Method and device for signal analysis, process identification and monitoring of a technical process

• US 5,748,508 A
• Filed: 06/23/1995
• Issued: 05/05/1998
• Est. Priority Date: 12/23/1992
• Status: Expired due to Fees

• Current Assignee: Michael-Alin Baleanu
• Original Assignee: Michael-Alin Baleanu
• Inventors: Baleanu, Michael-Alin
• Primary Examiner(s): Teska, Kevin J.
• Assistant Examiner(s): Walker, Tyrone V.
• Application Number:
  US08/494,594
• Time in Patent Office: 1,047 Days
• Field of Search: 364/578, 364/554, 364/571.02, 364/721, 364/724.17, 364/736, 364/481, 364/151, 395/2.09, 395/3
• US Class Current: 703/13
• CPC Class Codes: G01H 1/003 of rotating machines G01H1/10 takes precedence
• CPC Class Codes: G01M 7/00 Vibration-testing of structures; Shock-testing of structures G01M9/00 takes precedence
• CPC Class Codes: G07C 3/00 Registering or indicating the condition or the working of machines or other apparatus, other than vehicles

Abstract

Method and device for modelling the variables relevant to a process as a function of other parameters describing or influencing the process, termed regressors, by means of multiple regression for the purpose of process identification, monitoring, analysis and control or regulation. The classical method of stepwise multiple regression is expanded by the introduction of the so-called collinearity cone into a recursive method yielding all "best" collinearity-free regression models. The method is completed by giving consideration to the regression errors and by restriction to the absolutely necessary matrix elements. Stable regression models of various sizes are thus produced with little expenditure of time. Further, either linear or nonlinear regression functions permit a more accurate process analysis or modelling. By automatic learning in the case of newly occurring combinations of regressive values, it is also possible to apply the process to process monitoring, control and regulation.

Claims

1. Device for carrying out consistent, expanded, fast and collinearity-free multiple regression in recursive form, for the purpose of the analysis or control of a technical process, comprising:

a processor (2) calculating the best Q regression models from data in a memory for the correlation matrix (123) which can be obtained via the unit for correlation calculation (132) from the memory (122) for the covariance matrix, and the memory for the regressor variances (115) entered or calculated beforehand, which are utilized for correction of the diagonal elements of the correlation matrix, and

that the regression models are classified in terms of their residue variance by the regression unit (102) of the processor (2), which is bus-controlled by the control and addressing unit (101), the residual variance of the model, regressor numbers and coefficients being deposited in the memory (116) provided for that purpose, and

that a locking memory (110), a memory for the diagonal elements and the dependent column of the regression matrix (111), a memory for the recursion matrix (113) with the offset-memory (112) belonging to the recursion matrix and a memory for the regression numbers (114) are provided for the realization of the recursive method, and

that the regression unit (102) must perform several actions for the recursion level p, including

that first of all those regressors are determined which make a small contribution to the model, and if necessary the regressors are locked (110), the offset for the recursion matrix is stored (112), the regressor rows of the new recursion matrix (113) and the new diagonal elements and the dependent column (111) are calculated, and the residual regressor numbers (114) stored,

that those regressors, located within the collinearity cone of the current model, are determined and locked (110) with the help of the data from the memory for the regressor variances (115),

that the most significant regressor from those remaining is searched, that regressor is selected which yields the model with the smallest residual variance and at the same time passes the F test known from statistics, and

that, in the event of an unsuccessful search, the residual variance of the current model is utilized for the classification of this model between the best Qs, with the help of the residual variances of the models (116) already found, in which case, if the residual variance is smaller or equal to the largest stored residual variance, or currently less than Q models were found, the residual variance, the regression numbers and coefficients are stored in memory for the best models (116), and thereafter the regression unit (102) returns to the recursion level p-1, and its context is loaded from the corresponding memories (110 through

     114),

that, in the event of a successful search for a significant regressor, the group of regressors within the collinearity cone of the best regressor is found with the values of the regressor variances (115),

that each of these regressors is incorporated into the model individually, one after the other, while the remaining ones from the collinear group remain locked (110), and with the precondition that none of the regression coefficients already found undergoes a sign change, the rows of the recursion matrix (113) for the new regressor and the new diagonal elements and the dependent column (111) are calculated and the recursion level p+1 selected,

that, after processing of the collinear regressors, the change is made to the recursion level p-1, its context being loaded from the corresponding memories (110 through

     114).

2. Circuit arrangement according to claim 1 wherein several regression units with associated memories operate in parallel.

3. Circuit arrangement according to claim 1 wherein several regression units with associated memory operate in parallel in several planes, the number of the regression units in the lower planes can increase.

4. Device according to claim 1, for monitoring a technical process or of oscillations of a machine or components of the same, especially machines with rotating parts, with the use of sensors for the measurement of monitoring variables, and of parameters which describe or influence the loading state and relative positions of the components of the monitored machine, and by modelling the parameter behavior of the monitoring variables by means of consistent, expanded, fast and collinearity-free multiple regression in recursive form, characterized by the fact

that a measurement recording unit (A) is provided for the detection of the monitoring variables and a process capture system (B) for detection of the parameters,

that a bus-controlled monitoring unit (C) is connected to the computer (5) along with the measurement recording unit (A) and the process detection system (B),

that the monitoring unit (C) exhibits a control and state analysis unit (1), which, with the aid of the data in the operating-state control memory (11), has the following tasks:

a) Checking the operating state;

in case of normal state, go to step b), otherwise it triggers an alarm and stores it in the alarm memory (51),

b) if new, not yet learned operating states occur, it initiates formation of a new model in the regression unit (2), called a learning phase, if no alarms are present in the alarm memory (51), and then it updates the operating-state control memory (11),

c) initiates the formation of a control model in the regression control unit (3), if an already learned operating state is present,

that the new Q models are calculated according to the CEFSR process in the regression unit (2), and the coefficient and function-number memory (116) is updated, whereby the associated function generator (131) makes available a large number N, with N much larger than L, of additional, freely selectable, linear and nonlinear functions from the data of the L input channels, and the associated covariance matrix (122) and correlation matrix (123) memories are updated,

that the formation of the control model is carried out according to the CEFSR method in the regression controller unit (3), employing the parameter functions from the coefficient and function number memory (116) determined in the regression unit (2), as well as the current measurement data, whereby a function generator (131) identical with that in the regression unit (2), is used for calculation of the values in the control-covariance matrix memory (31), and thereafter the model coefficients and function numbers obtained are stored in the ring memory (32) for the purpose of detecting trend developments in the Q models,

that first of all the Q estimates of the monitor variables, or estimate of a process model obtained by utilization of the redundancy of the Q regression model, are calculated in the predictor and comparator unit (4), then the residues of the monitoring variables or the contributions of the individual parameters are formed and, with the fulfillment of various alarm criteria, such as threshold crossings, as well as model changes, alarm messages are formed, which are stored in the alarm memory (51) and employed for failure or process diagnosis,

and that the computer (5) processes the visualization of the residues and model changes with the help of the data in the ring memory (32) as well as the processing or forwarding of the alarms present in the alarm memory (51).

5. Circuit arrangement according to claims 1, 2 or 3, characterized by the fact

that, from the memory (121) of the input matrix U with L rows corresponding to the number of the input channels and V columns corresponding to the size of the sample, a number of additional, freely selectable, linear and nonlinear functions are made available as new repressors by the function generator (131), and thus a preliminary matrix X is generated with N rows, N being much larger than L, and the same column number, with which the covariance matrix XX'"'"' is calculated, and

that the determined Q regression models, or a process model, obtained by utilizing the redundancy of the Q regression models, is employed for the control or analysis of a process.

6. Device according to claim 4, characterized by the fact

that the residue of the monitoring variable is fed to an adjustment device, capable of changing one or more processing input values, for the purpose of regulating the process.

7. Device according to one of the claims 1, 2, 3 or 4, characterized by the fact that the performance characteristics of the circuit arrangements are realized on a computer or in parallel-computer architecture.

8. Method for the analysis or control of a technical process with the aid of regression models, which are calculated in recursive form by means of consistent, expanded, fast and collinearity-free multiple regression, in short CEFSR, for the variables relevant to the process, as a function of other parameters describing or influencing the process, termed regressors, characterized by the fact

that a number of supplementary, freely selectable, linear and nonlinear functions of the regressors are calculated from the entry matrix U with L rows corresponding to the number of input channels and V columns corresponding to the size of the sample, and thus a preliminary matrix X is generated with N rows, with N much greater than L, and the same number of columns as U, which is used to calculate the covariance matrix XX'"'"',

that the correlation matrix is calculated from the covariance matrix and the diagonal elements are corrected with the help of the individual regressor variances,

that a regression model for the variable declared as dependent is calculated with the covariance matrix by means of a recursive method containing the following steps for one recursion level,

that first of all, the regressors are determined, which make a small contribution to the model, in order to remove them from the model and lock them for the further recursion levels, in which case, if there are such regressors, the recursion matrix as well as the diagonal elements and the independent column of the regression matrix must be calculated anew for recursion level p,

that the regressors located inside the collinearity cone of a regression space are locked,

that the most significant regressor from the remaining ones is found by means of the F test,

that, in the event of an unsuccessful search, the residual variance is used by the current model for the classification of this model between the best Qs, with the aid of the residual variances of the models already found, in which case, if the residual variance is smaller or equal to the largest stored residual variance, or at present fewer than Q models were found, the residual variance, the regressor numbers and coefficients are stored, and where, as the next step, the recursion level p-1 is selected,

that, when the search is successful, the set of regressors is determined, which lie within the collinearity cone of the best regressor,

that each of these regressors is incorporated into the model individually, in sequence, while the others, belonging to the collinear set, are locked and, with the precondition that none of the regression coefficients already determined has undergone a sign change, the rows of the recursion matrix for the new regressor and the diagonal elements and the independent column is calculated and the recursion level p+1 selected,

that, after the processing of the collinear regressors, a change is made to recursion level p-1.

9. Method according to claim 8 for monitoring a technical process or the vibrations of a machine or structural parts of the same, especially a machine with rotating parts, with the use of sensors for the measurement of monitoring variables and of parameters describing or influencing the process, as well as of a measurement-data processing unit, which makes the values of the monitoring variables and of the parameters, called regressors, simultaneously available in digital form to a computer, characterized by the fact

that an individual model formation is begun anew with the occurrence of new operating states for the respective monitoring variables after checking for a normal state and a checking for completeness in the regressor space, if no alarms are present,

that, in the case of a new operating state, and updated covariance matrix, thus containing learning values, is formed from the current measurement values of the monitoring variables and the either linear or nonlinear parameter functions produced by the function generator, new regression models are determined according to CEFSR,

that, in the case of an already learned operating state, a control model computation is carried out by means of the same method, in which case the significant parameter functions, already determined during the learning phase, are taken into account,

that at least one estimate of the monitoring variables is derived from the parameter measurements with the coefficients or parameter functions determined, of one or more of the available Q regression models, or of a process model obtained by use of the redundancy of the Q regression models, and the at least one estimate is subtracted from the actual measured value of the monitored variable and

that the resulting residues are taken into consideration as new monitoring variables, especially for alarm purposes and

that new or leaving parameter functions or contribution changes of the old ones are used to form alarm messages and for failure or process diagnosis.

10. Method according to one of the claims 8 or 9, characterized by the fact,

that the residue of the monitoring variable are passed to a adjustment device which can change one or more process input variables for the purpose of regulating the process.