Deep cepstrum-wavelet autoencoder: A novel intelligent sonar classifier. (15th September 2022)

• Record Type: Journal Article
• Authors: Jia, Hailong; Khishe, Mohammad; Mohammadi, Mokhtar; Rashidi, Shima
• Is Part Of: Expert systems with applications. Volume 202(2022)
• Publication Date: 2022-09-15
• Subjects: Underwater target classification -- Deep cepstral liftering -- Autoencoders -- Wavelet network; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.117295
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: A novel deep wavelet autoencoder is proposed. Cepstral lifters reduce the multi-path distortion and time-varying effects. Wavelet autoencoder is used to extract fluctuating frequency signature of vessels. Three underwater acoustic datasets are exploited. The results are evaluated, comparing ten state-of-the-art methods. Abstract: Different marine vessels belonging to the same class may have different and time-varying radiated noise due to different and changing machinery configurations. Further, the time-varying radiated noise, which is propagated underwater, suffers from random fluctuations in frequency, amplitude, and phase. The ambient noise and time-varying multi-path effects further complicate the classification problem. In this study, for the first time, a deep wavelet autoencoder is proposed, evolving by a two-hybrid cepstral concept to reduce the multi-path distortion and time-varying shallow underwater channel effects. In this approach, the autoencoder is an automatic feature extractor, choosing the best feature composition without information loss and human intervention, and wavelet networks extract the periodic frequency signatures of vessels, which are fluctuating by various machinery regimes. Three underwater acoustic datasets, including synthetic, ShipsEar, and real experimental datasets, are exploited to evaluate the performance of the proposed model so that the performance of the designed model is compared with ten benchmark methods. The results show that the designed model with an average accuracy of 96.11% and average giga-multiplier–accumulators equal to 0.019 represents the best performance in terms of accuracy and complexity compared to benchmark methods. Besides, the cepstral liftering procedure and the average cepstral feature can significantly alleviate the shallow underwater channel's multi-path distortion and time-varying channel effects. Furthermore, the proposed model is less sensitive to SNR level compare to other benchmark methods.