Hybrid IGWO-DNN approach for bearing fault diagnosis using time domain features from empirical mode decomposition

This study proposes a hybrid diagnostic framework that integrates Empirical Mode Decomposition (EMD) with a Deep Neural Network (DNN) whose architecture and learning configuration are optimized using the Improved Grey Wolf Optimizer (IGWO) for rolling bearing fault classification. To ensure computational efficiency and interpretability, the method relies exclusively on time-domain characterization. Vibration signals are decomposed by EMD into intrinsic mode functions (IMFs), and a compact set of statistical descriptors (including root mean square, kurtosis, peak-to-peak value, and skewness) is extracted from both the original signals and the IMFs to capture fault-related signatures without requiring predefined time–frequency bases. These features are then fed into an IGWO optimized multilayer perceptron, where IGWO automatically selects key design variables such as the number of hidden layers, neurons, activation function, and training solver. The proposed approach is evaluated on experimental data representing three bearing conditions (healthy, inner-race fault, and outer-race fault). The optimized IGWO-DNN achieves high diagnostic performance, reaching 98% test accuracy and perfect class-balanced test metrics (macro-precision/recall/F1/specificity = 1.000). Additional benchmarking highlights solver sensitivity and confirms the benefit of the IGWO-selected quasi-Newton training, retraining the same optimized architecture using Adam and SGDM yields noticeably lower class-balanced performance, while a multiclass AdaBoostM2 ensemble baseline remains significantly less accurate on the test set. ROC–AUC analysis further confirms the strong separability and robustness of the proposed IGWO-DNN across all classes. Overall, the results demonstrate that combining EMD-based time-domain features with IGWO-driven DNN optimization provides an effective and reliable solution for bearing fault diagnosis.