Search

EP-4419875-B1 - AUTONOMOUS DISCRIMINATION OF OPERATION VIBRATION SIGNALS

EP4419875B1EP 4419875 B1EP4419875 B1EP 4419875B1EP-4419875-B1

Inventors

  • BALASUBRAMANIAN, HARIHARAN
  • LIU, YIXIU
  • GERDES, Matthew, T.
  • WANG, Guang, C.
  • GROSS, KENNY, C.

Dates

Publication Date
20260513
Application Date
20221010

Claims (12)

  1. A computer-implemented method for autonomous discrimination of operation vibration signals, the method comprising: partitioning (310, 610) a frequency spectrum of output into a plurality of discrete frequency bins, wherein the output is collected from one or more vibration sensors (220, 230, 232, 233) monitoring a reference device (231); generating (315, 640-675) a representative time series signal for each frequency bin, wherein said generating (315, 640-675) comprises: obtaining time series signals for each frequency in said plurality of discrete frequency bins, based on data scanned or recorded from the reference device (231) while the reference device (231) is operated in a deterministic stress load test pattern (405); identifying (665), from said obtained time series signals, a set of highly-intercorrelated frequencies in each of said plurality of discrete frequency bins, wherein said representative time series signal is generated from an average of time series values for the identified set of highly-intercorrelated frequencies; generating (320) a power spectral density for each frequency bin by converting each representative time series signal from the time domain to the frequency domain; determining (325) a maximum power spectral density value for each frequency bin and a peak frequency value for each frequency bin at which the maximum power spectral density value occurs; selecting (330) a subset of the frequency bins that have maximum power spectral density values exceeding a first threshold; assigning (335) the representative time series signals from the selected subset of frequency bins as operation vibration signals indicative of operational load on the reference device (231); and configuring (340) a machine learning model based on at least the operation vibration signals.
  2. The computer-implemented method of claim 1, further comprising: computing (640) correlation coefficients between each frequency in a first frequency bin and each other frequency in the first frequency bin; computing (650) an intercorrelation score for each frequency in the first frequency bin from the correlation coefficients for the frequency; and identifying (665) said set of highly-intercorrelated frequencies in the first frequency bin, wherein the highly-intercorrelated frequencies have intercorrelation scores exceeding a second threshold.
  3. The computer-implemented method of claim 2, wherein the second threshold is having an intercorrelation score in at least a 90th percentile of the intercorrelation scores for the first frequency bin.
  4. The computer-implemented method of claim 1 or claim 2 or claim 3, wherein the deterministic stress load test pattern is one of a sinusoidal stress load, a stress load regularly cycling between on and off states, a stress load regularly cycling between high and low states, and a stress load regularly cycling between high and idle states.
  5. The computer-implemented method of claim 1 or claim 2 or claim 3, wherein the first threshold is having a maximum power spectral density value in at least an 80 th percentile of the maximum power spectral density values of all frequency bins.
  6. The computer-implemented method of any of claims 1 to 5, wherein the machine learning model is a non-linear non-parametric regression model.
  7. The computer-implemented method of any of claims 1 to 5, wherein the machine learning model is a multivariate state estimation technique model.
  8. The computer-implemented method of any of claims 1 to 5, wherein configuring the machine learning model further comprises training the machine learning model with the operation vibration signals collected during a training period of normal operation of the reference device (231); and wherein the method of further comprises: monitoring operation vibration signals of a target device with the trained model during a surveillance period of operation to detect vibration anomalies; and transmitting a predictive maintenance alert in response to detection of a vibration anomaly in one or more of the operation vibration signals of the target device.
  9. A non-transitory computer-readable medium that includes stored thereon computer-executable instructions for autonomous discrimination of operation vibration signals that when executed by at least a processor (1310) of a computing system cause the computing system to perform operations which include the operations recited in any one of claims 1 to 7.
  10. The non-transitory computer-readable medium of claim 9, wherein the instructions to configure the machine learning model further cause the computing system to train the machine learning model with the operation vibration signals collected during a training period of normal operation of the reference device; and wherein the instructions further cause the computing system to: monitor operation vibration signals of a target device with the trained model during a surveillance period of operation to detect vibration anomalies; and transmit a predictive maintenance alert in response to detection of a vibration anomaly in one or more of the operation vibration signals of the target device.
  11. A computing system comprising: a processor (1310); a memory (1315) operably connected to the processor (1310); a non-transitory computer-readable medium operably connected to the processor (1310) and memory (1315) and storing computer-executable instructions for autonomous discrimination of operation vibration signals that when executed by at least a processor (1310) of the computing system cause the computing system to perform operations which include the operations recited in any one of claims 1 to 7.
  12. The computing system of claim 11, wherein the instructions to configure the machine learning model further cause the computing system to train the machine learning model with the operation vibration signals collected during a training period of normal operation of the reference device; and wherein the instructions further cause the computing system to: monitor operation vibration signals of a target device with the trained model during a surveillance period of operation to detect vibration anomalies; and transmit a predictive maintenance alert in response to detection of a vibration anomaly in one or more of the operation vibration signals of the target device.

Description

BACKGROUND Accelerometers (vibration sensors) may be affixed to mechanical devices and used to detect problems in the devices that result in higher vibration levels, which can be precursors to failure. Present techniques for vibration-based condition monitoring of mechanical devices are rudimentary. For example, one approach for deriving diagnostic and prognostic information from accelerometer sensors is to place thresholds on the gross vibration levels, so that if the vibrations increase above the threshold, an alarm condition is actuated to raise a prognostic warning. Thresholds on vibration amplitudes are most appropriate for machines that have a constant load and run at a fixed speed for the life of the system, and for constant-load machines that happen to be in an environment with a stationary ambient vibration level, that is, where there are no other vibrating components that may provide a variable ambient vibration level. But, fixed workload components that run at fixed rotations per minute (RPMs) for life and are not mechanically coupled to any other components that add to the ambient vibration background are very rare. For rotating machinery with dynamic (that is, changing over time) workloads, variable-speed performance, and mounted into structures containing other varying vibration sources, thresholds on gross vibrational amplitudes are very inefficient for detecting the early onset of degradation. The inefficiencies are a consequence of the threshold boundaries. Threshold alarm boundaries have to be set higher than the highest peak for the component at its highest load, highest RPM setting, and when the ambient vibration levels are highest. This significantly lowers the early warning potential for prognostics because vibration due to incipient failure may not reach the threshold until significant degradation has occurred, and may never reach the threshold at all before total failure. The paper by Melih C. Yesilli et al., "On transfer learning for chatter detection in turning using wavelet packet transform and ensemble empirical mode decomposition", CIRP Journal of Manufacturing Science and Technology, vol. 28, 2019, pages 118-135, discloses the use of wavelet packet transform with recursive feature elimination for chatter detection in metal cutting. A time series is decomposed into wavelet packets. Informative wavelet packets, which are best suited to distinguish between stable cutting and chattering motion, are selected. This involves reconstructing a time domain signal for each wavelet packet and obtaining the corresponding FFT for each of the reconstructed signals. Both frequency domain as well as time domain features for chatter identification are extracted from the reconstructed signal from the informative wavelet packet. Different classifiers are trained, using recursive feature elimination. SUMMARY The present invention is defined by the independent claims. The dependent claims concern optional features of some embodiments of the invention. For the purpose of determining the extent of protection, due account shall be taken of any element which is equivalent to an element specified in the claims. In one embodiment, a computer-implemented method for autonomous discrimination of operation vibration signals is provided. The computer-implemented method includes partitioning a frequency spectrum of output into a plurality of discrete frequency bins (also referred to hereinbelow as "bins"). The output is collected from one or more vibration sensors monitoring a reference device. The computer-implemented method also includes generating a representative time series signal for each frequency bin. This generating step includes a sub step of obtaining time series signals for each frequency in said plurality of discrete frequency bins, based on data scanned or recorded from the reference device while the reference device is operated in a deterministic stress load test pattern, as well as a sub step of identifying, from said obtained time series signals, a set of highly-intercorrelated frequencies in each of said plurality of discrete frequency bins. The representative time series signal is generated from an average of time series values for the identified set of highly-intercorrelated frequencies. The computer-implemented method also includes generating a power spectral density for each bin by converting each representative time series signal from the time domain to the frequency domain. The computer-implemented method also includes determining a maximum power spectral density value for each bin and a peak frequency value for each bin at which the maximum power spectral density value occurs. The computer-implemented method also includes selecting a subset of the bins that have maximum power spectral density values exceeding a first threshold. The computer-implemented method also includes assigning the representative time series signals from the selected subset of bins as operation vibration signals in