To facilitate distinguishing between normal operating sounds and sounds indicative of wear, Toshiba has developed VAE-DE, an AI technology that uses a neural network, which is an AI computational model that mimics the structure of the human brain. This AI learns the characteristics of the operating sounds of equipment and combines high performance in detecting minute changes in operating sounds with robustness against ambient noise. This AI uses a variational autoencoder (VAE) network, a deep-learning method that automatically finds and learns data features, using unique criteria designed to separate normal sounds from sounds indicative of wear (Fig. 1). Conventional VAE rarely produces false detections even when normal sounds are mixed with noise, but there is still a risk that weak signs of wear will be overlooked as normal sounds (Fig. 2). The newly developed AI is based on a VAE method that is less prone to false positives due to noise and performs learning based on new proprietary criteria that separate normal sounds from sounds indicative of wear, so that only those faint sounds indicative of wear are judged as being outside the normal range of sounds.
Cooling fans were used to verify the effectiveness of this AI. Cooling fans are installed in many devices and are essential for their proper operation. These fans constantly rotate at high speeds, so they often deteriorate faster than other parts of the device they are installed in. However, the rate of deterioration greatly varies depending on the individual fan and the conditions under which the device is used. Condition-based maintenance is therefore desirable, so that cooling fans can be replaced before they fail and stop working.
By first recording and analyzing sounds from operating cooling fans, that was in actual use in power facilities for over several years, it was confirmed that deterioration trends occur probabilistically and can be detected as a slight increase in high-frequency sounds that are inaudible to humans. Next, the developed AI was applied to simulated data that reproduced sounds indicating wear, together with electrical noise generated from power equipment. Regarding the correlation between the deterioration trend, which indicates the actual state of equipment wear, and the “estimated wear value,” which is the value estimated from the operating sounds, the correlation coefficient was greatly improved from 0.144 to 0.905, indicating an extremely strong correlation. This confirms that the developed AI can infer the presence of wear with high accuracy while suppressing false positives due to noise, a feat that has been difficult to achieve in the past (Fig. 3).
The development of this AI will aid in the realization of effective condition-based maintenance from the onset of wear, the timing of which differs by device, by capturing early signs of equipment abnormalities or wear while minimizing false positives due to ambient noise.