Proposed framework for multi-level extreme machine learning for underwater thruster's fault classification using YIN fundamental frequency estimator and pitch sound

Abstract

Unmanned underwater vehicles (UUV) have been used for many ocean related applications such as scientific exploration, surveillance, marine development, marine engineering and military purposes. However, the risk of damage or loss of property for unmanned underwater vehicle is often a major concern and thruster had been found to be the most common source of faults. Based on the audio recordings using a hydrophone on different thruster's state such as broken propeller fin and propeller entangled with seaweed, fundamental frequency (by YIN fundamental frequency estimator), pitch and root-mean squared (RMS) are extracted for multi-level extreme learning machine (ELM) to perform classification of thruster's faults. However, multi-level machine learning is often built on binary classification for each level with same parameters for each level may not be the most optimal to build a multi-level machine learning. In this paper, a new framework for multi-level ELM is developed using chi-square value that can obtain the optimal parameters from the list of inputs at each level using least number of parameters. Based on this framework, ELM can achieve a 100% accuracy and a comparison to different machine learning algorithms such as Decision Tree and Support Vector Machine using the same parameters show similar accuracy.