Engine bearings are among the most critical components in rotating machinery, and their failure can often lead to the entire system stopping. Therefore, accurate and reliable monitoring of bearings is crucial for predictive maintenance and fault prevention. However, noisy signals, variable operating conditions, and the simultaneous occurrence of multiple fault types make it difficult to achieve high accuracy in fault diagnosis. This study proposes a new hybrid approach to overcome these challenges. The proposed method consists of five main stages. In the first stage, Empirical Mode Decomposition (EMD) based preprocessing is applied to remove noise and unwanted components from the vibration signals. In the second stage, feature extraction is performed from the processed signals using two different methods: Detrended Fluctuation Analysis (DFA) captures the fractal dynamics and statistical structure of the signals, while a Long Short-Term Memory (LSTM) network learns the complex time-dependent relationships of the signals. In the third stage, these obtained features are combined to obtain a more meaningful representation. In the fourth stage, the K-Means clustering method is used to group the features to increase the accuracy of the classifiers. In the final stage, three powerful classification algorithms, namely XGBoost, Random Forest, and Support Vector Machines (SVM), are used to classify bearing faults. The HUST engine bearing dataset used in this study enabled us to evaluate the performance of our method under realistic conditions because it contained data collected under different loads, speeds, and fault types. Experimental analyses showed that the proposed hybrid structure exhibited superior performance compared to single feature sets. In particular, the combined use of DFA and LSTM-based features increased the generalizability and diagnostic power of the system. Furthermore, classification accuracy was further improved thanks to the feature vectors supported by EMD preprocessing and K-Means clustering. The obtained results show that the best performance was achieved when DFA + LSTM features were used together with EMD and K-Means and with the XGBoost classifier. In this case, 98.46% accuracy, 97.20% sensitivity, 97.24% precision, and 97.27% F1-score values were achieved. This study demonstrates that the proposed hybrid approach yields reliable results under different operating conditions and in data sets containing multiple fault types. The proposed method provides a robust and generalizable solution for predictive maintenance applications in industrial systems, making significant contributions to the safety and availability of rotating machinery.
Er et al. (Tue,) studied this question.