ENHANCING FAULT CLASSIFICATION IN ROTATING MACHINERY USING A HYBRID MODEL WITH FOURIER-BASED ACTIVATION FUNCTION
TL;DR
The results underscore the effectiveness of incorporating the Fourier activation function into machine learning models, as it enhances the ability to capture complex fault dynamics and improves diagnostic accuracy in industrial applications.
Abstract
Accurate diagnosis of faults in rotating industrial machinery is very critical for operational efficiency, reducing downtime, and avoiding costly failures and the associated economic losses. This study makes use of a hybrid machine learning model designed to classify known fault types which include imbalance, overhang, underhang (bearing), and horizontal faults based on vibration signal data. Traditional hybrid models often struggle to capture the complex, oscillatory nature of vibration data using standard activation functions like ReLU, which can lead to poor representation of non-linear and periodic fault dynamics. The novelty in the proposed approach is the incorporation of the Fourier function as an activation function within the model architecture. This models periodic and non-linear behaviors, enhancing the detection of subtle patterns missed by standard functions. This function captures both periodic and non-linear behaviours in the data, enhancing the model’s ability to distinguish between subtle fault patterns that traditional activation functions might overlook. The model's performance was evaluated using cross-validation, which revealed high precision, recall, and F1-scores (all above 0.90) for imbalance and bearing faults. However, horizontal faults posed more classification challenges, with F1-scores ranging from 0.73 to 0.76, likely due to spectral overlaps and noise. Among the classifiers assessed, XGBoost achieved the lowest error rate at 0.0067, followed closely by Random Forest at 0.007 and CNN at 0.008. Notably, the hybrid model outperformed existing methods, such as Abbasi’s approach, which achieved 92% accuracy, by reaching a peak accuracy of 99% in fold three. These results underscore the effectiveness of incorporating the Fourier activation function into machine learning models, as it enhances the ability to capture complex fault dynamics and improves diagnostic accuracy in industrial applications.