Underwater acoustic signal clustering with probabilistic linear discriminant analysis and Dirichlet process

In underwater acoustic signal clustering system, traditional frame-by-frame hand-craft features failed to catch the invariant discriminative features because of the changeable environmental noise. In this paper, a probabilistic linear discriminant analysis(PLDA) based feature re-extraction model is introduced to extract shared features of multiple frame signal by making use of posterior expectation. PLDA based feature re-extraction method can not only unify the feature dimensions of signal with different last time and extract shared features of multiple frame by minimizing the covariance within the class. Then a Dirichlet process based infinite Gaussian mixture model(DPGMM) is introduced to deal with the variable number of classes of targets. Aggregation degree and adjust rand index(ARI) are defined to compare the performance of unsupervised clustering methods in this paper. The clustering results of measured data indicate that proposed method can achieve higher aggregation degree and ARI by comparing the shared features with MFCCs.