Hybrid Wavelet–LSTM and Weibull Risk Modeling for Early Fault Prediction Using Accelerometer-Based Vibration Data in Rotating Machinery

As industrial processes advance toward Industry 5.0, predictive maintenance is increasingly vital for mitigating unplanned downtime in rotating machinery. However, existing data-driven prognostics often struggle with high-frequency industrial noise, limited run-to-failure data, and the inability of standard recurrent models to accurately capture long-term degradation, leading to high error rates in precision-critical applications. To address these limitations, this study proposes a novel, noise-robust hybrid framework integrating an Attention-based CNN-Transformer architecture with a Dynamic Mixture Weibull distribution. Initially, an enhanced signal denoising pre-processing step is employed alongside Continuous Wavelet Transform to extract robust multi-scale features, even under extreme noise conditions (e.g., -5dB SNR). The extracted features are processed by a CNN-Transformer network, where convolutional layers capture local fault signatures, and the self-attention mechanism effectively models complex, long-term degradation trends. Furthermore, a Mixture Weibull survival model is integrated to provide probabilistic Remaining Useful Life boundaries, accommodating complex, non-linear failure behaviours. Unlike previous approaches limited to controlled laboratory setups, the proposed framework is rigorously validated on standard industrial datasets. Experimental results demonstrate a dramatic improvement in predictive accuracy, reducing the MAPE to under 15%. Statistical evaluations, including the Diebold-Mariano test, confirm the model's significant superiority over traditional LSTM and CNN-LSTM baselines, establishing it as a highly reliable and generalizable tool for early fault prediction in real-world industrial environments.