表題番号：2021C-514
日付：2022/04/07
研究課題深層学習を用いた気液二相流用非接触マルチメータの基礎研究
| 研究者所属（当時） | 資格 | 氏名 | |
|---|---|---|---|
| （代表者） | 理工学術院 基幹理工学部 | 教授 | 佐藤 哲也 |
- 研究成果概要
- The principle of a gas-liquid two phase flow multimeter using a capacitive void fraction sensor and deep learning technique was developed. The multimeter can be applied to various fluids and flow styles.
We created a database of the flow rate and void fraction of gas and liquid by an experiment in a horizontal pipe using silicon oil / air flow. As a deep learning method, a multi-class classifier has been developed using a bidirectional long short-term memory network (BLSTM), which is suitable for learning the data including time-series information. This classifier has two BLSTM layers with 320 hidden units.
As a result, the discrimination accuracy of the classifier is about 60%. Most of the misclassifications are those with adjacent gas phase flow rates, followed by those with adjacent void fraction ratio. To clarify the cause of the misclassification, the output of the network was analyzed by dimensional compression using the t-SNE method. In addition, using an index “ambiguity”, the relationship between misclassification and the mean value and fluctuation of the void fraction was investigated. We will improve the accuracy of classification and develop a regressor that detects continuous flow rates.