Método de classificação de fluxo utilizando sensoriamento acústico distribuído e redes neurais

Abstract

The work presented in this master’s dissertation proposes a classification technique to differentiate the flow type in a test tube, in one of the following classes: Only liquid, only gas and multiphase (composed by liquid and gas). The flow data were collected from experiments with different values of phase velocity for liquid and gas, using a Φ-OTDR distributed acoustic sensor (DAS), with heterodyne detection. The methodology used to perform the classification task is based on a transfer learning technique applied to neural networks, where a pre-trained model is used for extracting a vector of embeddings, which are used as input values for the training of a customized classification network. To generate the customized network training and test datasets, the experiment data were pre-processed, converted into audio files, and inferred by the pre-trained model, generating embedding vectors related to the respective flow type. After training, the classifier was able to achieve an accuracy above 95,0 % for the training dataset and 87,0 % for the test datasets for each class, showing that the methodology used adapts well to the proposed task and that it has potential to be applied in more complex classification tasks extending the number of flow classes.