2025_programme: Underwater Sound Source Classification Using Non-Negative Matrix Factorization-Based Continuous Observation

• Day: June 17, Tuesday
    Location / Time: D. CHLOE at 18:00-18:20
• Last minutes changes: -
• Session: 06. Enhancing underwater acoustic sensing through machine learning
  Organiser(s): Cheng Chi
  Chairperson(s): Peng Xiao
• Lecture: Underwater Sound Source Classification Using Non-Negative Matrix Factorization-Based Continuous Observation
  Paper ID: 2319
  Author(s): Xueyang Yu, Cheng Chi, Shenglong Jin, Donghao Ju, Shuqiu Li, Yu Li, Haining Huang
  Presenter: Haining Huang
  Abstract: Underwater sound source classification is a cutting-edge topic in underwater acoustic signal processing. Most of the existing methods of underwater source classification are based on machine learning. These methods train and test the classification models using single-frame samples, predicting for each sample. It should be noted that these classification methods do not take full account of practical application requirements. In practical applications, continuous observation of source signals over a certain duration is typically needed to make final classification. This paper proposes a non-negative matrix factorization (NMF)-based method for underwater source classification, considering the operation of continuous observation. The proposed method iteratively decomposes the original non-negative feature matrix of the signal sample set, and generates a non-negative basis matrix and a non-negative encoding matrix in the feature space. The template matrices of different source classes are column-wisely concatenated. When a new test sample of unknown source is input, it is expected to exhibit larger encoding coefficients in the template matrix corresponding to its class, resulting in a larger sum of the elements in the encoding vector. By summing the elements of the encoding vector and performing multi-sample continuous observation, the target type is classified. Experimental results demonstrate that, compared to the reference machine learning methods based on single-sample observation, the proposed method requires less training overhead, offers higher classification stability. When the continuous observation duration reaches 40 seconds, the proposed method achieves nearly 100% classification accuracy on the simulated experimental data, which is unattainable by traditional methods.
• Corresponding author: Mr Xueyang Yu
  Affiliation: Institute of Acoustics, Chinese Academy of Sciences
  Country: China