A neural network-aided microwave sensing approach for qualitative and quantitative analysis of adulteration in extra virgin olive oil

• Ano de defesa: 2024
• Autor(a) principal: Alarcon, Júlio César Picolo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: adulteração; adulteration; aprendizagem de máquinas; artificial neural network; autenticidade; authenticity; azeite de oliva extra virgem; dielectric spectroscopy; espectroscopia dielétrica; extra virgin olive oil; fraud; fraude; machine learning; microwave sensor; redes neurais artificiais; resonator; ressoador; sensor em microondas
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/18/18155/tde-09092024-101534/

 In this work, we develop a spoof localized surface plasmon resonator-based planar microwave sensor that, together with simple architectures of feedforward artificial neural networks (ANNs), is capable of detecting and quantizing adulterants in extra virgin olive oil (EVOO). We investigate four common adulterants of EVOO, namely soybean oil, corn oil, sunflower oil, and canola oil. The sensor incorporates a four-spiral resonator electromagnetically coupled to microstrip transmission lines operating at 546.8 MHz to detect changes in the complex permittivity of EVOO samples, caused by adulteration. A vector network analyzer (VNA) is used to measure the sensor\'s complex scattering parameters S11 and S21, that serves as inputs for two different multilayer Perceptron (MLP) ANNs. The first model, using only the real and imaginary components of S21 achieves an overall accuracy of 91.4% in detecting which adulterant oil is being applied to the EVOO test samples. In contrast, the second model, incorporating the real and imaginary components of both S11 and S21, attains an overall accuracy of 99.2% in predicting the adulterant used in the test samples. Additionally, we leverage the linear behavior found between the measured |S21| (in dB) and the adulteration levels, expressed as the percentile value of the volume of adulterant per volume of the sample (mL/mL), to develop first order polynomials using partial least squares regression (PLSR) to predict adulteration levels up to 50%. The maximum obtained root mean square error (RMSE) is 2.1% for canola oil adulteration prediction. PLSR yields RMSE values of 0.9% for soybean oil, 1.1% for corn oil, and 1.0% for sunflower oil adulteration. Additionally, in contrast to recent works involving EVOO adulteration that utilize multiple spectral components of EVOO\'s spectra at optical frequencies, we experimentally demonstrate that a single frequency component of the proposed microwave sensor\'s reflected and transmitted signals can provide us enough information to identify and quantify adulterated EVOO samples. This methodology offers both qualitative and quantitative analyses of EVOO, allowing the detection of adulterations as low as 5% with a simple, portable, and practical system.