A finite element-based data-driven fault diagnosis approach for structures using AI

This research presents an innovative AI-based finite element-informed data-driven method for fault diagnosis in mechanical structures and machine elements. An engine cylinder head of internal combustion engine as the case study is used here to showcase the efficiency and robustness of the proposed approach. By coupling high-fidelity finite element (FE) simulations with sophisticated signal processing and artificial intelligence, the method allows for accurate fault localisation and early identification of defects. Vibration signals at specific locations of the structure, produced via explicit dynamic FE simulation under controlled impact loading, are converted into the frequency domain through Fast Fourier Transform (FFT). A variance-based feature selection method determines five high-impact frequency fault signatures (FFSs), drastically lowering data dimensionality without compromising diagnostic accuracy. The features train a nonlinear Multi-Layer Perceptron (MLP) neural network, with fault localisation errors ranging below 7% for test cases. Compared to conventional techniques such as visual inspection and ultrasonic testing, the method surpasses them by allowing non-invasive real-time monitoring, predictive maintenance, operational reliability, and sustainability in automotive engineering. The reduced feature set prevents computational complexity, rendering the method viable for embedded systems and large-scale industrial applications and mass production lines.