dc.description.abstract | This thesis explores the application of unsupervised neural network models, specifically autoencoders and transformers, for detecting anomalies in fish nets using acceleration vibration signals. The study evaluates the comparative performance of these models in terms of
accuracy, precision, recall, and F1-score, and investigates their robustness to environmental noise and sensitivity to low wave height and length configurations. The transformer
model consistently outperformed the autoencoder, achieving an F1-score of 0.95 compared
to the autoencoder’s 0.90. This superior performance is attributed to the transformer’s advanced architecture, which effectively captures complex patterns in the data. Additionally,
the transformer model demonstrated greater robustness to environmental noise, maintaining
high performance levels even with noisy or incomplete data. Contrary to initial expectations,
both models performed well in low wave height and length configurations, highlighting their
adaptability to varying environmental conditions. The findings underscore the potential
of transformers in advancing structural health monitoring for aquaculture, while also acknowledging the autoencoder’s viability in resource-constrained scenarios due to its lower
computational requirements. Future research should focus on optimizing these models for
real-time deployment, exploring hybrid solutions, and ensuring their long-term viability in
diverse environmental conditions. | |