532
J. Kim and J.-M. Kim
Fig. 46.3 The picture of the experimental setup
46.4 Experimental Result and Discussion
To validate the performance of Grad-CAM for bearing fault diagnosis, AE signals
from healthy-state and three types of bearing fault are acquired using the testbed The
length of a measured AE signal is 1-s with 1 MHz sampling rate, and the number
of AE signals for each condition is 600, half of the dataset is selected for training
DNN, randomly. Remaining half of the dataset are used for validation. The trained
DNN achieved 99% accuracy of classification with the validation dataset as shown
in Fig. 46.4.
Figure 46.5 shows the importance weight over frequencies from Grad-CAM for
the trained DNN. As shown, DNN learned that the harmonics of defect frequencies
are important information to classify the FBC to each condition. In this experiment,
Fig. 46.4 Confusion matrix
of classification result