The research applies data-driven methods to locate faults in automotive cylinder heads by analyzing vibrations with supervised machine learning and signal processing techniques. The main goal focuses on determining structural fault positions by analyzing acceleration data through frequency domain features. The Fast Fourier Transform (FFT) processed time-series signals which originated from finite element simulation of controlled impact tests. The transformation process revealed amplitude spectra which allowed researchers to identify the most significant frequency components. The variance-based feature selection method identified five frequency bins which showed the highest variability between faulty cases to detect unique fault location characteristics. The features were used as inputs to create two predictive models which included a linear regression model with Least Squares Estimation (LSE) and a nonlinear Multi-Layer Perceptron (MLP) neural network. The models received their training and testing data from a predefined dataset split to maintain consistent evaluation results. The MLP model performed better than the LSE model according to the results. The MLP achieved a Mean Absolute Error (MAE) of 1.362 mm, whereas the LSE resulted in an MAE of 8.098 mm, both over a 270 mm fault range. The research shows that nonlinear models outperform linear approaches because they better detect complex vibration patterns. The MLP model achieved stable training convergence which confirms its reliable performance. The research confirms that FFT signal processing combined with machine learning algorithms produces precise fault detection results. The proposed method provides a non-destructive and expandable system for automotive component health monitoring which requires additional validation to transition to real-world diagnostic applications.