Deep LSTM-Based Bearing Condition Monitoring with Batch Normalization Enhancement

Abstract

Accurate diagnosis of bearing faults is critical for ensuring the reliability and safety of industrial electric motors, particularly under complex operating conditions where traditional signal processing techniques often exhibit limited robustness. To overcome these challenges, this study proposes an advanced deep learning–based diagnostic framework that integrates signal decomposition, feature optimization, and temporal modeling for intelligent bearing fault identification. Vibration signals acquired from rotating machinery are first processed using Ensemble Empirical Mode Decomposition (EEMD) to effectively suppress noise and isolate meaningful intrinsic mode functions. Relevant features are then selected based on correlation coefficient analysis, and Principal Component Analysis (PCA) is applied to reduce feature dimensionality while preserving essential fault-related information. The refined feature set is subsequently fed into a Long Short-Term Memory (LSTM) network enhanced with Batch Normalization to capture temporal dependencies and stabilize the training process. The inclusion of Batch Normalization improves convergence behavior and enhances model generalization under varying operating conditions. Experimental evaluation demonstrates that the proposed framework achieves superior diagnostic performance, reaching 100% classification accuracy and outperforming conventional fault diagnosis approaches. Owing to its robustness, low sensitivity to noise, and strong temporal learning capability, the proposed method provides an effective and reliable solution for real-time bearing fault diagnosis in industrial motor applications.