Calibration transfer in near infrared spectroscopic analysis using adaptive artificial neural network

• Main Author: Yap, Xien Yin
• Format: Thesis
• Language: English; English; English
• Published: 2021
• Subjects: QA801-939 Analytic mechanics
• Online Access: http://eprints.uthm.edu.my/8368/1/24p%20YAP%20XIEN%20YIN.pdf; http://eprints.uthm.edu.my/8368/2/YAP%20XIEN%20YIN%20COPYRIGHT%20DECLARATION.pdf; http://eprints.uthm.edu.my/8368/3/YAP%20XIEN%20YIN%20WATERMARK.pdf; http://eprints.uthm.edu.my/8368/

Near Infrared Spectroscopy (NIRS) has been implemented in various areas due to its non-invasive and rapid measurement features. A NIRS calibration model can be transferred among different instruments using calibration transfer methods. According to review paper reported by J. Worksman, the most popular calibration transfer methods of spectra standardization methods require primary and secondary instruments to acquire transfer samples at the same samples. However, if the primary instrument is broken, then the existing model cannot be transferred using these methods. Artificial Neural Network (ANN) that has the capability of adapting new environmental conditions. Thus, this study aims to investigate the feasibility of an adaptive ANN (AANN) as an alternative in transferring models from primary to secondary instruments with transfer samples collected on secondary instruments only. First, ANN was developed and optimized using primary instrument’s spectrum. Then, the optimized ANN was adapted to secondary instruments using transfer samples collected on secondary instruments, in which the weights and biases of the ANN were updated. Finding show that the excellent results were obtained using proposed AANN and 20 transfer samples, with the best averaged root mean squared error of prediction (RMSEP) of 0.1017% and the best averaged correlation coefficient of 0.7898, followed by Direct Standardization – Artificial Neural Network (DS-ANN) and Direct Standardization – Adaptive Artificial Neural Network (DS-AANN) in corn oils prediction applications. The proposed AANN outperformed previous works Piecewise Direct Standardization – Partial Least Squared (PDS-PLS) with RMSEP of 0.1321% and 0.1150%, and correlation coefficient of 0.7780 and 0.7785, for m5/mp5 and m5/mp6 respectively. Hence, proposed AANN has the capability to transfer the existing calibration model to secondary instruments without the involvement of primary instrument.