Hierarchical complexity-based AI model for efficient feature extraction in maritime acoustic signal recognition

Learning models based on hierarchical complexity reflect the way humans naturally acquire knowledge, and experimental evidence suggests they hold promise for improving the efficiency of model training in artificial intelligence (AI). This research presents an innovative approach to developing an AI model capable of classifying maritime acoustic signals, for ship identification or structural integrity assessment. Acoustic signal analysis is critical in maritime environments, as sound travels effectively underwater, offering potential for applications where above-water technologies are not possible. Nonetheless, decoding these acoustic signals is a complex task that presents significant computational challenges. This study applies the Model of Hierarchical Complexity (MHC) to maritime acoustic signal recognition. A domain-specific Order of Hierarchical Complexity was proposed, and three training configurations on a ResNet-18 backbone were evaluated under identical architecture and hyperparameters: traditional non-structured learning, two-stage (binary, multiclass), and full three-stage MHC-structured training. The dataset was strongly imbalanced across the 12 classes (11 vessel types and background), reflecting real maritime traffic; this realism introduced constraints on rare categories, lowering the model performance. The full MHC configuration achieved the best overall metrics (accuracy 0.82 and the highest macro-averaged precision, recall, and F1) with 1% better training time, comparable to the other tests. Improvements were concentrated in well-represented classes, indicating that MHC-structured training can organize learning without additional computational cost but does not, by itself, overcome class imbalance.