We were tasked to assess the suitability of deep-learning methods for complex high-frequency signals such as were produced by recent automated underwater vehicles. Such vehicles transmit detailed data that is considerably more complex than traditional sensors. We interpreted the task as including several subgoals. First, we need to determine distinctive features of these signals. Second, we need to distinguish different signal sources from each other. Third, we need to distinguish periods of time within those signals and make guesses as to what is happening in each. We used an approach of extracting features from both the time domain (wavelets were the most helpful) and the frequency domain (logarithmically spaced frequency components were the most helpful). We trained several kinds of machine-learning models and demonstrated excellent performance in distinguishing the test signals.
展开▼
