| CPC G01M 7/025 (2013.01) | 4 Claims |
|
1. A method for vibroacoustic analysis implemented by at least one processor, the method comprising:
extracting feature values from time-series vibroacoustic data measured in equipment in operation;
accumulating feature values corresponding to a state in which the equipment operates normally and general measurement data on the equipment;
constructing through machine learning a probability distribution model using the feature values corresponding to the state in which the equipment operates normally and the general measurement data;
feeding the probability distribution model with the general measurement data on the equipment and with the feature values;
calculating a degree of anomaly based on the probability distribution model;
issuing an alarm if the degree of anomaly is not within a normal range;
pre-calculating a damaged-case dataset of degrees of anomaly for each location in the equipment that correspond to cases where the location is damaged;
comparing the damaged-case dataset with an actual measurement dataset of degrees of anomaly calculated from general measurement data on the equipment to obtain a degree of similarity for each location in the equipment;
outputting, on charts, the locations, in order ranked by the degrees of similarity for each location in the equipment, as locations having high likelihoods of being damaged, an abscissa of each of the charts representing a frequency level and an ordinate of each of the charts representing a degree of anomaly; and
implementing maintenance of the equipment based upon the locations having the high likelihoods of being damaged so as to prevent secondary damage of the equipment,
wherein the extracting includes:
extracting an effective value, a peak value, a largest peak difference, skewness, kurtosis, an average crest factor, and an absolute average amplitude as feature values from measured acceleration in the time-series vibroacoustic data measured in the equipment in operation;
performing spectral analysis on the time-series data to extract an order and a peak value of a maximum peak, an order and a peak value of a second peak, an order and a peak value of a third peak, an order and a peak value of a fourth peak, . . . and an order and a peak value of a n-th peak as the feature values for the data obtained;
performing a wavelet transform on the time-series data to decompose the time-series data into scale levels corresponding to a plurality of frequency bands and obtain power time-series data for each of the scale levels;
extracting an effective value WP1, a largest value WP2, a highest crest factor WP3, and an absolute average value WP4 as feature values for each of the frequency bands from the power time-series data obtained for each of the scale levels, wherein
WP1=√Σi=1nxi2/n,
WP2=max(xi),
WP3=WP2/WP1,
WP4=Σi=1n|xi−x|/x; and
performing spectral analysis on the power time-series data obtained for each of the scale levels to extract an order and a peak value of a maximum peak, an order and a peak value of a second peak, an order and a peak value of a third peak, an order and a peak value of a fourth peak, . . . and an order and a peak value of an n-th peak as the feature values for the data obtained.
|