This thesis investigates how analysis level, temporal stability, and model agreement jointly influence unsupervised anomaly detection in Automatic Identification System (AIS) trip data. Using Danish fishing vessel trajectories collected between December 2024 and May 202 we segmented and transformed 9,189 trips into 27 behavioral features covering speed, AIS gaps, port novelty, and route geometry. Four unsupervised models were applied, Isolation Forest, Gaussian Mixture Models, Local Outlier Factor, and Mahalanobis Distance, at four analysis levels: global, departure harbor, arrival harbor, and vessel. Results show that discriminative features vary by scope. At the vessel level, anomalies are best explained by trajectory-shape variability, whereas at harbor levels AIS gaps and trip duration dominate. Temporal stability analyses revealed substantial drift at harbor and global levels, while vessel-level features remained more stable. Model agreement was moderate: IF and MD aligned most consistently, while GMM produced distinct anomaly profiles. These findings highlight that anomaly detection is inherently context-dependent. For operational use, combining multiple models and accounting for feature drift is essential to ensure robustness, interpretability, and long-term reliability. Limitations include the reliance on public Danish AIS data and the absence of ground-truth anomalies; future work should test adaptive baselines on broader fleets.