Wavelet-Based Spectral Prediction with Local PLS Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multivariate calibration methods for spectroscopic measurements, such as NIR and MIR, face challenges including high computational costs, large file sizes, and the need for extensive data protection mechanisms due to the direct use of spectra, which complicates the creation and updating of calibration libraries.
Innovation Solution
The method employs wavelet transformation of spectra to create a library of wavelet coefficients, allowing for local PLS modeling using k nearest neighbors, eliminating the need for direct spectrum access and reducing computational complexity by filtering and selecting relevant wavelet coefficients.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If direct spectra are used for multivariate calibration, then measurement precision is maintained, but device complexity and data protection requirements increase significantly
Solution Approach 1:
The patent extracts only the essential calibration information from the original spectra by transforming them into wavelet coefficients. The calibration library stores only these transformed coefficients rather than the complete spectral data, thereby extracting the necessary predictive information while discarding redundant data that would require protection.
Solution Approach 2:
The patent creates a transformed copy of the spectral data in the form of wavelet coefficients. This copy contains the essential calibration information needed for predictions but is fundamentally different from the original spectra, making it impossible to reconstruct the original data from the stored coefficients.
2Measurement precision
If complete spectrum libraries are distributed in networks, then measurement precision is maintained, but loss of time increases due to lengthy distribution processes
Solution Approach 1:
The patent extracts only the essential calibration information from the original spectra by transforming them into wavelet coefficients. The calibration library stores only these transformed coefficients rather than the complete spectral data, thereby extracting the necessary predictive information while discarding redundant data that would require protection.
Solution Approach 2:
The patent changes the parameter representation of spectral data from raw intensity values across many wavelengths to wavelet coefficients that compactly represent the same information. This parameter transformation reduces the data size while preserving the essential features needed for calibration and prediction.
3Measurement precision
If non-linear calibration models like SVR or ANN are used, then measurement precision improves, but productivity decreases due to significantly longer calculation times
Solution Approach 1:
The patent segments the calibration process into two distinct phases: an offline phase where the wavelet-based calibration model is built and stored in the library, and an online phase where predictions are made using the pre-built model. This segmentation allows complex non-linear modeling to be performed once during calibration while enabling rapid predictions during actual use.
Solution Approach 2:
The patent performs the computationally intensive calibration work in advance during the offline phase, creating a pre-built wavelet-based model that is stored in the library. When actual predictions are needed, the system only needs to perform the much faster wavelet transformation and model application, having already done the heavy computational lifting beforehand.
Data Source
AI summary
Techniques are disclosed for predicting a property value of a sample of a particular sample type. A wavelet transformation is applied to the obtained spectrum to compute a sample set of wavelet coefficients in a plurality of wavelet bands. Calibration sets of wavelet coefficients are computed by wavelet transformations of respective NIR/MIR spectra obtained from calibration samples of at least the particular sample type. Each calibration set is associated with one or more reference property values of the respective calibration sample. The system creates a local PLS model of the to-be-determined property value of said sample by selecting a predefined number k of nearest neighbors of the sample set, and computes the local PLS model based on the selected k nearest neighbors and their associated reference property values. The property value of said sample is predicted by applying the local PLS model to the sample set.