Neural network based system for equipment surveillance

• US 5,745,382 A
• Filed: 08/31/1995
• Issued: 04/28/1998
• Est. Priority Date: 08/31/1995
• Status: Expired due to Term

• Current Assignee: ARCH Development Corporation (University of Chicago)
• Original Assignee: ARCH Development Corporation (University of Chicago)
• Inventors: Wegerich, Stephan W., Gross, Kenneth C., Vilim, Richard B.
• Primary Examiner(s): Cosimano, Edward R.
• Application Number:
  US08/521,892
• Time in Patent Office: 971 Days
• Field of Search: 128/695 R, 364/550, 364/551.01, 395/21, 395/23, 395/904, 395/906, 395/912, 395/914
• US Class Current: 706/16
• 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
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G21D 3/001 Computer implemented control
• CPC Class Codes: Y02E 30/00 Energy generation of nuclear origin
• CPC Class Codes: Y02E 30/30 Nuclear fission reactors
• CPC Class Codes: Y10S 706/904 Manufacturing or machine, e.g. agricultural machinery, machine tool
• CPC Class Codes: Y10S 706/906 Process plant

Abstract

A method and system for performing surveillance of transient signals of an industrial device to ascertain the operating state. The method and system involves the steps of reading into a memory training data, determining neural network weighting values until achieving target outputs close to the neural network output. If the target outputs are inadequate, wavelet parameters are determined to yield neural network outputs close to the desired set of target outputs and then providing signals characteristic of an industrial process and comparing the neural network output to the industrial process signals to evaluate the operating state of the industrial process.

Claims

1. A method of performing surveillance of transient signals of an industrial device to determine an operating state thereof, comprising the steps of:

(a) reading into a memory training data;

(b) determining neural network weighting values by the steps comprising;

(1) solving a set of linear equations for obtaining the neural network weighting values;

(2) computing a neural network output;

(3) evaluating the neural network output to determine whether the output is close to a set of target outputs;

(4) continuing steps (1)-(3) until achieving the target outputs; and

(5) providing a neural network output;

(c) providing signals characteristic of an industrial process;

(d) comprising the neural network output to said industrial process signals to ascertain the operating state of the industrial process comprising the steps of;

(1) hypothesizing said industrial process signals belong to a particular class;

(2) at least one of translating and scaling stored neural network training signals; and

(3) determining error between the neural network output for said industrial process signals and said stored neural network training signals.

2. The method as defined in claim 1 further includes additional steps after said step (b) when the target outputs are not sufficiently close to the target outputs, said steps comprised of:

commencing iterating on wavelet parameters; and

calculating the neural network output for the wavelet parameters until achieving a desired set of the wavelet parameters such that the network output converges to the desired set of target outputs.

3. The method as defined in claim 1 wherein the network weighting values are adjusted to enable meeting the target outputs by minimizing an error equation given by:

##EQU22## where t_p.sbsb.1 =target value at p^th output mode when input signal 1 is presented to the neural network.

4. The method as defined in claim 1 wherein n potential classes exist and the industrial process signals are tested n times to determine the proper class for the process signals.

5. The method as defined in claim 1 where output values u_p1 of the neural network are given in terms of the weighting values w_kp, and coefficients thereof including a_k, b_k and c_k as follows:

##EQU23## where i₁ (t)=value of input signal 1 at time t.

6. The method as defined in claim 5 wherein a defined best set of w_kp values is found by least squares followed by applying a gradient descent method to determine the values of said a_k, b_k and c_k.

7. The method as defined in claim 1 wherein the industrial process signals comprise biological signals.

8. The method as defined in claim 7 wherein the biological signals comprise heart signals.

9. The method as defined in claim 1 wherein the industrial process signals comprise signals having substantially constant signals either side thereof characteristic of equilibrium signals for the industrial process.

10. The method as defined in claim 9 wherein further including the step of determining onset of equilibrium signals from the industrial process and implementation of a SPRT analysis.

11. The method as defined in claim 1 wherein said step (b) comprises training on translated and scaled versions of the training data.

12. The method as defined in claim 11 wherein the training process includes setting the target outputs lower, thereby establishing a gradient.

13. The method as defined in claim 11 wherein a cartesian coordinate system is established with an original signal at point (0,1), a first axis corresponding to a translate dimension and a second axis corresponding to a scale dimension and said training data establishing a well-shaped response surface, thereby enabling evaluation of a presented signal to determine whether it is a member of the same class as the training data.

14. The method as defined in claim 13 wherein the output of the network is input to a two layer perceptron network thereby providing additional degrees of freedom for enhanced shaping of the response surface.

15. The method as defined in claim 13 wherein error between the training data and the presented signal is least for the class to which it belongs.

16. The method as defined in claim 13 wherein a genetic search method is used to find a global extremum in said response surface.

17. A method of performing surveillance of transient signals of an industrial device to determine an operating state thereof, comprising the steps of:

(a) reading into a memory training data;

(b) determining neural network weighting values by the steps comprising;

(1) solving a set of linear equations for obtaining the neural network weighting values;

(2) computing a neural network output;

(3) evaluating the neural network output to determine whether the output is close to a set of target outputs;

(4) continuing steps (1)-(3) until achieving the target outputs; and

(5) providing a neural network output;

(c) providing signals characteristic of an industrial process;

(d) calculating the neural network output for the wavelet parameters until achieving a desired set of the wavelet parameters which yield the neural network output close to a desired set of target outputs by performing the following steps;

(1) hypothesizing said industrial process signals belong to a particular class;

(2) translating and scaling stored neural network training signals; and

(3) determining error between the neural network output for said industrial process signals and said stored neural network training signals.

(e) providing signals characteristic of an industrial process;

(f) comparing the neural network output to said industrial process signals to ascertain the operating state of the industrial process.

18. The method as defined in claim 17 wherein the industrial process signals comprise transient signals.

19. The method as defined in claim 17 further including the step of annunciating an alarm condition upon detecting a deviation of the industrial process signals from a desired one of the particular class.

20. The method as defined in claim 17 wherein said industrial process signals are selected from the group consisting of engine startup signals, biological pulse signals, chemical process startup signals, power plant startup signals and electronic systems transient signals.