MP-ResNet: A dual-rotor system vibration source localization method based on parallel attention mechanism and residual neural network collaboration

Abstract

Vibration source localization in dual-rotor aero-engine systems presents significant challenges due to complex coupling effects and interference patterns between rotating components. To address these challenges, a novel MP-ResNet neural network model is proposed, which integrates parallel attention mechanisms (MSCA-ParNet) with residual neural networks for enhanced vibration source analysis and fault diagnosis of dual-rotor systems. A comprehensive validation strategy is implemented through two experimental platforms. Initially, a multi-source excitation vibration simulation test bench is established, where sinusoidal excitations of different frequencies are applied to bearing seats, and vibration response signals are collected from the outer casing to train the model for rapid classification and vibration source localization. Comparative analysis demonstrates that MP-ResNet significantly outperforms traditional neural networks (GoogleNet, AlexNet, DenseNet) and existing attention-based networks (CBAM-ResNet, SE-ResNet, ECA-ResNet), achieving 98.4% recognition accuracy with training loss as low as 0.091 on the simulation platform. Further validation on a dual-rotor aero-engine vibration simulation test bench across eight operating conditions confirms superior performance with 97.08% overall accuracy, 97.64% precision, and 97.18% F1 score, demonstrating accuracy improvements of 1.25–4.58% over competing methods. These results demonstrate that MP-ResNet provides a robust and effective solution for dual-rotor vibration source localization, showing strong potential for advancing industrial fault diagnosis applications.