Radial basis function neural network autoassociator and method for induction motor monitoring

• US 5,574,387 A
• Filed: 06/30/1994
• Issued: 11/12/1996
• Est. Priority Date: 06/30/1994
• Status: Expired due to Fees

• Current Assignee: Siemens Corporate Research Incorporated (Siemens AG)
• Original Assignee: Siemens Corporate Research Incorporated (Siemens AG)
• Inventors: Petsche, Thomas, Garrett, Charles
• Primary Examiner(s): Wieder, Kenneth A.
• Assistant Examiner(s): BOWSER, BARRY
• Application Number:
  US08/269,465
• Time in Patent Office: 866 Days
• Field of Search: 324/772, 324/72.5, 73/116, 361/93, 361/94, 361/87, 361/31, 361/96, 318/800, 318/473, 318/806, 364/508, 364/513, 395/22
• US Class Current: 706/18
• CPC Class Codes: G01R 31/343 in operation
• CPC Class Codes: G06F 18/2414 Smoothing the distance, e.g. radial basis function networks [RBFN]
• CPC Class Codes: G06F 2218/12 Classification; Matching
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06N 3/065 Analogue means
• CPC Class Codes: H02H 1/0092 concerning the data processing means, e.g. expert systems, neural networks
• CPC Class Codes: H02H 7/085 against excessive load H02H6/00 takes precedence
• CPC Class Codes: Y10S 706/911 Nonmedical diagnostics
• CPC Class Codes: Y10S 706/912 Manufacturing or machine, e.g. agricultural machinery, machine tool

Abstract

A method for detecting a departure from normal operation of an electric motor comprises obtaining a set of normal current measurements for a motor being monitored; forming clusters of the normal current measurements; training a neural network auto-associator using the set of normal current measurements; making current measurements for the motor in operation; comparing the input and output of the auto-associator; and indicating abnormal operation whenever the current measurements deviate more than a predetermined amount from the normal current measurements. The method models a set of normal current measurements for the motor being monitored, and indicates a potential failure whenever measurements from the motor deviate significantly from a model. The model takes the form of an neural network auto-associator which is "trained"--using clusters of current measurements collected while the motor is known to be in a normal operating condition--to reproduce the inputs on the output. A new set of FFT'"'"'s of current measurements are classified as "good" or "bad" by first transforming the measurement using a Fast Fourier Transform (FFT) and an internal scaling procedure, and then applying a subset of the transformed measurements as inputs to the neural network auto-associator. A decision is generated based on the difference between the input and output of the network.

Claims

1. Apparatus for detecting a departure from normal operation of an electric motor, comprising:

sensing means for measuring a set of current values for a motor being monitored;

first processing means coupled to said sensing means for deriving frequency spectral components associated with said set of current values;

a neural network auto-associator coupled to said signal processing means and further coupled to said sensing means for receiving at least one of at least a portion of said spectral components and at least a portion of said current values as an input vector and having output terminals for providing an output vector, said neural network having undergone a training phase; and

second processing means coupled to said output terminals for comparing said input and output vectors for providing an error metric, wherein said first processing means clusters said frequency spectral components into a smaller number of clusters during said training phase.

2. Apparatus in accordance with claim 1, wherein said first processing means performs a Fast Fourier Transform (FFT).

3. Apparatus in accordance with claim 1, wherein said current values correspond to instantaneous values of current.

4. Apparatus in accordance with claim 1, wherein said clustering is performed by use of a K-means clustering algorithm.

5. Apparatus in accordance with claim 1, wherein said first processing means scales said frequency spectral components in accordance with predetermined weights.

6. Apparatus in accordance with claim 2, wherein said input vector comprises a subset of said frequency spectral components.

7. Apparatus in accordance with claim 2, wherein said second processing means provides a set of error metrics and includes means for evaluating said set of error metrics for providing a determination in accordance with predetermined criteria of whether motor operation is normal or not.

8. Apparatus in accordance with claim 1, wherein neural network includes an autoassociator which is trained using a set of current values obtained from a motor known to be operating normally.

9. Apparatus in accordance with claim 1, wherein neural network includes radial basis function means for processing said input vector.

10. Apparatus in accordance with claim 1, wherein neural network includes a hidden layer.

13. Apparatus in accordance with claim 1, wherein K equals one of 4096 and 16384.

14. Apparatus in accordance with claim 1, wherein each said frequency spectral component is scaled by the inverse of the magnitude of the largest of said frequency spectral components, whereby relative magnitudes of said frequency spectral components remain constant and in the range of magnitudes of 0-1.

17. Apparatus in accordance with claim 13, wherein said hidden layer of said neural network is defined by ##EQU14## where Y_i (i=0, . . . , L) is the ith output node of the network;

H is the number of nodes in the hidden layer;

w_i,j is a weight on the output of hidden unit j when applied as input to output unit i;

Θ

_i is the threshold for output unit i;

Θ

_i⁰ is the threshold for hidden unit i;

c_i is the center of hidden unit j and σ

_j is its width; and

σ

(u) = (1+e^-u)^-l where e is the base of the natural logarithm.

18. Apparatus in accordance with claim 14, wherein, given values for x_i and x_i , said second processing means computes said error metric as d, where ##EQU15##

11. Apparatus for detecting a departure from normal operation of an electric motor, comprising:

sensing means for measuring a set of current values for a motor being monitored;

first processing means coupled to said sensing means for deriving frequency spectral components associated with said set of current values in accordance with a Fast Fourier Transform (FFT), for scaling said frequency spectral components in accordance with predetermined weights;

a neural network auto-associator coupled to said signal processing means and further coupled to said sensing means for receiving at least one of a selected portion of said frequency spectral components and a portion of said current values as an input vector and having output terminals for providing an output vector, and including a hidden layer, said autoassociator having been trained in a training phase, using a set of current values obtained from a motor known to be operating normally, said first processing means having clustered said frequency spectral components into a smaller number of clusters during said training phase; and

second processing means coupled to said output terminals for comparing said input and output vectors for providing an error metric.

12. Apparatus in accordance with claim 11 wherein said values are collected beginning at some time arbitrarily called 0 (zero) and at time K-1, the K measures c₀, . . . , c_K-1 are transformed using said (FFT) to produce an estimate of magnitudes of the frequency spectral components.

15. Apparatus in accordance with claim 11, wherein Φ

is defined as the i^th component of said frequency spectral components (FFT components), and for each FFT component, said first processing means provides a quantity x_i, where ##EQU13## where τ

is a user defined index set that selects some subset of the components of the FFT for defining said portion of said frequency spectral components, τ

containing L elements, said L elements being normalized with respect to the largest and smallest elements thereof, a resulting set of L scaled magnitude components x₀, . . . , x_L-1, being defined to represent the maximum RMS value of said current values being monitored.

16. Apparatus in accordance with claim 12, wherein for each FFT component, said first processing means provides a vector X, where

X=x₀, . . . , x_L, whereby vector X corresponds to L scaled components of the FFT plus the scaled RMS current, scaling being in accordance with respective scaling factors.

19. Apparatus in accordance with claim 15, wherein said second processing means repeats the foregoing procedure a total of T times to generate T error metrics d₁, . . . d_T. The system then indicates that the motor is "good" if ##EQU16## and that it is bad otherwise.