ER-R: Improving regression by deep learning and prior knowledge utilization for fluorescence analysis

Abstract

Linear regression is a dominant estimation technique in chemometrics, where there is a need for inexpensive and reliable sensors for water monitoring. However, most problems are nonlinear, such as the estimation of concentration in solution from an emitted fluorescence spectrum (EFS). Even if an estimation method gives desirable results, at some point it will be used under field conditions, where poor signal quality and less control over environmental effects are expected, leading to poor performance. In this study, we overcome these problems by implementing deep neural network (DNN) models and transfer learning technique for EFS analysis. The proposed models, R (Regression module) and ER (Encoder-Regression), outperformed linear methods and a naive DNN approach for high-quality laboratory-sampled data with a maximum mean relative error of ∼11%, vs. a minimum mean relative error of 184% for the linear methods. In the case of low-quality data, which were simulated based on a real-use case, the lowest error of the linear methods climbed to 263%, whereas the proposed ER model error remained at 9%. At low concentrations, ER gave the best results for all datasets: ∼3.46 ppb in the high-quality datasets, and 2.4 ppb in the low-quality datasets.