Accurate peak detection in vibration spectra is key to early fault diagnosis but remains challenging due to noise and unknown spectral content. This paper presents a hybrid approach combining the benefits of U-Nets and Region-based Convolutional Neural Networks for blind peak detection, trained on complex simulated 1D data with extension to real-world signals. The model combines classification and regression to identify and localise peaks without prior knowledge of the signal or machine. Its performance is benchmarked against the Exact Order Model Estimation of Signal Parameters via Rotational Invariance Technique and a Topological Peak Identification approach using labelled simulated data, demonstrating robustness and practical effectiveness. The method is further tested on real offshore wind turbine data, where it successfully detects visually identifiable peaks, confirming its applicability in real-world scenarios.
