Vibration-based structural health monitoring (SHM) has become a practical tool for assessing joint in￾tegrity in offshore wind jacket-type support structures, where bolted connections are critical for relia￾bility and maintenance planning. However, robust generalization remains challenging when signals are highly correlated within the same experimental run and when adjacent damage states present similar spectral signatures, potentially inflating performance if the data split is not performed at the experi￾ment level. This work proposes a supervised frequency-domain classification pipeline for bolt-loosening detection using a public dataset acquired on an instrumented jacket structure under four integrity con￾ditions: Healthy, 6 Nm, 9 Nm, and NoBolt. To mitigate information leakage, the dataset is split at the experiment/file level prior to windowing. Each triaxial record is segmented into overlapping windows of 1024 samples with a hop size of 512. For every window, a real-valued fast Fourier transform (rFFT) is computed per channel and the magnitude spectrum is used as the descriptor (513 bins per channel, 1539 features per window). Features are standardized using parameters fitted exclusively on the training set. Classification is performed using a Support Vector Machine (SVM) with a Radial Basis Function (RBF) kernel, with hyperparameters selected through a validation/cross-validation strategy and performance assessed on a held-out test set. The proposed approach achieves test accuracy above 95% with strong balanced performance across classes. Remaining errors are mainly concentrated between 6 Nm and NoBolt, consistent with similar spectral patterns under severe loosening/failure. These results indicate that FFT-based descriptors combined with an RBF-SVM provide a compact, reproducible, and high-performing baseline for SHM of bolted joints in jacket-type structures.