Automatic control loop monitoring and diagnostics

• US 5,838,561 A
• Filed: 09/23/1996
• Issued: 11/17/1998
• Est. Priority Date: 04/29/1996
• Status: Expired due to Term

• Current Assignee: FPInnovations
• Original Assignee: Pulp & Paper Research Institute of Canada
• Inventors: Owen, James Gareth
• Primary Examiner(s): RUGGIERO, JOSEPH F
• Application Number:
  US08/722,920
• Time in Patent Office: 785 Days
• Field of Search: 364/148, 364/152, 364/156, 364/178, 364/179, 364/176, 364/551.01, 364/552, 364/554, 364/468.15, 364/468.16, 364/468.17, 364/582
• US Class Current: 700/32
• CPC Class Codes: G05B 19/0428 Safety, monitoring G05B19/0423 takes precedence
• CPC Class Codes: G05B 23/0235 based on a comparison with predetermined threshold or range, e.g. "classical methods", carried out during normal operation; threshold adaptation or choice; when or how to compare with the threshold
• CPC Class Codes: G05B 23/024 Quantitative history assessment, e.g. mathematical relationships between available data; Functions therefor; Principal component analysis [PCA]; Partial least square [PLS]; Statistical classifiers, e.g. Bayesian networks, linear regression or correlation analysis; Neural networks
• CPC Class Codes: G06F 2218/12 Classification; Matching

Abstract

A method of diagnosing a malfunction of a process control system which includes at least one closed loop control loop comprising measuring a histogram of tracking error of the control loop, determining distortion of the tracking error relative to a Gaussian distribution, and indicating a malfunction in the process in the event a deviation from the Gaussian distribution of the tracking error exceeds predetermined limits.

Claims

1. A method of diagnosing a malfunction of a process control system which includes at least one closed loop control loop comprising measuring a histogram of tracking error of said control loop, determining distortion of said tracking error relative to a Gaussian distribution, and indicating a malfunction in the process in the event a deviation from said Gaussian distribution of said tracking error exceeds predetermined limits, wherein said distortion (K) is measured by subtracting from a height of a tracking error histogram bar of said histogram centered on zero, a number of samples multiplied by an area between a pair of limits defining a normal density about a mean of said histogram, and then indicating a malfunction in the process in the event a value of K is different from 0 by a predetermined amount.

2. A method as defined in claim 1 in which the malfunctioning indicating step is carried out by:

(a) if a sequence of observed tracking errors is x(k), where the index k ranges from 1 (the start of the sequence) to n (the end of the sequence), generate a new sequence y(k) by the following rule;

if x(k) is within the center bar of the histogram, then y(k);

=1, otherwise y(k);

=0, then subtract the sample mean of y from each element y(k),

(b) compute an autocorrelation function for the sequence y(k) up to a fixed lag N, wherein an autocorrelation function sequence is R(j) where j ranges from -N to +N,

(c) determine a variance using the transform ##EQU7## where R(.) is the estimated autocorrelation sequence, n is the number of samples,

S=the sum of the sequence y(k) prior to removal of the sample mean in (a), which is the height of the center bar of the histogram, and the expected value of S is equal to the area under the Gaussian curve in the range of the center histogram bar multiplied by the number of samples,

(d) determine a confidence value C, ##EQU8## χ

=absolute value of the difference between the expected value of S and its computed value divided by the estimated standard deviation of S, and

(e) set confidence bands and determine whether C is contained within the bands.

3. A method of automatic assessment of control loop performance of an industrial machine comprising:

(a) collecting operating data comprising time series of controlled variable measurements and control loop set points simultaneously from predetermined control loops, for a period of at least approximately 100 times a longest time constant of said predetermined control loops,

(b) subtracting measured variable data from set point data to obtain tracking errors,

(c) determining an amount by which observed variance of a tracking error exceeds a minimum value, after non-linear elements have been removed from a loop, exploiting prior estimates of process time constant and dead-time to provide a raw index,

(d) testing for any interactions between control loops which may be inflating in an abnormal manner an estimate of said raw index,

(e) determining a modified raw index for a particular loop in the event said inflated estimates are detected, and

(f) distinguishing between control loops that are malfunctioning, those that are not malfunctioning, and those that are possibly malfunctioning and are perturbed by interacting malfunctioning control loops, based on said raw index and said modified raw index.

4. A method as defined in claim 3 further comprising:

(i) computing a histogram of the tracking error for each potentially malfunctioning loop,

(ii) quantifying kurtosis of each histogram by determining height of a center bar of said histogram relative to an expected height wherein the expected height is determined by an assumption that tracking error is normally distributed with a sample mean and sample variance,

(iii) calculating a statistical significance of said kurtosis in a downward direction, taking into account any inter-sample correlation of tracking error time-series, and

(iv) displaying said statistical significance as a diagnostic measure indication.

5. A method as defined in claim 3, in which said raw index (IN) is determined by observing tracking error variation (O) in said loop, determining a variance of conditional expectation (V) from measurements and samples in the past, and processing O/V to obtain IN.

6. A method as defined in claim 5, wherein V is determined by:

(a) subtracting a measured variable from a set point to obtain a tracking error,

(b) extracting prior measured variable data for said loop,

(c) in the event a sample time T of said variable is at least approximately 0.1 times the dominant open loop time constant of said loop, finding best least squares approximation of a vector of tracking error observations with linear combinations of tracking errors and successive tracking errors delayed by successive sample period delays, and

(d) in the event a sample time of said variable is smaller than approximately 0.1 times said dominant loop time constant, resample said variable at a sample interval approximately 5-10 times shorter than said dominant open loop time constant, and then find the best least squares approximation as in step (c).

7. A method as defined in claim 6 in which the step of calculating said raw performance index is comprised of determining an amount by which observed variance of a tracking error exceeds a minimum value, after non-linear elements have been removed from a loop, from prior estimates of process delayed time constant.

8. A method as defined in claim 6 in which step (d) is comprised of digital antialias filtering said variable and resampling said variable at a longer sample interval I which is approximately 5-10 times shorter than said open loop time constant, and expressing said longer sample intervals as said sample intervals.

9. A method as defined in claim 8 in which a cut off frequency of the antialiasing filter is between 10-20 times the reciprocal of the open loop time constant in radians per sample, and in which the resampling interval is approximately 0.1-0.2 times an estimated process time constant.

10. A method as defined in claim 6 including the step of determining the existence of drift in tracking error using a statistic D wherein D=maximum of |m1|, |m3|) wherein m1 and m3 are means of the first third and third third of a time series of said tracking error, and providing a warning signal if D>

standard deviation of the tracking error x (12M/L)^1/2, where m is a predetermined large number and L is the number of samples in the time series.

17. A method as defined in claim 7 including determining a category of malfunction of a process carried out by said industrial machine comprising:

(a) tracking error variations of narrow spectral bandwidth in each of plural control loops of said process,

(b) comparing spectral peaks of said error variations to detect coincidences of peaks which are indicative of interaction between said plural control loops, and

(c) quantifying effects of said error variations which have said coincidences of peaks, and as a result determining malfunctioning of a control loop.

18. A method as defined in claim 17 performed on each control loop deemed to be malfunctioning,

in which said error variation tracking step is comprised of evaluating a Fourier transform of said tracking error variations, windowing products of the Fourier transformation, choosing a Daniel window having bandwidth W periodogram ordinates such that W is a smallest integer which satisfies Wexp^W >

L,

and in which the spectral peak comparing step is comprised of estimating the frequency f1 of a first maximum of the estimated power spectrum, evaluating a center frequency f2 of any center peak by testing for a second maximum over frequencies excluded from an interval around the first maximum, evaluating a variance associated with said primary and secondary maxima by computing an area under the power spectrum estimate over an interval of fixed bandwidth about said frequencies f1 and f2, and indicating the presence of a spectral resonance at a corresponding frequency f1 or f2 in the event either variance exceeds predetermined proportions of overall tracking error variance.

19. A method as defined in claim 18 including the step of forming classes of control loops by associating control loops in a class, wherein said frequencies f1 and f2 of a control loop are adjacent to either said f1 or f2 of another control loop by a small predetermined amount.

20. A method as defined in claim 19 including the steps of determining a modified index of performance as a determination of malfunction for control loops belonging to either a class of potentially interacting control loops or a class of control loops in which a resonance is identified to be outside a predetermined range of wavelengths which resonance is caused by other than a limit cycle generated due to the presence of a severe nonlearity in a control loop.

21. A method as defined in claim 20 in which said determining step is comprised of processing the signal values:

(modified index)=(raw index)×

(1-proportion of the tracking error variance associated with a resonance suspected of being imposed externally to a control loop).

11. A method of automatic assessment of control loop performance of and industrial machine comprising:

(a) identifying a current control loop in a group of control loops,

(b) obtaining operating data and prior dynamic information for said control loop,

(c) calculating a raw performance index for said control loop,

(d) indicating said current control loop as potentially malfunctioning in the event said raw performance index is greater than a predetermined threshold,

(e) in the event said control loop is indicated as potentially malfunctioning, computing a fast Fourier transform of a tracking error, and filter products of said transform to remove spurious peaks,

(f) identifying primary and secondary spectral peaks contributing more than a threshold variance in a predetermined bandwidth for said control loop,

(g) selecting another control loop in said group of control loops and repeat steps (a)-(g) until a last control loop in said group has been processed,

(h) divide potentially malfunctioning loops with approximately coincident spectral peaks into possibly interacting classes,

(i) determine a modified performance index for all control loops belonging to a class, and

(j) apply a histogram test to spectral peaks of all control loops in a class to determine a category of malfunction.

12. A method as defined in claim 11 including determining a drift of said operating data, and indicating from a value or a trend of said drift whether an upset condition exists, and in the event an upset condition exists, providing a warning indication.

13. A method as defined in claim 12 including carrying out the step of determining the presence of said drift by determining a statistic D for the drift tracking error of said loop, wherein D=maximum of |m1|, |m3|), wherein m1 and m3 are means of a first third and third third of a time series of tracking error, and indicating an upset condition if D>

standard deviation of the tracking error x (12M/L)^1/2 where M is a predetermined large number and L is the number of samples in the tracking error time eries.

14. A method of determining a category of malfunction of a process comprising:

(a) tracking error variations of narrow spectral bandwidth in each of plural control loops of said process,

(b) comparing spectral peaks of said error variations to detect coincidences of peaks which are indicative of interaction between said plural control loops, and

(c) quantifying effects of said error variations which have said coincidences of peaks, and as a result determining malfunctioning of a control loop.

15. A method as defined in claim 14 performed on each control loop deemed to be malfunctioning,

in which said error variation tracking step is comprised of evaluating a Fourier transform of said tracking error variations, windowing products of the Fourier transformation, choosing a Daniel window having bandwidth W periodogram ordinates such that W is a smallest integer which satisfies Wexp^W >

L,

and in which the spectral peak comparing step is comprised of estimating the frequency f1 of a first maximum of the estimated power spectrum, evaluating a center frequency f2 of any center peak by testing for a second maximum over frequencies excluded from an interval around the first maximum, evaluating a variance associated with said primary and secondary maxima by computing an area under the power spectrum estimate over an interval of fixed bandwidth about said frequencies f1 and f2, and indicating the presence of a spectral resonance at a corresponding frequency f1 or f2 in the event either variance exceeds predetermined proportions of overall tracking error variance.

16. A method as defined in claim 15 including the step of forming classes of control loops by associating control loops in a class, wherein said frequencies f1 and f2 of a control loop are adjacent to either said f1 or f2 of another control loop by a small predetermined amount.