Bearing fault classification using TDA-guided feature fusion

Abstract

The application of Topological Data Analysis (TDA) to bearing fault diagnosis is emerging as a new research direction. However, the excessive time consumption of persistent homology in the TDA pipeline poses a new challenge. This paper improves the method from three aspects: the size of the Takens sliding window, the normalization strategy, and the robustness against noise interference. First, each channel of the signal is segmented into non-overlapping windows with a length of 1024. Then, a TDA-guided fused feature representation is constructed by integrating persistent homology statistics, H1 persistence images, and compact amplitude-sensitive signal descriptors. After comparing several classifiers, Random Forest (RF) is selected as the final classification model. Experimental results show that the 2 ws Takens window setting achieves the highest average test accuracy on the dataset under four operating conditions. Ablation experiments with individual TDA settings on the 1730 rpm subset reveal that per-window z-score normalization reduces the test accuracy to 59.75%, whereas preserving the original signal scale yields an accuracy of 88.56%. The final fused model achieves test accuracies of 100.00%, 100.00%, 98.73%, and 99.57% at rotational speeds of 1730, 1750, 1772, and 1797 rpm, respectively, with an average macro F1-score of 99.57%. Additional experiments with Additive White Gaussian Noise (AWGN) verify the robustness of the proposed method: it remains stable under 12 dB noise, maintains usable discriminative performance at 6 dB noise, and experiences a sharp performance drop only at 0 dB noise.