FRF-based crack localization in AMB-Supported rotors using neural networks

Well-established procedures exist for monitoring and diagnosing faults in rotating machinery, and many techniques for detecting rotor cracks have been explored in the literature. However, limited progress has been made in developing non-invasive methods capable of accurately localizing rotor cracks and assessing their severity without requiring rotor disassembly or direct physical inspection. This paper presents a novel, non-invasive approach for crack localization in flexible rotors supported by Active Magnetic Bearings (AMBs), based exclusively on frequency responses acquired through AMB excitation. The methodology involves constructing a physics-informed fault dictionary using frequency responses simulated on a high-fidelity digital twin of the rotor system, obtained through established modeling procedures, under various crack locations and severities. These responses exhibit characteristic shifts in resonance and antiresonance frequencies, which are used to define distinct fault classes. Neural network classifiers were trained on the simulated dataset, with a 1D Convolutional Neural Network (1D-CNN) used as the primary model and an Autoencoder + Multilayer Perceptron (AE + MLP) used as a comparative baseline, to evaluate their ability to automatically identify the fault zone. The entire framework was validated experimentally on a dedicated AMB-supported test rig, confirming the ability of the proposed method to detect and localize cracks without requiring additional sensors or plant disassembly. The 1D-CNN achieved a classification accuracy of 99.4% on simulated test data, while the AE + MLP baseline reached 98.3%. Experimental validation on a dedicated AMB-supported test rig showed correct localization for all tested crack cases.