Document Type : Applicable
Authors
1 Ms. C student, Mechanical Engineering of Biosystems Department, Faculty of Agriculture, Razi University, Kermanshah, Iran
2 Associate Professor, Mechanical Engineering of Biosystems Department, Faculty of Agriculture, Razi University, Kermanshah, Iran
3 Postdoctoral Researcher, Department of Signal and Information Processing for Sensing Systems, Institute for Bioengineering of Catalonia (IBEC)
Abstract
Introduction: Coffee is a common drink which is obtained from the roasted and ground beans of the coffee plant. Coffee beverages are widely consumed as a stimulant, a property largely attributed to the presence of caffeine, which is the most active pharmaceutical ingredient consumed worldwide. When the fruit of the coffee plant ripens, the coffee beans are harvested, processed, and finally dried. Dried coffee beans are roasted to different degrees and graded depending on the desired aroma and taste. It is very important to detect natural and unnatural impurities and adulteration in coffee.
Materials and Methods: An odor machine system based on eight MOS sensors was used to investigate the effect of bread storage time based on odor characteristics. The designed system includes a data acquisition system, sensors, sensor shield, sample container, power supply, connections, electric valves, air pump, and air filter. The sensor array consisted of 8 MOS sensors, including MQ136, TGS822, MQ9, MQ3, TGS813, TG2620, TG2602, and MQ135, each reacting to specific volatile compounds. These sensors are widely used in olfactory machines because of their high chemical stability, high durability, low response to moisture, and affordable prices. They are the most commonly used sensors in electronic nose systems. Sensors are the main components of an electronic nose system; therefore, it is necessary to select sensors able to detect differences among samples. In this research, the use of electronic nose technology and artificial intelligence was evaluated to detect common adulteration in Arabica coffee (Medium Dark). Robusta coffee samples with weight percentages of 10, 40, 30, 20, and 50% were used for experiments and adulteration. An electronic nose equipped with eight metal oxide sensors was used to carry out experiments related to odor. The data received from 8 sensors was first recorded and stored as raw data. In this research, the fractional method was used to normalize the data. Preprocessed data were used as the input matrix for multivariate analytical methods. The unsupervised multivariate principal component analysis (PCA) method was used to analyze the data. The LDA method was used to reduce classification differences and expand the differences between different groups. The artificial neural networks (ANN) method was used for classification. All calculations and analyses were done using Excel 2016, Unscrambler x10.4, and MATLAB software. Model evaluation criteria are used to evaluate algorithm performance in supervised learning. To analyze the system's performance, common criteria including Specificity, Recall, Precision, Accuracy, Area Under the Curve (AUC), and F-score were used.
Results and Discussion: The results of PCA showed that 87% of the total variance of the data was explained by PC1, and 8% by PC2, and the two main components constituted 95% of the total variance of the normalized data. Based on the results, pure Robusta coffee (B) was located on the right side of the PCA diagram and completely separate from other levels of adulteration. Also, pure Arabica coffee (A) was placed in the vicinity of counterfeit samples, and all counterfeit samples showed the same behavior as Arabica coffee, which is very difficult to distinguish. The loading diagram was examined to determine the role of sensors in separating the groups. Based on the loading diagram for coffee adulteration detection, the sensors that had the highest value on the principal component were MQ9, TGS822, and MQ136. Other sensors also showed a high correlation with the smell of the samples. In other words, other sensors could be neglected. The models of artificial neural networks analysis were evaluated by the correct classification rate (CCR), root mean square error (RMSE), and coefficient of determination (R2). According to the results obtained for 7 different coffee groups, the 7-8-8 structure had the best results. This structure has 8 neurons in the input layer (number of sensors), a hidden layer with 8 neurons, and 7 neurons in the output layer (7 groups). The average values of the class obtained from the ANN model for the parameters of accuracy, precision, recall, specificity, area under the curve (AUC), and F-score were equal to 0.984, 0.952, 0.943, 0.990, 0.971, and 0.942, respectively. Also, the ANN method showed higher accuracy than the LDA method.
Conclusion: The electronic nose showed that it is a fast and effective tool for detecting adulteration substances in coffee.
Keywords
Main Subjects
)