FTIR Analysis Device Using Segmented Reference Samples

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

FTIR analyzers face challenges in achieving high measurement accuracy due to overfitting and pseudo-correlations when analyzing exhaust gas, as they struggle to separate the contributions of individual hydrocarbons in complex gas mixtures, leading to decreased analysis accuracy.

Innovation Solution

An analysis device that employs a machine learning model using both a first reference sample with multiple components and a second reference sample, which includes either one or multiple components from the first sample, to improve accuracy by learning the individual contributions of each hydrocarbon and avoiding pseudo-correlations, thereby enhancing the robustness of the model against changes in the measurement sample's composition.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If machine learning is performed using only the optical spectrum of exhaust gas as training data, then the model can be trained with available data, but it becomes difficult to separate out and learn the contributions of each hydrocarbon in the THC concentration

Engineering Contradiction:

Improveease of training data preparationVSAvoidanalysis accuracy

Core Design Contradiction:

Ease of manufactureVSMeasurement precision

Solution Approach 1:

The training data is segmented into two distinct types: first reference sample data containing optical spectra of exhaust gas with known THC concentrations, and second reference sample data containing optical spectra of individual hydrocarbon components with known concentrations. This segmentation allows the machine learning model to separately learn the spectral characteristics of each hydrocarbon component and their individual contributions to THC concentration, thereby improving analysis accuracy while maintaining ease of training data preparation

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary action by preparing pure hydrocarbon component samples in advance and obtaining their optical spectra before mixing them into exhaust gas samples. This preliminary characterization of individual component spectra enables the machine learning model to distinguish and quantify each hydrocarbon's contribution to the total THC concentration, resolving the overfitting problem that occurs when training only on mixed exhaust gas spectra

Inventive Principle:

Principle #10Preliminary action

2Adaptability or versatility

If a machine learning model is trained on mixed hydrocarbon data, then training is simplified, but the model suffers from overfitting when hydrocarbon composition deviates from training data

Engineering Contradiction:

Improverobustness to composition changesVSAvoidmeasurement accuracy

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

By segmenting the training data into individual hydrocarbon component spectra (second reference sample data) and mixed exhaust gas spectra (first reference sample data), the model learns both the unique spectral signatures of each component and their combined effects. This segmentation enhances the model's adaptability to different hydrocarbon compositions while maintaining reliable measurement accuracy across varying sample types

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent changes the parameters of the training data by varying the hydrocarbon composition ratios in the first reference samples while maintaining pure component spectra in the second reference samples. This parameter variation during training improves the model's adaptability to different exhaust gas compositions, preventing overfitting to specific composition patterns and ensuring reliable measurements across diverse real-world conditions

Inventive Principle:

Principle #35Parameter changes

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

This approach significantly improves the analysis accuracy of total hydrocarbon concentrations in exhaust gas by accurately learning the contributions of each component and avoiding pseudo-correlations, resulting in a more robust and precise measurement model.

Implementation Method 1

a correlation data storage portion that stores correlation data that shows a correlation between spectral data for a reference sample in which total analysis values for a predetermined plurality of components are already known, and a total analysis value of the reference sample

Methodology Applied

Scientific EffectMachine learning:

Data Source

PatentUS20230296501A1Analysis device, analysis method, program for analysis device, learning device for analysis, learning method for analysis, and program for learning device for analysis

Publication Date: 2023.09.21 HORIBA LTD

AI summary

An analysis device analyzes a measurement sample based on spectral data obtained from that measurement sample. This analysis device includes a correlation data storage portion that stores correlation data that shows a correlation between spectral data for a reference sample in which total analysis values for a predetermined plurality of components are already known, and a total analysis value of the reference sample, and a calculation main unit that applies the correlation data stored in the correlation data storage portion to the spectral data obtained from the measurement sample, and then calculates the total analysis values of the predetermined plurality of components contained in the measurement sample. The reference sample includes a first reference sample that contains the predetermined plurality of components, and a second reference sample that is consisting of either one or a plurality of the components contained in the first reference sample.