CNN transfer learning method based on polynomial reconstruction algorithm for Raman spectroscopy

Abstract

The invention discloses a CNN transfer learning method based on a polynomial reconstruction algorithm for Raman spectroscopy. The method comprises the following steps: (1) constructing a big data set by utilizing a public Raman spectrum database; (2) reconstructing a big data set by using a reconstruction algorithm; (3) training the CNN by using the big data set; and (4) applying the CNN to data sets measured by different Raman spectrometers to realize transfer learning. According to the reconstruction algorithm provided by the invention, smoothing and interpolation processing can be carried out on the spectrum at the same time, sampling errors are reduced to the greatest extent while only one-time sampling is needed and noise is filtered out, so that the transfer learning performance of the CNN is improved, and the result is superior to that of an interpolation algorithm. The CNN established based on the method only needs 75% of calibration data to obtain relatively high transfer learning performance, and is superior to existing CNN models. The invention provides a simpler and more efficient way for transfer learning of CNN in Raman spectroscopy.

Description

Technical field [0001] The present invention belongs to the technical field Raman spectroscopy and depth learning techniques, and in particular relates to a method of transfer learning CNN Raman spectroscopy for reconstruction algorithm based on the polynomial. Background technique [0002] With the development of deep learning techniques, neural network technology in the field of Raman spectroscopy has become a popular research, especially in the multi-classification problems, neural network models are often able to show more excellent than the traditional machine learning methods of performance, Raman spectroscopy in combination with depth learning techniques for identification and classification of the substance has become an inevitable trend. Convolutional neural network (CNN) model as a popular deep learning model has demonstrated its unique advantages in the application of Raman spectroscopy, the CNN model but has a good performance usually requires a lot of training data f...

AI Technical Summary

Problems solved by technology

However, more preprocessing steps mean multiple sampling of the original spectrum, which will inevitably introduce more sampling errors and reduce the performance of the CNN model.

Images