Ultraviolet spectral transfer based on a convolutional variational autoencoder model for detecting chemical oxygen...

Area of Science:

• Environmental Science
• Analytical Chemistry
• Spectroscopy

Background:

• Spectral data shifts between different UV instruments hinder accurate chemical oxygen demand (COD) detection in rivers.
• Existing predictive models often fail due to instrument-specific spectral variations.

Purpose of the Study:

• To develop a non-linear spectral transfer method to overcome challenges in sharing predictive models across different UV instruments.
• To improve the accuracy and reliability of COD detection in river water analysis.

Main Methods:

• A convolutional variational autoencoder (CVAE) was employed for non-linear spectral transfer.
• The CVAE utilizes an encoder-decoder architecture to extract key spectral features and reconstruct data.
• Variational inference was incorporated to model data distribution and capture global spectral features.

Main Results:

• The proposed CVAE method significantly reduced the root mean square error of prediction compared to traditional methods (spectral space transformation, piecewise direct standardization, deep autoencoders, and convolutional autoencoder).
• The CVAE achieved prediction error reductions of 1.0242, 0.4571, 0.4201, and 0.1965 over the comparative methods.
• The method effectively addressed model sharing issues between different measurement devices.

Conclusions:

• The CVAE-based spectral transfer method offers a robust solution for instrument-independent COD prediction.
• This approach enhances the applicability of predictive models in environmental monitoring and water quality assessment.
• The study provides new insights for spectral data processing and multivariate calibration in analytical chemistry.