Why does TC-CNN outperform CWT+2D-CNN in fault diagnosis, especially under high imbalance ratios?

TC-CNN outperforms CWT+2D-CNN under high imbalance because it fuses FFT-based frequency features and GST-based time-frequency features from two channels. This complementary, multi-dimensional feature set lets TC-CNN extract enough fault information from very few fault samples, so its precision and F1 stay nearly stable even as the imbalance ratio rises from 6:1 to 51:1, whereas CWT+2D-CNN's performance drops sharply as fault sample counts shrink.

According to the chapter, TC-CNN integrates frequency spectrum information from FFT and time-frequency images from GST through two parallel CNN channels, and these differently dimensioned features complement one another. This enriches the fault characteristic data and makes fault samples easier to recognize, which is especially useful when the number of fault samples is very small. When the normal-to-fault ratio is low, for example 3:1, TC-CNN does not significantly improve over CWT+2D-CNN because enough fault samples are available and the feature-rich advantage is not critical. However, as the imbalance ratio increases from 6:1 up to 51:1, TC-CNN's fault diagnostic performance remains relatively unchanged, while existing approaches, including CWT+2D-CNN, degrade dramatically because their models cannot be trained well when fault data decreases with increasing imbalance. The reason given in the chapter is that TC-CNN can derive more fault information from fewer fault samples through its complementary FFT and GST features, reducing the impact of data imbalance. Even though all models suffer from extremely imbalanced datasets, TC-CNN is more stable and resilient.