Rotor imbalance fault classification through unified feature fusion: Combining statistical signal processing with adaptive deep learning features

The most important element of efficient fault classification lies in the identification of relevant features that can serve as representations for various fault categories. For time series classification, the utilization of solely basic Statistical Features (SFs) does not yield high accuracy in classification. On the other hand, the exclusive dependence on machine learning-based feature extraction, without the incorporation of prior knowledge, compromise the model’s ability to generalize. To tackle these concerns, in this work, a hybrid methodology has been proposed, wherein the utilization of basic SFs is combined with Deep Learning (DL) model’s capability to facilitate comprehensive feature extraction, while simultaneously upholding the robustness of the model. A modified DL architecture based on the second generation wavelet decomposition theory is designed to extract Adaptive Latent Features (ALFs) from rotational time series for the purpose of classifying rotor imbalance faults. The validity of this approach is confirmed by assessing its practicality and effectiveness over a benchmark as well as on a custom dataset obtained from an experimental test-bed for rotor structural faults classification. The experiment demonstrates that the proposed method achieves 99.63% accuracy on the custom dataset, outperforming Support Vector Machine (56.89%), Extreme Gradient Boosting (72.38%), and Lifting Net (94.93%). The validation on the benchmark datasets highlights the effectiveness of the proposed method in distinguishing between different types of mechanical faults, with less than 1.5% of misclassification.