Design of an Iterative Model for Fault Severity Classification Using CNN-LSTM and Attention Mechanisms

Accurate classification of ball bearing fault severity is imperative for predictive maintenance and mechanical system reliability. A new hybrid deep learning methodology at the junction of deep convolutional neural networks and long short-term memory recurrent neural networks, empowered by attention mechanisms and domain-specific feature engineering, is proposed to further enhance fault severity classification accuracy. The proposed approach will address the limitations of the available methodologies, which were oriented either to the spatial or temporal analysis of vibration signals. Spatial patterns are captured by the CNN, while at the same time, temporal dependencies are dealt with by the LSTM. Further emphasis on only the fault-related important features is brought out by an attention layer. Additionally, domain-specific feature engineering is performed, wherein statistical and frequency domain features obtained using the Fast Fourier Transform aid in increasing the accuracy of classification. The model works with an accuracy of up to 94%, outperforming both standalone CNN and LSTM models. Finally, explainable AI methods—more specifically, SHAP values and LIME—improve the interpretability of the results from the model by explaining how features contribute to predictions and confirming the central role of temporal patterns. This methodology advances fault severity classification with pragmatic implications for predictive maintenance and mechanical system reliability improvement through the combined power of CNN, LSTM, attention mechanisms, and domain-specific feature engineering processes.