How does TC-CNN maintain high fault diagnosis performance even when the ratio of healthy to faulty samples reaches 50:1, and what happens to other methods under such imbalance?

TC-CNN keeps high fault diagnosis performance at a 50:1 healthy-to-fault ratio by combining FFT-based frequency spectra and GST-based time-frequency images in two parallel CNN branches. Fusing these complementary feature sets enriches fault-related information, making the minority fault samples easier to identify even when very few are available. Other methods degrade sharply under this imbalance, producing low precision and F1 scores because they cannot learn enough from the small number of faulty training samples.

When the training set contains roughly one faulty sample for every 50 healthy samples, TC-CNN still reports a diagnostic accuracy around 97.80% and an F1 score around 0.98, only slightly lower than its balanced-dataset result. This stability comes from its architecture: one channel uses FFT to extract frequency-domain features, and the other uses GST to extract time-frequency image features. The two channels run in parallel through convolutional networks, and their extracted features are fused at a feature layer, so fault characteristics are supplemented from different dimensions and become more distinguishable despite data scarcity. By contrast, conventional methods such as DBN, single-channel CNN variants, and some hybrid approaches suffer severe performance loss as imbalance grows. They tend to produce much lower accuracy and F1 scores, typically falling below 53% accuracy and below 0.5 F1 in the 50:1 case, because their models cannot be trained adequately to recognize the rare fault samples. DBN in particular cannot represent fault information well and its training becomes increasingly difficult with fewer samples. Thus, although all models feel the effects of extreme imbalance, TC-CNN remains relatively stable because FFT and GST allow it to extract more useful fault information from fewer samples than other methods.