Takanori Hasegawa, Jun Ogata, Masahiro Murakawa, Tetsunori Kobayashi, Tetsuji Ogawa, ``Adaptive training of vibration-based anomaly detector for wind turbine condition monitoring,’’ Proc. Annual Conference on PHM Society, pp.177-184, Oct. 2017.
5. Effective Use of Existing Detector
Target wind turbine Other wind turbine
• Base anomaly detector is built using large‐
scale data from non‐target device.
• Base anomaly detector is adapted to target
device using small amount of data.
5
base model
adaptive
training
target model
large
data
small
data
6. Effective Use of Existing Detector
Target wind turbine Other wind turbine
• Base anomaly detector is built using large‐
scale data from non‐target wind turbine.
• Base anomaly detector is adapted to target
wind turbine using small amount of data.
6
base model
adaptive
training
target model
large
data
small
data
Efficient development of robust
anomaly detector
24. Measuring Anomaly Using GMM
Likelihood of GMM:
: # of Gaussians
Θ , , Σ : parameter set of GMM
: mixture weight for ‐th Gaussian
: mean vector ‐th Gaussian
Σ : full covariance matrix for ‐th Gaussian
Anomaly score:
24
FLAC feature space
Anomaly
• GMM represents normal status of machinery.
• Negative logarithmic likelihood of input for normal status GMM
can measure anomaly of device operating.