Modified LinkNet and DenseNet-Based Hybrid Architecture for Combined Fault Detection in an Electromechanical System

Abstract

Purpose: Electromechanical systems are widely used in industries for various applications. Electromechanical systems mostly have an electric motor as a prime mover and a mechanical load, such as a rotor, gearbox, pumps, etc., coupled to it. Vibration monitoring of mechanical systems is a very effective condition-monitoring technique to detect faults. Electromechanical systems may have combined faults, i.e., faults in motors and faults in loads. Diagnosing combined faults is challenging due to overlapping symptoms and their compounded effects. Methods: To address this issue, a modified LinkNet and DenseNet (MLiDNet) fault classification model is proposed. This model integrates signal processing techniques, entropy-based feature extraction, and hybrid deep learning (DL) classifiers. Signal pre-processing has employed the improved Synchro-squeezing wavelet transform (ISSWT). Feature extraction includes entropy-based features such as norm, improved spectral, threshold, and wavelet energy entropy. The hybrid DL classifier includes MLiDNet. Results and Conclusions: The results demonstrate that the MLiDNet achieves a maximum accuracy of 99.79% with 99.68% precision and 99.37% F-measure with a minimum computational time compared to other traditional models. The findings suggest that the proposed method has the potential to serve as an effective alternative for accurate fault detection of combined faults in electromechanical systems. © Springer Nature Singapore Pte Ltd. 2025.