Application of ANFIS Ensemble with Sparsemax Activation in Ship Acoustic Classification

Resumo

Resumo

This work proposes an interpretable classification system based on fuzzy inference for identifying ship classes from underwater acoustic signals. The methodology segments audio recordings into overlapping 3-second windows, followed by the extraction of time-frequency features (RMS, zero-crossing rate, spectral centroid, bandwidth, roll-off, and MFCCs). After normalization and dimensionality reduction via PCA, classification is performed by an ensemble of seven ANFIS models, each initialized differently and combined through majority voting.Each ANFIS employs 30 fuzzy rules with Gaussian member-ship functions and is trained using stochastic gradient descent, with sparsemax activation to produce sparse probability distributions. The system achieved an accuracy of 91%, outperforming KNN and SVM, and approaching the performance of Random Forest (92%). In addition to its strong performance metrics, the system stands out for the interpretability of its fuzzy activations, making it a promising approach for maritime acoustic monitoring applications.

palavras-chave

Palavras-chave

underwater acoustic signalsfuzzy inferenceinterpretable classificationmajority votingtime-frequency featuresMel-frequency cepstral coefficientsGaussian membership function

links

Links e arquivos