PSI - Issue 77

Hugo Mesquita Vasconcelos et al. / Procedia Structural Integrity 77 (2026) 601–610 Hugo Mesquita Vasconcelos/ Structural Integrity Procedia 00 (2026) 000–000

609

9

minority examples, it used the same file but different 1-second recordings, yielding similar data, increasing their frequency but not their variability, which allowed the network to memorize repeated patterns instead of learning generalizable class distinctions. Due to hardware and monitoring limitations, it was not possible to directly assess the power consumption during training. Only the total training time could be measured, which does not accurately reflect the real energy usage, as it can be influenced by several hardware, firmware, and different system usage factors. Based on the available measurements, the full MHC configuration showed only about a 1% decrease in training time compared to the traditional approach. Interestingly, the two-stage configuration recorded the longest training time among the three, suggesting that the observed variations are possibly related to computational overheads rather than intrinsic differences in power efficiency. When analyzing the detailed per-class results presented in Table 1, the general trend indicates that the MHC configuration yields marginal improvements in well-represented classes, while differences remain small or inconsistent for underrepresented categories. These variations can be explained by how the hierarchical heads influence feature learning over the α – θ – β schedule. During early epochs, the binary head (α -phase) should emphasize distinguishing vessel presence, which tends to stabilize the observed b ackground recall. As θ increases, the middle head enforces grouping-based consistency across vessel sizes, slightly improving recall for large, diverse categories such as cargo and passengership , where intra-class variation benefits from intermediate supervision. For small or infrequent classes like fishing and dredger , improvements are limited because the middle-level grouping cannot fully compensate for the low sample count. The drop in tanker precision likely reflects interference between correlated “big” subclasses, as the shared middle representation promotes coarse similarity before fine differentiation resumes in the β -phase. Meanwhile, classes with few samples ( pilotvessel , pleasurecraft , rescue , sailing ) remain near zero in all configurations, indicating that hierarchical training alone cannot overcome extreme data imbalance. These patterns suggest that the hierarchical structure primarily acts as a regularizer—promoting smoother feature transitions between related vessel types—while preserving stability for dominant classes and exerting minimal effect where the data are too sparse to benefit from shared representations. 5. Conclusions and Future Work The MHC was applied in the maritime acoustic vessel recognition, a OHC for the domain was proposed, and using ONC recordings an test was developed to evaluate if hierarchical learning improved the efficiency of AI model training. Three different approaches were performed. In summary, the proposed structured MHC learning produced small but consistent gains in learning stability and class balance, most notably among well-represented vessel categories. Although the overall improvements were modest, the results indicate that the hierarchical structure supports smoother convergence and slightly better recall in classes with greater internal variability. In the future, the hierarchical complexity concept could extend beyond training into deployment. A lightweight, low-power front-end preoperational computational system could handle preliminary differentiation—such as separating background from vessel—with low power consumption and activate a more capable module only for detailed classification. Acknowledgements This work was developed within the scope of the project MAUSER, which has received funding from the FCT (Fundação para a Ciência e a Tecnologia) under grant 2024.07588.IACDC. Declaration of generative AI and AI-assisted technologies in the manuscript preparation process