Method of monitoring machine condition

• US 8,426,771 B2
• Filed: 08/19/2010
• Issued: 04/23/2013
• Est. Priority Date: 02/11/2010
• Status: Active Grant

• Current Assignee: Korea Advanced Institute of Science and Technology
• Original Assignee: Korea Advanced Institute of Science and Technology
• Inventors: Hwang, Yoha, Lee, Jong Min
• Primary Examiner(s): Hoang, Huan
• Application Number:
  US12/859,753
• Publication Number:
  US 20110192827A1
• Time in Patent Office: 978 Days
• Field of Search: 219/130.01, 219/130.33
• US Class Current: 219/130.01
• CPC Class Codes: B23K 9/007 Spot arc welding
• CPC Class Codes: B23K 9/0953 using computing means

Abstract

Various embodiments of a method for monitoring a machine condition are provided. An embodiment of the present invention provides a method of monitoring a machine condition, comprising the following steps: modeling a normal signal model performed by detecting a signal for monitoring condition of a normal machine and converting the detected signal to a normal signal model in time domain using a hidden Markov model (HMM) algorithm; calculating a probability value data of the monitoring signal at a subject machine performed by detecting a signal for monitoring condition of the subject machine in real-time and converting the detected signal to the probability value data relative to the normal state signal model using the HMM algorithm; and determining a section having deficiency where the probability value data of the monitoring signal at the subject machinery is not maintained constantly relative to the normal signal model.

Claims

1. A method of monitoring a machine condition, the method comprising:

modeling a normal signal model performed by detecting a signal for monitoring a condition of a normal machine and converting the detected signal to data of the normal signal model in a time domain using a hidden Markov model (HMM) algorithm;

calculating a probability value data of the monitoring signal at a subject machine performed by detecting a signal for monitoring a condition of the subject machine in real-time and converting the detected signal to the probability value data relative to the normal signal model using the HMM algorithm;

determining a section having a deficiency where the probability value of the monitoring signal at the subject machine is not maintained constantly relative to the normal signal model; and

updating the normal signal model to include the probability value data of a section determined to have the deficiency when the deficiency does not occurred at the section.

2. A method of monitoring a welding deficiency, comprising

modeling a normal welding signal model performed by detecting welding signals at a welding portion in a time domain during a welding operation, checking normal welding sections without a deficiency, extracting a normal welding signal data from the normal welding sections, and converting the normal welding signal data to the normal welding signal model of the welding operation using a hidden Markov model (HMM);

calculating probability value data of the welding signal at a subject welding portion relative to the normal welding signal model performed by detecting the welding signal at the subject welding portion in real-time and converting the detected signal at the subject welding portion to the probability value data using the HMM algorithm; and

determining a section having a deficiency where the probability value data of the welding signal at the subject welding portion is not maintained constantly relative to the normal welding signal model.

3. The method of claim 2, wherein the welding signal is a voltage signal.

4. The method of claim 2, wherein the welding signal is a current signal.

5. The method of claim 2, wherein the normal welding sections are checked by at least one of an optical inspection, non-destructive inspection and a section inspection.

6. The method of claim 2, wherein the normal welding signal data is pre-processed to a time function.

7. The method of claim 6, wherein the normal welding signal data is pre-processed by filtering or time weighting.

8. The method of claim 6, wherein the pre-processed normal welding signal data is converted to characteristic vector array by a transformation.

9. The method of claim 8, wherein the characteristic vector array is one or combination of a mean value, a standard deviation, a power spectral coefficient, a filter bank, a cepstral coefficient, a wavelet coefficient and a coefficient of auto-regressive (AR) model.

10. The method of claim 2, wherein the normal welding signal model and the probability value data of the welding signal at the subject welding portion are one of an initial probability, a state transition probability, a symbol observation probability, a state mean vector and a state and covariance matrix, and

wherein the HMM algorithm is one of a forward/backward algorithm, a scaled forward/backward algorithm, a k-means clustering method, a modified k-mean clustering method, an expectation-maximize (EM) algorithm and a Baum-Welch algorithm.

11. The method of claim 2, wherein the welding signal detected at the subject welding portion is pre-processed to a time function.

12. The method of claim 11, wherein the welding signal detected at the subject welding portion is pre-processed by filtering or time weighting.

13. The method of claim 11, wherein the pre-processed welding signal at the subject welding portion is converted to characteristic vector array by a transformation.

14. The method of claim 13, wherein the characteristic vector array is one or combination of a mean value, a standard deviation, a power spectral coefficient, a filter bank, a cepstral coefficient, a wavelet coefficient and a coefficient of auto-regressive (AR) model.

15. The method of claim 2, wherein the normal welding signal model for a welding operation time is obtained by collecting a plurality of normal welding signal models at the normal welding sections extracted from one welding operation.

16. The method of claim 2, wherein the normal welding signal model for a welding operation time is obtained by collecting a plurality of normal welding signal models at the normal welding sections extracted from a plurality of welding operations.

17. The method of claim 2, wherein if the section determined as having the deficiency does not have deficiency, then the normal welding signal model is updated to include the probability value data.

18. A method of monitoring a welding deficiency, comprising:

modeling a normal welding voltage signal model performed by detecting welding voltage signals at a welding portion in a time domain during a welding operation, checking normal welding sections without a deficiency, extracting a normal welding voltage signal data from the normal welding sections, and converting the normal welding voltage signal data to the normal welding voltage signal model of the welding operation using a hidden Markov model (HMM);

calculating probability value data of the welding voltage signal at a subject welding portion relative to the normal welding voltage signal model performed by detecting the welding voltage signal at the subject welding portion in real-time and converting the detected signal at the subject welding portion to the probability value data using the HMM algorithm; and

determining a section having a deficiency where the probability value data of the welding voltage signal at the subject welding portion is not maintained constantly relative to the normal welding voltage signal model,

wherein the normal welding voltage signal data is pre-processed to a time function and the pre-processed normal welding voltage signal data is converted to a characteristic vector array by a transformation, and

wherein the welding voltage signal detected at the subject welding portion is pre-processed to a time function, and the pre-processed welding voltage signal at the subject welding portion is converted to a characteristic vector array by the transformation.

19. A method of monitoring a welding deficiency, comprising:

modeling a normal welding current signal model performed by detecting welding current signals at a welding portion in a time domain during a welding operation, checking normal welding sections without a deficiency, extracting a normal welding current signal data from the normal welding sections, and converting the normal welding current signal data to the normal welding current signal model of the welding operation using a hidden Markov model (HMM);

calculating probability value data of the welding current signal at a subject welding portion relative to the normal welding current signal model performed by detecting the welding current signal at the subject welding portion in real-time and converting the detected signal at the subject welding portion to the probability value data using the HMM algorithm; and

determining a section having a deficiency where the probability value data of the welding current signal at the subject welding portion is not maintained constantly relative to the normal welding current signal model,

wherein the normal welding current signal data is pre-processed to a time function and the pre-processed normal welding current signal data is converted to a characteristic vector array by a transformation, and

wherein the welding current signal detected at the subject welding portion is pre-processed to a time function, and the pre-processed welding current signal at the subject welding portion is converted to a characteristic vector array by the transformation.