Method for estimating and reducing undertainties in process measurements

• US 20040054507A1
• Filed: 09/24/2003
• Published: 03/18/2004
• Est. Priority Date: 10/26/2000
• Status: Active Grant

• Current Assignee: Triant Technologies, Inc. (PDF Solutions)
• Original Assignee: Triant Technologies, Inc. (PDF Solutions)
• Inventors: Mott, Jack Edward
• Granted Patent: US 7,016,816 B2
• Time in Patent Office: Days
• US Class Current: 702/196
• CPC Class Codes: G05B 17/02 electric

Abstract

A reference matrix contains valid measurements characterizing operation of a multivariate process (220). Modeling parameters of the reference matrix are derived (222-232). The final model parameters, balanced with respect to measuring and modeling uncertainties (234, 202), are applied to model (204) a new set of measurements (200). If the new set has no faults (206) then all modeled values and modeling uncertainties (208) can be used to control the process (218). If the new set has only one fault (210) then the modeled value and modeling uncertainty of the faulted measurement plus the measured values and measuring uncertainties of the unfaulted measurements (212) can be used to control the process (218) while repair procedures are implemented for the identified fault (216). If the new set has more than one fault (214) then the process (218) should be shut down, and repair procedures should be implemented (216) for all identified faults.

Claims

1. A method of controlling an m variable multivariate system or process, characterized by:

(a) obtaining (220) historical reference data comprising numerical measurements of said m variables, said measurements collectively encompassing a plurality of variations within one or more operational states of said system or process;

(b) modeling (222) said historical reference data to produce modeled values, modeling uncertainties and measuring uncertainties of said reference data set measurements;

(c) deriving (224-230) a final model of said historical reference data by sequentially repeating said modeling until successively derived sums of all of said measurement uncertainties are approximately equal and successively derived sums of all of said modeling uncertainties are approximately equal;

(d) deriving (232) a set of modeling parameters characteristic of said final model; and

,

(e) combining said measuring uncertainties with a new data set (200) comprising numerical measurements of said m variables, said new data set measurements collectively encompassing said plurality of variations within said one or more operational states of said system or process to control (218) continued operation of said multivariate system or process.

2. A method as defined in claim 1, further characterized by:

(a) determining (232) whether said modeling uncertainties exceed said measurement uncertainties;

(b) for said modeling uncertainties that do not exceed said measurement uncertainties;

(i) determining (206) whether there are no faults in said new data set measurements; and

,

(ii) if there are no faults in said new data set measurements, applying (208) all of said modeled values of said reference data set measurements and all of said modeled values of said new data set measurements to control (218) said multivariate system or process.

3. A method as defined in claim 1, further characterized by:

(a) determining (232) whether said modeling uncertainties exceed said measurement uncertainties;

(b) for said modeling uncertainties that do not exceed said measurement uncertainties;

(i) determining (206) whether there are no faults in said new data set measurements; and

,

(ii) if there are no faults in said new data set measurements, applying (208) all of said modeling uncertainties of said reference data set measurements to control (218) said multivariate system or process.

4. A method as defined in claim 1, further characterized by:

(a) determining (232) whether said modeling uncertainties exceed said measurement uncertainties;

(b) for said modeling uncertainties that do not exceed said measurement uncertainties;

(i) determining (210) whether there is a fault in only one of said new data set measurements; and

,

(ii) if there is a fault in only one of said new data set measurements, applying (212) said modeled value of said faulted one of said new data set measurements to control (218) said multivariate system or process.

5. A method as defined in claim 1, further characterized by:

(a) determining (232) whether said modeling uncertainties exceed said measurement uncertainties;

(b) for said modeling uncertainties that do not exceed said measurement uncertainties;

(i) determining (210) whether there is a fault in only one of said new data set measurements; and

,

(ii) if there is a fault in only one of said new data set measurements, applying (212) said modeling uncertainty of said faulted one of said new data set measurements to control (218) said multivariate system or process.

6. A method as defined in claim 1, further characterized by:

(a) determining (232) whether said modeling uncertainties exceed said measurement uncertainties;

(b) for said modeling uncertainties that do not exceed said measurement uncertainties;

(i) determining (210, 214) whether there is a fault in more than one of said new data set measurements; and

,

(ii) if there is a fault in more than one of said new data set measurements, determining the identity of each faulted one of said more than one of said new data set measurements; and

,

(iii) applying (214) said identity of said each faulted one of said more than one of said new data set measurements said modeling uncertainties of all of said new data set measurements to repair (216) said multivariate system or process.

7. A method as defined in claim 1, further characterized by:

(a) determining (232) whether said modeling uncertainties exceed said measurement uncertainties;

(b) for said modeling uncertainties that do not exceed said measurement uncertainties;

(i) determining (210) whether there is a fault in only one of said new data set measurements;

(ii) if there is a fault in only one of said new data set measurements;

(1) determining the identity of said faulted one of said new data set measurements; and

,

(2) applying (212) said identity of said faulted one of said new data set measurements to control (218) said multivariate system or process.

8. A method as defined in claim 1, further characterized by:

(a) determining (232) whether said modeling uncertainties exceed said measurement uncertainties;

(b) for said modeling uncertainties that do not exceed said measurement uncertainties;

(i) determining (210) whether there is a fault in only one of said new data set measurements;

(ii) if there is a fault in only one of said new data set measurements;

(1) determining the identity of said faulted one of said new data set measurements; and

,

(2) applying (212) said identity of said faulted one of said new data set measurements to repair (216) said multivariate system or process.

9. A method as defined in any one of claims 1, 2, 3, 4, 5, 6, 7 or 8, wherein said multivariate system or process is any physical process.

10. A method as defined in any one of claims 1, 2, 3, 4, 5, 6, 7 or 8, wherein said multivariate system or process is any informational process.

11. A method as defined in any one of claims 1, 2, 3, 4, 5, 6, 7 or 8, wherein said multivariate system or process is any manufacturing, operational, or informational process.

12. A method as defined in any one of claims 1, 2, 3, 4, 5, 6, 7 or 8, wherein said multivariate system or process is any one of an aircraft, aeronautics, biology, chemical, electric power, food product, genetic, metals, oil refining, pharmaceutical, plastic, pulp and paper, semiconductor, econometric, financial, investment, loan, or psychometric process.

13. A method as defined in claim 1, further characterized by before said modeling of said historical reference data to produce said modeled values of said reference data set measurements, comparing (32) said reference data set measurements to a corresponding set of predefined criteria to determine whether said reference data set measurements adequately characterize variations within said reference data set measurements.

14. A method as defined in claim 1, further characterized by, during said deriving of said final model, comparing (32) said sums of all of said measurement and modeling uncertainties to a corresponding set of predefined criteria to determine whether said reference data set measurements adequately characterize variations within said reference data set measurements.

15. A method as defined in claim 1, further characterized by, during said deriving of said final model, comparing (94, 96, 98) said sums of all of said measurement and modeling uncertainties to a corresponding set of predefined criteria to determine whether said sums of all of said measurement and modeling uncertainties are normally distributed.

16. A method as defined in claim 1, further characterized by, during said applying of said modeling parameters to said new data set, comparing (146, 148, 152) said new data set measurements to a corresponding set of predefined criteria to determine whether said new data set measurements adequately characterize variations within said new data set measurements.

17. A method as defined in claim 16, further characterized by, after determining that said new data set measurements adequately characterize said variations within said new data set measurements, associating all of said measurement uncertainties with all of said new data set measurements.

18. A method as defined in claim 17, further characterized by comparing (174, 180) said new data set measurements to said corresponding set of predefined criteria to determine whether said new data set measurements contain any faults.

19. A method as defined in claim 17, further characterized by comparing (174, 180) said new data set measurements to said corresponding set of predefined criteria to determine (192) whether said new data set measurements contains one and only one fault and to identify that one of said new data set measurements which is faulted.

20. A method as defined in claim 17, further characterized by comparing (174) said new data set measurements to said corresponding set of predefined criteria to determine (176) whether said new data set measurements contains more than one fault and to identify each one of said new data set measurements which is faulted.

21. A method as defined in claim 18, further characterized by, if said new data set measurements contain no faults, replacing said new data set measurements with said modeled values of said new data set measurements.

22. A method as defined in claim 18, further characterized by, if said new data set measurements contain no faults, replacing said measurement uncertainties with said modeling uncertainties.

23. A method defined in claim 19, further characterized by, if said new data set measurements contain one and only one fault replacing said faulted one of said new data set measurements with said modeled one of said new data set measurements corresponding to said faulted one of said new data set measurements.

24. A method defined in claim 19, further characterized by, if said new data set measurements contain one and only one fault replacing said measurement uncertainty of said faulted one of said new data set measurements with said modeling uncertainty of said faulted one of said new data set measurements.

25. A method of modeling an ii, variable multivariate system or process, said method characterized by:

(a) forming (10) an m row by n column matrix X having;

(i) n column vectors X_j respectively comprising numerical measurements of said m variables, said measurements collectively encompassing a plurality of variations within one or more operational states of said system or process, each one of said column vectors having elements x_ij where i=1 to m and j=1 to n;

(ii) m row vectors _iX, each said row vector _iX comprising elements x_ij having a range _ranx_i and an average _avex_i; and

,

(b) deriving (12) a similarity x_ij#x_ik=max(0, 1, −

|x_ij−

x_ik|/w_i) between a j^th example of an i^th one of said measurements and a k^th example of said i^th one of said measurements, where w_i is a weighting factor having an initial value equal to _ranx_i.

26. A method as defined in claim 25, further characterized by:

(a) selecting (20) an element x_ab common to both a row vector _aX and a column vector X_b of said matrix X;

(b) removing (22) said element x_ab from said column vector X_b to produce an m−

1 element column vector X′

_b;

(c) removing (24) said row vector _aX from said matrix X to produce an m−

1 row by n column matrix X′

;

(d) deriving (26), for all j=1 to n, similarities X′

_j^T#X′

_b between said column vector X′

_b and a column vector X′

_j in said matrix X′

, wherein;

X′

_j^T#X′

_b=(x_1j#x_1b+x_2j#x_2b+ . . . +x<sub>n−

1j</sub>#x<sub>a−

1b</sub>+x_a+1j#x_a+1b+ . . . +x_mj#x_mb)/(m−

1);

(e) selecting (28) the largest n′

of said n similarities, wherein 2<

n′

<

m and wherein each one of said n′

selected similarities is not equal to unity and not equal to zero; and

,

(f) determining (30, 32) that said matrix X inadequately characterizes said system or process in at least one of said operational states if said n′

similarities do not exist.

27. A method as defined in claim 26, further characterized by:

(a) transforming (38, 40) said matrix X′

into an m−

1 row by n′

column matrix X″

by removing from said matrix X′

all of said column vectors X′

_j lacking said n′

similarities;

(b) deriving (42), for all j=1 to n′ and

all k=1 to n′

, similarities X″

_j^T#X′

_k between said column vector X′

_j and said column vector X′

_k in said matrix X″

wherein;

X″

_j^T#X″

_k=(x_1j#x_1k+x_2j#x_2k+ . . . +x<sub>a−

1j</sub>#x<sub>a−

1k</sub>+x_a+1j#x_a+1k+ . . . +x_mj#x_mk)/(m−

1);

(c) deriving (44), for all j=1 to n′

, similarities X″

_j^T#X′

_b between said column vector X″

_j in said matrix X″ and

said column vector X′

_b, wherein;

X″

_j^T#X′

_b=(x_1j#x_1b+x_2j#x_2b+ . . . +x<sub>a−

1j</sub>#x<sub>a−

1b</sub>+x_a+1j#x_a+1b+ . . . +x_mj#x_mb)/(m−

1);

(d) deriving (46) an n′

element column vector C=(X″

^T#X″

)<sup>−

1</sup>(X″

^T#X′

_b), wherein;

(i) X″

^T#X″

is a square matrix of n′

² similarities X″

_j^T#X″

_k;

(ii) X″

^T#X′

_b is a column vector of n′

similarities X″

_j^T#X′

_b;

(iii) said n′

elements of said column vector C have a sum c_ab;

(iv) n′

squares of elements of said column vector C have a sum c²_ab;

(e) defining (48) a model y_ab of said element x_ab, wherein;

y_ab=_aXC+(1−

c_ab)_avex_a;

(f) producing (50) pluralities y_ij, c_ij, and c²_ij for said models y_ab said sum c_ab, and said sum c²_ab for all i=1 to m and for all j=1 to n;

(g) assembling (52) said plurality y_ij of models into an m row by n column model matrix Y comprising;

(i) n column vectors Y_j each one of said column vectors Y_j having elements y_ij;

(ii) m row vectors _iY, each said row vector _iY comprising elements y_ij having an average _avey_i;

(h) deriving (54) similarities x_ij#y_ij between corresponding elements in said matrix X and said matrix Y, wherein;

x_ij#y_ij=max(0, 1|x_ij−

y_ij|/w_i);

(i) deriving (56) average similarities s_i between elements in corresponding rows of said matrix X and said matrix Y for each one of said rows, wherein;

s_i=((x_i1#y_i1)+(x_i2#y_i2)+ . . . +(x_in#y_in))/n;

(j) deriving (64) root-mean-square deviations t_i between elements in corresponding rows of said matrix X and said matrix Y for each one of said rows, wherein;

t_i=(((x_i1−

y_i1)²+(x_i2−

y_i2)²+ . . . +(x_in−

y_in)²)/n)^1/2;

(k) defining (66) a quantity M;

for each row of said matrix X wherein;

M_i=_ranx_i/t_i;

(l) selecting (70) an integer M which approximates the largest one of said M_i; and

,

(m) redefining (72) said weighting factor w_i as w_i=Mt_i.

28. A method as defined in claim 27, further characterized by:

(a) sequentially repeating (78, 80, 84) said method with said redefined weighting factor and without reselecting said integer M, until successive values of 1−

s_i differ by less than 10%; and

,

(b) determining (86) that said matrix X contains inaccurate measurements if said successive values of 1−

s_i do not differ by less than 10% after a selected number of sequential repetitions of said method.

29. A method as defined in claim 28, further characterized by:

(a) concluding (92, 94, 98) that differences between said x_ij and said y_ij are approximately normally distributed with a mean of zero if said 1−

s_i values lie within 10% of 1−

_exps, wherein;

_exps=1−

(2/π

)^1/2/M;

(b) determining (100) a limit _minc_i, below which 1% of said column vector sums c_ij lie;

(c) deriving (102) averages _avec²_i of c²_ij;

(d) locating (104) maxima _maxc²_i of c²_ij;

(e) initially assigning (106) measuring uncertainty upper limit values equal to t_i;

(f) initially assigning (108) modeling uncertainty upper limit values equal to t_i;

(g) determining (114, 120) that modeling uncertainty limit values are less than measuring uncertainty limit values if _maxc²_i<

1;

(h) assigning (122) measuring uncertainty lower limits and reassigning (124) said modeling uncertainty upper limits as u_i if _maxc²_i<

1, wherein u_i=t_i(1/2)_1/2; and

,

(i) assigning (126) said modeling uncertainty lower limits as v₁ if _maxc²_i<

1, wherein v_i=t_i(_avec²_i/2)^1/2.

30. A method as defined in claim 29, further characterized by:

(a) acquiring (132) a column vector X having m numerical elements x_i;

(b) selecting (134) an element x_n common to said row vector _aX and to said vector X;

(c) removing (136) said selected element x_a from said column vector X to produce an m−

1 element column vector X′

;

(d) removing (142) said row vector _aX from said matrix X to produce an m−

1 row by n column matrix X′

;

(e) defining similarities x_ij#x_i between the said x_ij and said elements x_i wherein;

x_ij#x_i=max(0, 1−

|x_ij−

x_i|/w_i);

(f) deriving (138, 140), for all j=1 to n, similarities X′

_j^T#X′

between a column vector X′

_j in said matrix X′ and

said column vector X′

, wherein;

X′

_j^T#X′

=(x_ij#x₁+x_2j#x₂+ . . . +x<sub>a−

1j</sub>#x<sub>a−

1</sub>+x_a+1j#x_a+1+ . . . +x_mj#x_m)/(m−

1);

(g) selecting (144) the largest said n′

of said n similarities wherein each one of said n′

selected similarities is not equal to unity and not equal to zero;

(h) determining (146, 148) that said vector X is an invalid description of the process characterized by said matrix X if n′

of said similarities X′

_j^T#X′

not equal to zero and not equal to unity cannot be formed;

(i) transforming (154, 156) said matrix X′

into an m−

1 row by n′

column matrix X″

by removing from said matrix X′

all column vectors in said matrix X′

lacking said n′

selected similarities;

(j) deriving (158), for all j=1 to n′ and

all k=1 to n′

, similarities X″

_j^T#X″

_k between said column vector X″

_j and said column vector X″

_k in said matrix X″

wherein;

X″

_j^T#X″

_k=(x_1j#x_1k+x_2j#x_2k+ . . . +x<sub>a−

1j</sub>#x<sub>a−

1k</sub>+x_a+1j#x_a+1k+ . . . +x_mj#x_mk)/(m−

1);

(k) deriving (164), for all j=1 to n′

, similarities X″

_j^T#X′

between said column vector X″

_j in said matrix X″ and

column vector X′

wherein;

X″

_j^T#X′

=(x_1j#x₁+x_2j#x₂+ . . . +x<sub>a−

1j</sub>x<sub>a−

1</sub>+x_a+1j#x_a+1+ . . . +x_mj#x_m)/(m−

1);

(l) deriving (166) a column vector C with n′

elements, wherein;

C=(X″

^T#X″

)<sup>−

1</sup>(X″

^T#X′

) wherein;

(i) X″

^T#X″

is the square matrix of n′

² similarities X″

_j^T#X″

_k;

(ii) X″

^T#X′

is the column vector of n′

similarities X″

_j^T#X′

;

(iii) n′

elements of said column vector C have a sum c_a;

(m) producing (168) a model y_a of element x_a wherein;

y_a=_aXC+(1−

c_a)_avex_a;

(n) producing (170) pluralities y_i and c_i for said model y_a and said sum c_a for all i=1 to m;

(o) collecting (172) said models y;

into a column vector Y;

(p) determining (174, 180, 182) that said vector X contains no faults if more than one of said sums c_i are greater than _minc_i;

(q) associating all said y_i, all said t_i, all said u_i, and all said v_i with all said x_i if said vector X contains no faults;

(r) determining (174, 180, 192) that said vector X contains only one fault if only one of said sums c_i is greater than _minc_i;

(s) associating (186) said y_i, said u_i and said v_i with said x_i for which said sum c_i is greater than _minc_i if said vector X contains only one fault;

(t) associating (196) said t_i and said u_i with said x_i for which said sums c_i are not greater than _minc_i if said vector X contains only one fault; and

,

(u) determining (174, 176) that said vector X contains more than one fault if none of said sums c_i are >

_minc_i.