A broadband absorption spectrum processing method based on end-to-end deep learning networks

Absorption spectrum technology is an ideal candidate for engine temperature measurement, new material design, and chemical reaction mechanism analysis. However, to date, the existing methods for processing the absorption spectrum are very cumbersome and time-consuming, typically involving steps, such as filtering and denoising, baseline correction, and comparison of measured data with theoretical databases, which results in poor real-time performance. Herein, we propose an absorption spectrum processing method based on a neural network that can directly input the collected unprocessed absorption spectrum and obtain the test results without other steps. To address the challenge of limited spectral features in the absorption spectrum and to enhance the processing accuracy, we designed a novel neural network, to the best of our knowledge, by combining a long short-term memory (LSTM) network and a fully convolutional network (FCN). The final experimental results show that the proposed method achieves accurate temperature processing results for H 2 O and CO 2 , with an accuracy rate of 98.4%, and the average processing time for each absorption spectrum is 0.00036 s, which, to the best of our knowledge, is the fastest processing speed for absorption spectrum. As a result, the presented method shows great potential for real-time online processing of the absorption spectrum.