|
ИСТИНА |
Войти в систему Регистрация |
ИСТИНА ИНХС РАН |
||
Solution of Multi-parameter Inverse Problem by Adaptive Methods: Efficiency of Dividing the Problem Spaceстатья
Информация о цитировании статьи получена из
Web of Science,
Scopus
Статья опубликована в журнале из списка Web of Science и/или Scopus
Дата последнего поиска статьи во внешних источниках: 27 марта 2018 г.
- Авторы: Alexander Efitorov, Tatiana Dolenko, Sergey Burikov, Kirill Laptinskiy, Sergey Dolenko
- Журнал: Lecture Notes in Computer Science
- Том: 10614
- Год издания: 2017
- Первая страница: 751
- Последняя страница: 752
- DOI: 10.1007/978-3-319-68612-7
- Аннотация: The considered multi-parameter inverse problem (IP) is determination of concentrations of salts or ions in multi-component water solutions of inorganic salts by Raman spectroscopy with subsequent spectra analysis by a non-linear adaptive method (multilayer perceptron type artificial neural networks (ANN)) or by a linear adaptive method (partial least squares (PLS) method based on principal component analysis) [1]. Dividing the problem space into parts by data clustering simplifies the problem within each cluster but reduces the number of samples. This study compares efficiency of application of this approach for problems with various complexity (determination of concentrations of five salts, or ten salts, or ten ions) and with various distributions of samples over concentration range of the components [2]. It has been demonstrated, that the approach is efficient for IP with medium complexity (5 salts). The best clustering method was Kohonen self organized map (SOM). However, division into physically grounded sections of problem space by classification followed by applying a linear PLS regressor within each class provided better results than clustering; uniform sample distribution over the concentration range also required nonlinear data preprocessing within each class. For a more complex problem (10 salts or 10 ions), the single regressor approach with strongly non-linear ANN regressors turned out to perform better. The main reason of the observed effects is decreasing representativity of data within each section with increasing number of sections; so the results should be checked on other problems with much larger amount of data available.
- Добавил в систему: Ефиторов Александр Олегович