Date of Award
Summer 2022
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Electrical and Computer Engineering
First Advisor
Ye, Dong Hye
Second Advisor
Lee, Chung Hoon
Third Advisor
Frigo, Fred
Abstract
In the research on detecting different ppb-level metal ions in drinking water in multiple areas. Through the block loop gap resonator (BLGR), this kind of metal ion sensor can continuously measure the concentrations of different metal ions in the water sample. The BLGR can classify different ions and measure the Cu and Pb ions in different ppb-level water in the city. According to the different magnitude and phases in continuous frequency in the testing. A better model can be trained by using deep learning, the convolution neural network. A neural network is often regarded as a black-box model because its very strength in modeling complex interactions also makes its operation almost impossible to explain. However, the neural network is an effective tool to make evaluations and predictions, they can perform very well in this kind of area. The neural network can play a better model learning ability in the case of processing a large number of data, especially in the detecting of the metal composition of water quality in multiple areas, more data points can be better evaluated and predicted. In this project, we are going to improve the explainability of neural networks applied in detecting ppb-level water systems. We propose to apply the Layer-wise relevance propagation (LRP) algorithm to explain the relevance of different frequency features. This algorithm can highlight the features that contribute most to deep network learning. We can effectively explain which frequencies are more decisive by the high contribution characteristics to the LRP output. We believe that being able to explain machine learning-based decisions greatly improves our analytical capabilities for ppb-level water detecting in this project.