<p>Detecting faults in mechanical components is essential for minimizing repair costs, avoiding unexpected downtime, and ensuring operational safety. However, traditional fault detection techniques often fail to accurately identify faults in rotating machinery due to the complexity and non-stationary nature of vibration signals. To address these limitations, this study proposes a deep learning-based framework for the classification of roller bearing faults. The approach integrates Ensemble Empirical Mode Decomposition (EEMD), Principal Component Analysis (PCA), and a Bidirectional Long Short-Term Memory (Bi-LSTM) network. The process begins with EEMD, which decomposes the raw vibration signal into multiple Intrinsic Mode Functions (IMFs), helping to reduce noise and highlight relevant signal characteristics. A correlation-based method is then used to select the most informative IMFs. Next, PCA is applied to the selected features to reduce dimensionality while retaining essential temporal patterns. These optimized features are used to train a Bi-LSTM model, which learns from both past and future data sequences. To further enhance performance, Batch Normalization (BN) is applied within the Bi-LSTM model. For comparison, the same extracted features are used to train other deep learning models, including LSTM, GRU, and Bi-GRU. The proposed Bi-LSTM model with BN achieves a testing accuracy of 100%, significantly outperforming LSTM (84.90%), GRU (82.92%), and Bi-GRU (83%). In terms of accuracy, the Bi-LSTM model demonstrates superior performance, making it a more reliable and effective solution for real-time fault diagnosis of rolling bearings.</p>