Real-time fault detection in sensor data streams is critical for the reliability of intelligent monitoring and cyber-physical systems. However, the multi-dimensional, noise-prone, and non-stationary nature of multi-modal sensor data often limits the effectiveness of classical threshold-based or manual feature extraction methods. This study proposes a robust hybrid methodology that integrates advanced signal processing with end-to-end deep learning models. Multi-sensor signals—including vibration, acoustic emission, torque, and stator current—are processed through a specialized framework based on CNN and LSTM architectures combined with neuro-fuzzy logic. This approach enables the automatic capture of temporal dependencies without manual labeling, ensuring resilience against data loss and environmental noise. Experimental results demonstrate that energy and statistical indicators effectively serve as early markers of degradation. Furthermore, the study confirms that the generalization ability and diagnostic accuracy of the proposed deep learning models significantly improve with increasing data volume, achieving an accuracy of over 95%, thereby outperforming traditional machine learning algorithms.