing the big window historical data and initializing detection systems; (2), acquiring a small window process datum from each to-be-detected control circuit in real time in an online manner, and updating the big window data by the aid of the small window process data; (3), combining the small window data with the original big window data, processing the small window data and the big window data and recombining the small window data with the big window data to obtain various frequency recombined sub-signals of the updated big window data; (4), computing monitoring statistics corresponding to the various recombined sub-signals and judging real-time detection results; (5), repeatedly carrying out steps (2), (3) and (4) and integrating all judgment results to obtain online detection results. The method has the advantages that intermittent and multi-period oscillation behavior of the industrial control circuits can be quantitatively detected by the aid of the method, compositions such as multiple-oscillation, intermittent oscillation and non-stationary signals in the intermittent and multi-period oscillation behavior can be differentiated from one another, and regular degrees and periods of various oscillation components can be acquired.

Description

The online test method of process control loops interrupted oscillation

Technical field

The present invention relates to the Performance Evaluation field in industrial control system, particularly relate to a kind of online test method of process control loops interrupted oscillation.

Background technology

Modern industry process device has scale, and degree is high greatly, comprehensively, manipulation is complicated, variable is many, and long-play is in the inferior feature of closed-loop control.The chemical process that industry is common, often comprise thousands of control loops, and these loops interacts due to coupled relation.Cross due to process control loops middle controller adjust, the ubiquity of the characteristic such as external disturbance and variable valve nonlinear operation, the oscillation behavior of control loop occurs often, and this greatly have impact on economic benefit and stability that industrial flow equipment runs.

Preliminary oscillation test is accurately carried out to industrial flow equipment and can reduce waste product turnout, reduce disqualification rate, increase the reliability in industrial flow equipment running process, security, reduce manufacturing cost simultaneously.Many controllers can also keep good performance at initial operating stage, but As time goes on, and due to the impact of external interference factor or equipment self problem, the performance of controller can reduce gradually and even lost efficacy.Be embodied in control loop process and all kinds of oscillation behavior occurs, wherein may comprise multiple vibration, interrupted oscillation, the composition such as non-linear, thus threaten the safe and stable operation of industrial process.Meanwhile, because apparatus of load in real time environment and operating mode often change, industrial process also shows the one side of Non-stationary Data characteristic, is embodied in the local mean value transport phenomena of process data.For important control loop, the oscillating characteristic of its operational process of Timeliness coverage contributes to engineering staff and carries out fault diagnosis and investigation.Therefore, in industrial control system Performance Evaluation process, design effective on-line monitoring means, in time, all kinds of oscillationg components of non-stationary process data in control loop are accurately detected, and distinguish frequency ranges different separately, for controller performance assessment and control loop fault diagnosis all important in inhibitings.

Existing process control loops oscillation test technology, the overwhelming majority is the off-line analysis method based on stationary process data.Also some detection method of oscillations for non-stationary process data have been there are in nearest Two decades years.Three kinds can be roughly summarized as: the analytical approach of Kernel-based methods data Time-domain Statistics by its main thought; The analytical approach of the autocorrelation function (ACF) of Kernel-based methods data; And the signal decomposition method of Kernel-based methods data (comprising empirical mode decomposition EMD and base conversion decomposition).But the detection method of Kernel-based methods data Time-domain Statistics or autocorrelation function domain analysis has three shortcomings in commercial Application: one, the method needs to treat measure loop or process has certain priori, and some parameter is also empirically determine; Two, to the industrial process that non-stationary and many oscillation period exist, cannot realizing automatically detecting without intervening, needing to design wave filter targetedly and carrying out data tranquilization process and be separated with oscillationg component; Three, most detection algorithm quantitatively cannot calculate the regular degree of oscillationg component.The signal decomposition method of current Kernel-based methods data is compared above-mentioned detection method and is existed progressive, but its limitation is mainly reflected in: the subsignal number of plies redundancy that existing signal decomposition technical point solution obtains is various, many subsignals are caused to lack the support of actual physics meaning, not there is good representativeness, and these methods are also poor to the degree of fitting of non-stationary signal trend, computation complexity is also higher.In addition, in existing interrupted oscillation, multicycle oscillation test technology, mostly require that method off-line carries out, still can show good method in on-line operation few.

Publication number is the online test method that the Chinese invention patent of CN103970124A discloses a kind of process control loops multicycle oscillation behavior, comprises the steps: in control loop to be detected, online real time collecting one group of process data; Online in real time to the essential time scale decomposition that process data is improved, and calculate the monitoring statistic of each decomposition of gained corresponding to subsignal in real time; Judge whether each monitoring statistic exceedes the threshold value Ω of setting, obtains on-line checkingi result according to all judged results.When there is the behavior of interrupted oscillation in industrial process loop, only carry out merely essential time scale decomposition, the subsignal component belonging to different frequency scope easily mixes mutually, and this is totally unfavorable to our oscillation test; In addition, irregular spike disturbance ubiquity in industrial process, when this disturbance has higher magnitude, probably causes the decomposition level of essential time scale to increase emptily situation.

In the practical application of process oscillation detection algorithm, detect process control loops and whether there is oscillation behavior, and the rule degree index of qualitative assessment oscillation behavior, generally be applicable to exist the process data of multicycle vibration, interrupted oscillation, non-stationary and non-linear component, and the lot data that can rely on a less moving window realizes on-line checkingi, existence for the vibration of Accurate Diagnosis industrial process has very important Practical significance, is also conducive to the control performance qualitative assessment of industrial process.

Summary of the invention

The invention provides a kind of process control loops interrupted oscillation online test method, the process control loops process that there is the behaviors such as multicycle vibration, interrupted oscillation can be applicable to.

An online test method for process control loops interrupted oscillation, comprises the steps:

(1) at control loop to be detected, gather large windows history data in advance, described historical data is decomposed by essential time scale, half-wave information extraction, based on after the Robust clustering of frequency and signal restructuring process, obtains the initial estimation of each frequency content in former data;

(2) in control loop to be detected, after online real time collecting wicket process data is placed on former large window data, large window slides backward, and forms a new large window data with the wicket process data of Real-time Collection;

(3) the wicket data of Real-time Collection are decomposed through essential time scale, after half-wave information extraction, do Robust clustering process in conjunction with remaining half-wave information in existing large window and recombinate, the recon signal obtained is the real-time estimation of each frequency content in the large window data after renewal;

(4) calculate the monitoring statistic corresponding to each subsignal, judge whether each sub-recombination signal is in oscillatory regime, the judged result of comprehensive each recon signal, is the real-time testing result of current large window data;

(5) repeat step (2) ~ step (4), get final product the operation conditions of real-time follow-up control loop to be detected.

The present invention directly adopts the measurable variable of chemical process as process data, and these data are obtained by field real-time acquisition, namely and along with passage of time, constantly gather and renewal process data to supervisory system.After the former large window data of each renewal, first essential time scale decomposition is carried out to data in gathered wicket, obtain decomposing subsignal set { x ⁱ, then the robust K average value processing based on frequency is done to it; In conjunction with the current cluster structures of all the other half-wave information updatings in large window, and respective sub-signal of recombinating obtains recon signal; Then each recon signal is calculated corresponding monitoring statistic the computation complexity of this statistic is minimum, also can carry out in real time large batch of multi-group data simultaneously.Finally, judged by defined threshold Ω, when a certain recon signal corresponding monitoring statistic when exceeding this threshold value, illustrate that this recon signal and original signal vibrate.

In step (1), the collection method of large windows history data is: in each sampling interval, record the process data in control loop to be detected, and the data collected in each sampling interval are added on the process data end previously gathered, until collected data meet prespecified large window size.

Sampling interval refers to the sampling interval of performance evaluation system.Process data x constantly updates along with passage of time, often through the time span of a sampling interval, all has new process data to add the end of the process data previously gathered to.The sampling interval of performance evaluation system is general identical with the control cycle in industrial control system, also can be chosen as the integral multiple of control cycle, specifically limits according to the requirement of real-time of performance monitoring and industry spot and memory data output and determines.

For reaching the real-time performance of detection system, ensure the reliability of testing result simultaneously, the size of wicket may be selected to be about 0.2 ~ 0.3 times of large window size, and the selection of large window size then considers according to maximum cycle vibration in control loop, industry spot requirement of real-time, industry vibration degree of admission and memory data output.

As preferably, in step (1) and step (3), original signal obtains the decomposition subsignal of different frequency after essential time scale is decomposed, and wherein the stop condition of essential time scale decomposition method is the index of oscillation I < 0.7 of residual components.

Essential time scale is used to decompose in the present invention, according to prior art (list of references: Frei M G, Osorio I.Intrinsic time-scale decomposition:time-frequency-energy analysis and real-time filtering of non-stationary signals [J] .Proceedings of the Royal Society A:Mathematical, Physical and Engineering Science, 2007,463 (2078): 321-342) implement.For the end condition (the essential time scale of improvement is decomposed) that essential time scale is decomposed, namely this index of index of oscillation I < 0.7 of residual components is decomposed, according to prior art (list of references: Online detection of time-variant oscillations based on improved ITD.Zixu, Guo; Lei, Xie; Taihang, Ye; Alexander, Horch.Control Engineering Practice vol.32issue 8August, 2014.p.64-72) obtain.

The so-called essential time scale improved is decomposed, refer to and to improve on original essential time scale decomposition base, remain former methodical all mathematics and calculate feature, just carry out simplifying and revising on end condition, the decomposition method improved is for same process data, compared to former method, the decomposition subsignal quantity of acquisition is less, is more suitable for analyzing original signal oscillation behavior.This decomposition computation complexity is very low, and online carrying out that therefore can be real-time, completes calculating in each sampling interval, decomposes subsignal arrangement set { x ⁱ.Namely retain original subsignal structure and extracting method constant, and the condition its original method being stopped decomposition is revised as the index of oscillation I < 0.7 of residual components.

As preferably, in step (1) and step (3), described half-wave is the signal segment that each layer decomposes in subsignal between two continuous zero cross points; Half-wave information extraction comprises: reject and represent noise signal; Collect residue and decompose the wavelength of all half-waves in subsignal and initial, end time thereof.

The extraction of half-wave voltage signal in the present invention, what first carry out is noise level rejecting machine system.Decompose through essential time scale in the decomposition subsignal obtained, some frequency of decomposing subsignal is very high, and its number of times passing through zero cross point is far away more than the decomposition subsignal of other levels.Therefore, can be similar to and be thought noise level, namely in process control loops, the composition of high frequency like this does not belong to vibration category.

In the present invention, as preferably, the basis for estimation of noise signal is statistic A, and its computing formula is:

$A=\frac{C_k}{N},$

Wherein C _kfor decomposing the zero cross point number of subsignal autocorrelation function, N is hits; If statistic A is greater than boundary value A _lim, then think that this decomposition subsignal is noise signal.A _limgeneral value, can according to Operating condition adjustment 0.3 ~ 0.5; Disallowable noise signal, the cluster after no longer participating in and restructuring link.

In the present invention, as preferably, in step (1) and step (3), based on the K means clustering algorithm of Robust clustering for being individuality with extracted half-wave, being feature with the wavelength of each half-wave of frequency;

K means clustering algorithm in the present invention is according to prior art (list of references: Some methods for classification and analysis of multivariate observations.Macqueen, J.In:Proceedings of 5th Berkeley Symposium on Mathematical Statistics and Probability; 1967.p.281-297) obtain.

And so-called robust K mean cluster, refer to and to improve on the basis of original K mean cluster, determine at the selection of cluster number, initial cluster center and avoid being improved in local optimum etc. three, making traditional K mean algorithm have more robustness:

As preferably, in step (1):

After optimum cluster number k equals cancelling noise signal, residue decomposes the number of subsignal;

Initial cluster center P _ifor each layer of correspondence decomposes the median of the half-wave wavelength that subsignal comprises ${\mathrm{\&mu;}}_T^k=\mathrm{median}\left(T\right);$

The avoiding method of its local optimum is: restrained center after K convergence in the mean as initial center again cluster to verify cluster whether Yi Da global optimum.

Online real time collecting wicket process data, refers to and record process data in control loop to be detected in each sampling interval preset, and the process data collected in each sampling interval is added on the process data end previously gathered.When after the small window size that collected hits reaches prespecified, no longer add data to current wicket, then newly-built next wicket.

And to upgrade former large window data method be the wicket real time data of collection is added on former large window procedure data end, and leave out the data of a part foremost in this large window, leave out the size that data size equals used wicket.

All half-waves that wicket data extract combine all the other half-waves in existing large window and do the Robust clustering process based on frequency.

As preferably, in step (3):

Existing window cluster structures to be upgraded before cluster, comprise optimum cluster number K ^*with the renewal of initial cluster center;

Optimum cluster number K ^*for:

K ^*＝argmax{S(K _old-1)，S(K _old)，S(K _old+1)}，

Wherein S (*) is clustering result quality index-isolation, K _oldcomprise by former large window the cluster number of data;

Basis prior art (the list of references: Well-separated clusters and optimal fuzzy partitions.Dunn of isolation S (*), J.C.Cybernetics and Systems vol.4 issue1,1974.p.95-104) obtain.K ^*the Analysis of Deep Implications of calculating like this is, with the renewal of wicket data, the change of cluster structures is inviolent, and the cluster number of its optimum is only fluctuate ± 1 on former cluster numbers basis.So do the calculated amount that can reduce whole process, if cluster structures change is more violent in actual condition, can K be expanded ^*hunting zone to find optimum cluster number.

The renewal of initial cluster center: initial cluster center, selects according to current optimum cluster number.

As preferably, in step (3), the method for class initial cluster center is selected to be according to current optimum cluster number:

If K ^*=K _old, cluster number does not change, and class initial center is C _new=C _old;

When cluster number does not change, the situation of maximum probability is: the cluster structures of large window data does not change, and therefore selects former large window cluster result as existing large window initial cluster center.Certainly, even if cluster structures there occurs change (as: a newly-increased class, Geju City class disappears) in large window simultaneously, robust K mean cluster result can converge to final cluster centre equally.

If K ^*=K _old+ 1, have more a class in existing large window procedure data, then such initial cluster center is chosen as:

$C_{\mathrm{add}}\&LeftArrow;\underset{1\&le;j\&le;n}{\arg \max }\left\{\underset{1\&le;i\&le;m}{\min }\right\{\frac{\mathrm{dis}(c_i,c_j)}{\mathrm{rad}\left(c_i\right)}\left\}\right\};$

Wherein, m is the optimum cluster number of former large window data, c _ifor corresponding cluster centre; N is the cluster number of wicket data, c _jfor corresponding cluster centre; Dis (c _i, c _j) be Euclidean distance between two cluster centres; Rad (c _i) be class c _icorresponding class radius;

Its Analysis of Deep Implications, for pick out an initial center in existing wicket, adds the K of former large window _oldindividual cluster centre is as the K of existing large window _old+ 1 cluster initial center, and center selected in wicket is and former large window K _oldthe center that individual cluster centre is the most irrelevant in relative distance.

If K ^*=K _old-1, existing large window procedure data have lacked a class, then should leave out a class in former cluster structures:

$C_{\mathrm{del}}\&LeftArrow;\underset{1\&le;i\&le;n}{\arg \max }\left\{\underset{1\&le;j\&le;m}{\min }\right\{\frac{\mathrm{dis}(c_j,c_i)}{\mathrm{rad}\left(c_j\right)}\left\}\right\};$

Wherein, m is the optimum cluster number of former large window data, c _ifor corresponding cluster centre; N is the cluster number of wicket data, c _jfor corresponding cluster centre; Dis (c _i, c _j) be Euclidean distance between two cluster centres; Rad (c _i) be class c _icorresponding class radius;

Its Analysis of Deep Implications is at former large window K _olddelete an initial center in individual cluster centre, remaining cluster centre is as the K of existing large window _old-1 initial cluster center, and the center of deleting is a center the most irrelevant in relative distance with the cluster centre of existing wicket.

As preferably, in step (1) and step (3), the method obtaining recon signal is: after the process of K mean cluster, will to belong in of a sort all half-wave computing with words to same subsignal to obtain recon signal.Be specially; Initial, end time according to each half-wave voltage signal determine half wave characteristic, same class half-wave is recombinated in same layer recon signal, and the part null value lacking half-wave in same layer recon signal substitutes, the part of overlapping half-wave is then by these half-wave linear, additive.Recon signal is by frequency descending sort from high to low.

The existing large window of restructuring, through all kinds of half-wave being in different frequency scope of cluster gained, can obtain each layer frequency recon signal upgrading rear process data.

As preferably, in step (4), monitor statistic computing formula be:

${\widehat{\mathrm{\&eta;}}}^k=\frac{\sqrt{N-1}}{\sqrt{{\mathrm{\&chi;}}_{N-\mathrm{1,1}-\mathrm{\&alpha;}/2}^2}}\frac{{\mathrm{\&mu;}}_T^k}{{\mathrm{\&sigma;}}_T^k},$

Wherein, N is the data length of this recon signal, when be degree of confidence being 1-α, degree of freedom is that the card side of N-1 distributes critical value (card side's distribution critical value can table look-up acquisition), for the median of half-wave wavelength in this recon signal, for utilizing Q _nthe robustness variance that algorithm for estimating calculates.

For each recon signal, its monitoring normalized set method comprises the steps:

Step (4-1), obtain each layer recon signal zero passes through an intervening sequence, for a kth recon signal zero to pass through an intervening sequence be T for it ^k;

Step (4-2), calculates zero and passes through an intervening sequence T ^kmedian

Step (4-3), calculates zero and passes through an intervening sequence T ^krobustness variance

Step (4-4), according to median with robustness variance calculate monitoring statistic

Wherein in step (4-3), the Qn estimation technique can be utilized to calculate its robustness variance

The concrete mode obtaining on-line checkingi result according to judged result is: if one of them monitoring statistic exceed threshold value Ω, then judge the recon signal that this control loop is corresponding there is vibration, if there is multiple recon signal to there is oscillation behavior in the process data gathered, then variable oscillation behavior when judging that this control loop exists the multicycle.Wherein during the multicycle, variable oscillation may comprise the compositions such as multiple vibration, interrupted oscillation.

In the present invention, as preferably, described threshold value Ω is 3, namely when explanation in there is oscillation behavior.

Afterwards, the process data of every Real-time Collection wicket, is updated in former large window data immediately.Thereafter upgrade existing window cluster structures, detect each layer frequency component corresponding to existing window procedure data and whether there is oscillatory occurences, reach the on-line checkingi object of process control loops discontinuous, multicycle oscillation behavior with this.

The beneficial effect that the present invention compared with prior art has:

(1) algorithm encourages without the need to outside additional signal, also can not introduce additional disturbance to system, can realize the Detection and diagnosis of non-intrusion type;

(2) the essential time scale adopted decomposes the automatic separation achieving non-stationary component in process data, and compared to prior art, its decomposition efficiency is higher, and computation complexity is lower;

(3) decompose to each layer the half-wave voltage signal that subsignal extracts and adopt Robust clustering process, the half-wave misplaced everywhere can be recombinated the level that it should belong to, can obtain physical significance more obviously, more reliable result;

(4) quantizating index detection can be carried out to the interval of process control loops, multicycle oscillation behavior, for the assessment of loop performance to be detected and source of trouble diagnosis provide abundant Data support;

(5) adopt the method for data driven type completely, without the need to process priori, without the need to designing filter in advance, also do not need to carry out manual intervention;

(6) computation complexity is low, convenient operation, and algorithm is write simple and easy, is beneficial to and implements on existing DCS workstation or control system host computer.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the chemical process in present example of the present invention;

Fig. 2 is the process data of one group of furnace outlet temperature control loop of Real-time Collection in present example of the present invention;

Fig. 3 is the implementation result of the present invention in the present example within the scope of 1 ~ 2000Samples;

Fig. 4 is the present invention's monitoring statistic that each recon signal is corresponding within the scope of 1 ~ 2000Samples in the present example and threshold value Ω position, wherein do not illustrate noise component;

Fig. 5 is the discomposing effect that essential time scale decomposes in the present example within the scope of 1 ~ 2000Samples;

Fig. 6 is the discomposing effect that essential time scale decomposes in the present example within the scope of 1 ~ 2800Samples;

Fig. 7 is the implementation result of the present invention in the present example within the scope of 1 ~ 3000Samples;

Fig. 8 is the present invention's monitoring statistic that each recon signal is corresponding within the scope of 1 ~ 3000Samples in the present example and threshold value Ω position, wherein do not illustrate noise component;

Fig. 9 is method flow diagram of the present invention.

Embodiment

Be example for the Performance Evaluation of main heating furnace in certain large petrochemical plant delayed coking production run domestic below, the detection method of the process control loops interrupted oscillation that there is operation valve viscosity property is described in detail.

As shown in Figure 1, petrochemical process heating furnace is one of important step and main energy consumption element in production procedure, and the steady control of heater outlet temperature is for raising product quality and reduce energy consumption important in inhibiting.

Heating furnace supplies heat-obtaining by device in Gas, gas amount fluctuates according to the change of upstream oiliness, needing to control air intake makes device in Gas Thorough combustion to obtain maximum heat, should ensure certain air surplus, but too much Cryogenic air can take away furnace heat simultaneously, cause waste of fuel, loss economic benefit, therefore, using furnace outlet temperature as controlled variable, fuel device in Gas aperture carries out circuit controls as performance variable, and process exists random perturbation simultaneously.

Device in Gas degree adjustment valve (operation valve) belongs to the topworks of this control loop, and occur certain nonlinear characteristic after running a period of time, because reasons such as adjusting crossed by controller, persistent oscillation behavior easily appears in control loop.And external disturbance also can introduce this loop by coupling circuit, loop is caused to produce other hunting of frequencys.Furnace outlet temperature data under the process data that present example of the present invention gathers is device in Gas variable valve viscous situation, when having again the external disturbance of interval to introduce.Furnace outlet temperature data after centralization as shown in Figure 2, in Fig. 2, horizontal ordinate is sampled point ordinal number, unit is Samples (sampling interval of the corresponding data of 1 Sample), and ordinate is furnace outlet temperature under the nominal situation after centralization, and unit is DEG C.

As shown in Figure 9, the specific embodiment of the present invention is as follows:

Step (1):

In control loop to be detected, designing large window size is 600 Samples, and small window size is 200 Samples.As shown in Figure 2, front 600 sampled points of process data y that real-time online gathers will use as detection system initialization.Detail is as follows:

To the process data y collected _t, carry out respectively essential time scale decomposition, half-wave information extraction, based on the robust K mean cluster of frequency and signal restructuring.Wherein stop decompose condition be, the index of oscillation I < 0.7 of residual components; Half-wave information extraction refers to the spacing (sampling interval) of extracting each decomposition subsignal adjacent zero crossings point and initial, end time accordingly, rejects the subsignal belonging to noise, y simultaneously ¹be noise level (as shown in Figure 3); The K mean cluster of robust is with the One-dimensional clustering algorithm of each half-wave corresponding wavelength tolerance; Signal restructuring each class is recombinated respectively in descending order by frequency to arrive corresponding level recon signal.So can obtain the initial estimation of each layer frequency content of this large window data.

Step (2):

After initialization terminates, every real-time collecting small window size (200Samples) process data, upgrades data in large window simultaneously.Such as, existing wicket is 601st ~ 800Samples, then existing large window slides into 201 ~ 800Samples from 1 ~ 600Samples.Existing wicket data essence time scale is decomposed and after extracting half-wave information, in conjunction with the half-wave in 201 ~ 600Samples, does new round K mean cluster, wherein comprise:

Step (2-1), cluster number upgrades: in existing large window data, its optimum cluster number K ^*for:

K ^*＝argmax{S(K _old-1)，S(K _old)，S(K _old+1)}

Wherein S (*) is clustering result quality index-isolation, K _oldfor former large window comprises the cluster number of data.Now find to have more a classification (y ³the interrupted oscillation composition characterized, as shown in Figure 3), its optimum cluster number has become 3 from 2.

Step (2-2), K ^*=K _old+ 1, then adopt formula:

$C_{\mathrm{add}}\&LeftArrow;\underset{1\&le;j\&le;n}{\arg \max }\left\{\underset{1\&le;i\&le;m}{\min }\right\{\frac{\mathrm{dis}(c_i,c_j)}{\mathrm{rad}\left(c_i\right)}\left\}\right\}$

Increase a new initial cluster center, wherein, m is the optimum cluster number of former large window data, c _ifor corresponding cluster centre; N is the cluster number of wicket data, c _jfor corresponding cluster centre; Dis (c _i, c _j) be Euclidean distance between two cluster centres; Rad (c _i) be class c _icorresponding class radius.

Step (2-3), after obtaining optimum cluster number and corresponding initial center, does clustering processing to all half-waves.

Step (3):

After renewal cluster structures, the corresponding half-wave of the existing large window cluster result of restructuring, obtains upgrading rear each layer recon signal.

Step (4):

For each recon signal { y ², y ³, y ⁴, calculate corresponding monitoring statistic in real time following (the wherein y of account form ¹for noise level, do not do oscillation test):

Step (4-1), obtain each decomposition subsignal zero passes through an intervening sequence, decomposes subsignal y for kth ^k, zero to pass through an intervening sequence be T for it ^k;

Step (4-2), calculates zero and passes through an intervening sequence T ^kmedian

Step (4-3), utilizes Q _nalgorithm for estimating calculates zero and passes through an intervening sequence T ^krobustness variance

Step (4-4), according to median with robustness variance calculate monitoring statistic monitoring statistic computing formula as follows:

${\widehat{\mathrm{\&eta;}}}^k=\frac{\sqrt{N-1}}{\sqrt{{\mathrm{\&chi;}}_{N-\mathrm{1,1}-\mathrm{\&alpha;}/2}^2}}\frac{{\mathrm{\&mu;}}_T^k}{{\mathrm{\&sigma;}}_T^k},$

Wherein, N is the data length of this decomposition subsignal, when be degree of confidence being 1-α, degree of freedom is that the card side of N-1 distributes critical value.Degree of confidence 1-α gets 0.95 in embodiments of the present invention, and corresponding parameter alpha is 0.05, monitoring statistic each sampling interval real-time result of calculation as shown in Figure 4, if certain recon signal to statistical value more than 3, then have 0.95 degree of confidence to think and occurred oscillatory occurences in loop to be detected.

Step (5):

When existing large window is monitored after normalized set terminates in real time, waiting system collects the next wicket data being of a size of 200Samples, and the large window that constantly slides, realize the real-time follow-up of control loop operation conditions to be detected.Fig. 3 is the history process curve of wicket detection system when being updated to 1801 ~ 2000Samples, clearly can see y from figure ⁴corresponding signal is persistent oscillation composition in control loop, y ³then by the higher-frequency interrupted oscillation composition because causing such as external disturbance in process operation, it occurs from first wicket (601 ~ 800Samples), and disappear the 5th wicket (1401 ~ 1600Samples), its feature interpretation is consistent with observations in original signal y, thus demonstrates reliability of the present invention and validity.

Fig. 4 is restructuring subsignal { y ², y ³, y ⁴corresponding monitoring statistic result of calculation, as can be seen from the figure, monitoring statistic with exceed defined threshold Ω=3, threshold value is shown in dotted line, and this loop corresponding subsignal component y is described ³and y ⁴vibrate, thus confirm that this furnace outlet temperature data exists the oscillation behavior of two different cycles.

If only do essential time scale to process data y to decompose, its decomposition result as shown in Figure 5, can find out the subsignal component y belonging to different frequency scope easily from figure ², y ³and y ⁴mixed in together, wherein y ²comprise and belong to originally in y on a small quantity ³half wave component, simultaneously y ³and y ⁴respective components mixes mutually, and this is totally unfavorable to our oscillation test.And decomposition proposed by the invention, extraction, cluster and reconstitution steps can effectively address this problem, Fig. 5 fully illustrates the necessity of content proposed by the invention.

In addition, irregular spike disturbance ubiquity in industrial process, when this disturbance has higher magnitude, the decomposition level of essential time scale is probably caused to increase emptily situation, as shown in Figure 6, in 1 ~ 2800Samples, do simple essential time scale to process data y decompose, obtaining decomposing subsignal is { y ¹, y ², y ³, y ⁴, y ⁵.System does not produce the oscillation behavior of new level, but the irregular disturbing influence normal decomposable process of process signal, cause unnecessary subsignal y ⁵appearance.

According to way proposed by the invention, when wicket is updated to 2201 ~ 2400Samples, although this wicket essence time scale decomposition result obtains 5 straton signals, when doing clustering processing in conjunction with current all the other half-wave voltage signal of large window, its optimum cluster number K ^*=K _old=3 (y ¹for noise level) remain unchanged, all half-waves still cluster are also recombinated in three levels, as shown in Figure 7.Method proposed by the invention has good anti-interference, and this is also the embodiment of robustness in the present invention.

Fig. 8 is detection system process proposed by the invention 3000 Samples recon signal { y ², y ³, y ⁴corresponding monitoring statistic result.Persistent oscillation composition y ⁴corresponding monitoring statistic is steady in a long-term more than threshold value Ω=3, simultaneously interrupted oscillation composition y ³occurred twice during the course respectively, our monitoring statistic also real-time follow-up has arrived composition y ³situation of change.

Utilize the inventive method, quantitatively can detect the interrupted oscillation behavior of process control loops, obtaining the multicycle vibrates the regular degree of each oscillating component (as compositions such as multiple vibration, interrupted oscillations) and cycle.For the evaluation of oscillation behavior and source of trouble diagnosis provide abundant Data support.

Claims (10)

1. an online test method for process control loops interrupted oscillation, is characterized in that, comprises the steps:

(1) at control loop to be detected, gather large windows history data in advance, described historical data is decomposed by essential time scale, half-wave information extraction, based on after the Robust clustering of frequency and signal restructuring process, obtains the initial estimation of each frequency content in former data;

(2) in control loop to be detected, after online real time collecting wicket process data is placed on former large window data, large window slides backward, and forms a new large window data with the wicket process data of Real-time Collection;

(3) the wicket data of Real-time Collection are decomposed through essential time scale, after half-wave information extraction, do Robust clustering process in conjunction with remaining half-wave information in existing large window and recombinate, the recon signal obtained is the real-time estimation of each frequency content in the large window data after renewal;

(4) calculate the monitoring statistic corresponding to each subsignal, judge whether each recon signal is in oscillatory regime, the judged result of comprehensive each recon signal, is the real-time testing result of current large window data;

(5) repeat step (2) ~ step (4), get final product the operation conditions of real-time follow-up control loop to be detected.

2. the online test method of process control loops interrupted oscillation according to claim 1, it is characterized in that, in step (1), the collection method of large windows history data is: in each sampling interval, record the process data in control loop to be detected, and the data collected in each sampling interval are added on the process data end previously gathered, until collected data meet prespecified large window size.

3. the online test method of process control loops interrupted oscillation according to claim 1, it is characterized in that, in step (1) and step (3), original signal obtains the decomposition subsignal of different frequency after essential time scale is decomposed, and wherein the stop condition of essential time scale decomposition method is the index of oscillation I < 0.7 of residual components.

4. the online test method of process control loops interrupted oscillation according to claim 1, it is characterized in that, in step (1) and step (3), described half-wave is the signal segment that each layer decomposes in subsignal between two continuous zero cross points; Half-wave information extraction comprises: reject the subsignal representing noise; Collect residue and decompose the wavelength of all half-waves in subsignal and initial, end time thereof.

5. the online test method of process control loops interrupted oscillation according to claim 4, is characterized in that, the basis for estimation of noise signal is statistic A, and its computing formula is:

$A=\frac{C_k}{N},$

Wherein C _kfor decomposing the zero cross point number of subsignal autocorrelation function, N is hits; If statistic A is greater than boundary value A _lim, then think that this decomposition subsignal is noise signal; Described A _limvalue be 0.3 ~ 0.5.

6. the online test method of process control loops interrupted oscillation according to claim 1, it is characterized in that, in step (1) and step (3), based on the K means clustering algorithm of Robust clustering for being individuality with extracted half-wave, being feature with the wavelength of each half-wave of frequency;

In step (1):

After optimum cluster number k equals cancelling noise signal, residue decomposes the number of subsignal;

Initial cluster center P _ifor each layer of correspondence decomposes the median of the half-wave wavelength that subsignal comprises ${\mathrm{\&mu;}}_T^k=\mathrm{median}\left(T\right);$

In step (3):

Existing window cluster structures to be upgraded before cluster, comprise optimum cluster number K ^*with the renewal of initial cluster center;

Optimum cluster number K ^*for:

K ^*＝argmax{S(K _old-1)，S(K _old)，S(K _old+1)}，

Wherein S (*) is clustering result quality index-isolation, K _oldcomprise by former large window the cluster number of data;

Initial cluster center, selects according to current optimum cluster number.

7. the online test method of process control loops interrupted oscillation according to claim 6, is characterized in that, in step (3), the method according to current optimum cluster number selection class initial cluster center is:

If K ^*=K _old, cluster number does not change, and class initial center is C _new=C _old;

If K ^*=K _old+ 1, have more a class in existing large window procedure data, then such initial cluster center is chosen as:

$C_{\mathrm{add}}\&LeftArrow;\underset{1\&le;j\&le;n}{\arg \max }\left\{\underset{1\&le;i\&le;m}{\min }\right\{\frac{\mathrm{dis}(c_i,c_j)}{\mathrm{rad}\left(c_i\right)}\left\}\right\};$

If K ^*=K _old-1, existing large window procedure data have lacked a class, then should leave out a class in former cluster structures:

$C_{\mathrm{del}}\&LeftArrow;\underset{1\&le;i\&le;n}{\arg \max }\left\{\underset{1\&le;j\&le;m}{\min }\right\{\frac{\mathrm{dis}(c_j,c_i)}{\mathrm{rad}\left(c_j\right)}\left\}\right\};$

Wherein, m is the optimum cluster number of former large window data, C _ifor corresponding cluster centre; N is the cluster number of wicket data, c _jfor corresponding cluster centre; Dis (c _i, c _j) be Euclidean distance between two cluster centres; Rad (c _i) be class c _icorresponding class radius.

8. the online test method of process control loops interrupted oscillation according to claim 1, it is characterized in that, in step (1) and step (3), the method obtaining recon signal is: after the process of K mean cluster, will belong in of a sort all half-wave computing with words to same layer signal and obtain recon signal.

9. the online test method of process control loops interrupted oscillation according to claim 1, is characterized in that, in step (4), and monitoring statistic computing formula be:

${\widehat{\mathrm{\&eta;}}}^k=\frac{\sqrt{N-1}}{\sqrt{{\mathrm{\&chi;}}_{N-\mathrm{1,1}-\mathrm{\&alpha;}/2}^2}}\frac{{\mathrm{\&mu;}}_T^k}{{\mathrm{\&sigma;}}_T^k},$

Wherein, N is the data length of this recon signal, when be degree of confidence being 1-α, degree of freedom is that the card side of N-1 distributes critical value, for the median of half-wave wavelength in this recon signal, for utilizing Q _nthe robustness variance that algorithm for estimating calculates.

10. the online test method of process control loops interrupted oscillation according to claim 1, is characterized in that, described threshold value Ω is 3.