Development of a portable electronic nose for the classification of tea quality based on tea dregs aroma

Camellia sinensis is a species in the flowering plant family of Theaceae whose leaf buds and leaves are used to produce the popular beverage tea. Tea is famous worldwide, although in significantly varied amounts; tea is widely regarded as the most consumed beverage on earth after water [1]. The combination of the processing techniques and the characteristics of manufactured tea determines the tea quality. Tea has different quality characteristics, including aroma, taste, color, and appearance due to the manufacturing process. The different types of tea, including black tea, can be classified based on its processing method: CTC (crush, tear, and curl) and orthodox tea. Orthodox tea has increasingly become famous because of its antioxidant and antibacterial properties [2]. Moreover, in terms of its quality, global black tea can be categorized into four grades of tea: leaf grade, broken, fannings, and dust. Based on this global category, the tea industry subsequently classifies its black tea into the first grade (i.e., Broken Orange Pekoe [BOP], Broken Orange Pekoe Fanning [BOPF], Pekoe Fanning [PF], Fanning), second grade (i.e., PFII, FII, etc.), and off-grade (i.e., Bohea). This grading category defines the price of tea. Kumar et al. [3] revealed that the color of tea liquor, taste, and aroma predominantly contribute to the quality of tea. Aroma is one of the crucial aspects of tea quality, which can decide the acceptance or rejection of a tea [1]. Tea aroma depends upon certain volatile aromatic substances developed during the fermentation process [4]. During the fermentation process, the grassy smell of tea leaves is converted into a floral smell through a complex chain of biochemical reactions [5, 6]. Therefore, the fermentation process is the most crucial of all the processes in the formation of tea aroma. The determination of volatile compounds that actively contribute to the aroma of tea has been heavily studied in many varieties of tea using various methods [7,8,9,10,11].

The assessment of tea quality is one of the essential points since it affects the price of the tea product and customer acceptance. The traditional evaluation of tea quality is still carried out manually through organoleptic testing by expert tea testers by relying on their human senses. Unfortunately, the physiological and psychological factors during the test could impact the assessment’s results [12], as well as the environmental factor. This leads to inconsistent, subjective, and expensive assessment outcomes, and complicated and challenging processes for ordinary people to understand. Factors like sensitivity and fatigue also have an impact on the outcomes of the trained tea tester’s sensory analysis [13]. Therefore, testing in this way cannot be used to evaluate the quality of a product with a large number of samples [7]. Although sensory evaluation based on human smell is subjective, careful planning and comprehensive training of assessors enable it to become more objective but remain a more challenging alternative.

Additionally, some analytical techniques, including gas chromatography/mass spectrometry (GC/MS), a sophisticated tool that detects aroma compounds, can be used but requires specific analytical skills by an expert. Sample preparation costs are relatively expensive and inappropriate for real-time monitoring. In addition, since the tool is generally placed in laboratories, it is not lightweight and easy to use/carry outside. Therefore, there is a vital necessity to develop reliable, nonexpensive, precise, and rapid systems to serve as monitoring devices on the process of tea quality assessment. Due to their ability for odor analysis, electronic noses with multisensor systems have been demonstrated as an alternative way to detect and evaluate food quality [14].

The electronic nose (e-nose) is an electronic sensing device that mimics the human sense of smell and is used to detect, recognize, or quantify volatile substances [15]. Gardner stated that “an electronic nose is a system that uses an array of electronic chemical sensors and pattern recognition algorithm, capable of identifying simple or complicated odors,” [16]. An electronic nose generally consists of hardware and software. The hardware on the electronic nose is a combination of a series of gas sensors as an aroma extraction system that functions to produce a sensor array as the sensor response signal outputs. Furthermore, the software on the electronic nose is a control and measurement system with a pattern recognition method to evaluate the sensor array patterns for analyzing the aroma [15, 18, 19]. The arrangement of sensor nodes with various functions can be used to develop smart sensing systems for different applications [20]. Artificial intelligence (AI) can be adopted throughout the value chain of agricultural products, making work efficient and reducing costs [21, 22].

Machine learning techniques enable accurate odor identification with qualitative and quantitative analysis. The e-nose signal pattern analysis in the tea quality assessment has been done using general machine learning techniques consisting of preprocessing and feature extraction stages, machine learning modeling, and drift compensation. The identification performance depends on the e-nose system setup and target gas [23]. Xu et al. obtained perfect results for the qualitative identification of tea quality grades based on a feature-level fusion strategy with support vector machine (SVM) and random forest (RF) [24]. Zhi et al. and Dai et al. also applied procedures to tea quality identification including feature-level fusion (fuse the time-domain-based feature and frequency-domain-based feature) and decision-level fusion (D-S evidence to combine the classification results from multiple classifiers) [25, 26]. Experimental results showed that the fused features could better represent signal characteristics compared with the single features and indicated that nonlinear algorithms were more effective in feature selection than linear algorithms, and the highest recognition rate was reached by feature fusion and KLDA-K-Nearest Neighbor (KNN) model. Rouabeh et al. designed fast and low-cost multipurpose e-nose for rapid gas identification [27]. This research compared Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) to project the extracted dataset to demonstrate the discriminative skills of the chosen features and also compared the KNN and SVM radial basis function-based (SVM-RBF) classifiers. These obtained results show that the LDA performed better than the PCA in terms of discrimination and the SVM-RBF has the best, stable, and consistent classification accuracy. Benabdellah et al. identified bad odor diffuses from two rotten types of meat that have almost the same rottenness odor using an electronic nose system [28]. Classification and identification of these odors are realized by PCA and Discriminative Factorial Analysis (DFA) methods and perform a good separation between classes. Xu et al. also compared three data reduction methods including PCA, LDA, and multidimensional scaling (MDS), and two classifiers including SVM and logistic regression (LR) for qualitative evaluation of tea [29]. The results indicated that LDA outperformed original data, PCA, and MDS in both LR and SVM models, and SVM had an advantage over LR in developing classification models. Tudu et al. assessed eight normalization techniques and multilayer perceptron (MLP) classifier for black tea aroma classification using an electronic nose [30]. It is observed that the normalization techniques influence the results considerably and careful choice of the appropriate method is significant. However, none of the normalization techniques and machine learning approaches hold good results for each application, thus the performance differs based on the application [29, 30].

The results of manually performed preprocessing and feature extraction cannot guarantee the optimization of the machine learning model performance. The classification method without feature extraction is possible using deep learning algorithms. Unlike machine learning, the deep learning approach does not require us to manually extract or select features [31]. This is because deep learning is composed of several processing layers based on artificial neural networks to learn data representations with different levels of abstraction so that no additional processing is needed to represent the data [32]. MLP is a learning algorithm with one or many layers called a hidden layer that contains one or more neurons to carry out learning that can represent the characteristics of the input data. In addition, MLP can study nonlinear and complex models and adapt to data, so there is no requirement for an explicit functional form specification and model distribution [33].

On this basis, this research aims to develop a portable electronic nose system to assess tea quality based on the aroma of tea dregs by relying on the aromatic capture process through sensors. This is done by using a learning algorithm to classify tea aroma quality into three classes, i.e., fresh, burnt, and strange smell. Thus, the first part of this paper elaborates on the following steps: (1) developing a portable electronic nose system based on the fusion features from odor sensor arrangement; (2) data acquisition from various aromas of tea dregs with different quality; and (3) analysis the classification algorithms to determine tea quality by comparing the traditional machine learning and MLP algorithm. Furthermore, section II presents the portable electronic nose, data acquisition, classifiers, and experimental setup, while section III presents results and discussion, and section IV concludes this paper.

This section provides and discusses the evaluation results of the classification methods using machine learning and MLP models. Table 4 displays the evaluation performances for machine learning classifiers, including the average accuracy, precision, recall, and F1-score.

According to the evaluation results using machine learning classifiers as illustrated in Table 4, the NB classifier has the lowest average accuracy of 0.5972 with a standard deviation of 0.0241 and an average F1-score of 0.5432 with a standard deviation of 0.0181. Moreover, the DT classifier produced slightly better results than NB with an accuracy of around 0.5972, but with a higher standard deviation of 0.0992 and an F1-score of around 0.6002 with a standard deviation of 0.0932. The high standard deviation scores indicate that metric scores are still not constant and change significantly. The RF classifier performs slightly better than the DT, with an accuracy of 0.6667 ± 0.0556 and an F1-score of 0.6572 ± 0.0444. The best machine learning classifiers are SVM and KNN. The average accuracy of these two models is identical with around 0.7778. However, the standard deviation for KNN is 0 and the SVM is 0.0393. Thus, the KNN result is more consistent than the SVM. Furthermore, according to the F1-score, there is a slightly different result between KNN and SVM. KNN obtained an F1-score of 0.7575 ± 0.0112, while a higher score was obtained with the SVM which gathered an F1-score of 0.7702 ± 0.0433.

Table 5 demonstrates the evaluation performances for MLP with variations of the number of hidden layers, the number of hidden nodes, and the activation function. The evaluation metrics include the average accuracy, precision, recall, and F1-score.

Evaluation performances in Table 5 indicate that the MLP model using the ReLU activation function is better than the Sigmoid activation function. It also revealed that increasing the number of hidden layers and nodes does not necessarily enhance its performance. The best architecture obtained the highest average accuracy of 0.8750 with a standard deviation of 0.0241 and F1-score of 0.8703 with a standard deviation of 0.0246 when using the MLP model with three hidden layers and 100 hidden nodes in each layer.

Overall, the performance evaluation results using the MLP model are higher than the performance results of traditional machine learning classifiers. The highest performance results in experiments using machine learning classifiers were achieved with the SVM method with an F1-score of 0.7702 ± 0.0433. Meanwhile, the performance results when using the MLP model with three hidden layers and 100 hidden nodes in each layer produced an F1 score of 0.8703 ± 0.0246.

This study developed a portable electronic nose that can assess tea quality based on tea dregs’ aroma by relying on the aromatic capture process through sensors and then using MLP to classify the quality of tea aroma. Tea quality was divided into three classes, i.e., fresh, burnt, and strange smell. The sensors used in the TeaNose were TGS 2602, TGD 2620, MQ-7, MQ-9, TGS 2600, and TGS 2611. The dataset in this study consisted of three aroma quality classes, each consisting of 30 samples. A MLP with variations of the activation function, the number of hidden layers, and the number of nodes were used in this study. The MLP was then compared with five machine learning classifiers. According to the machine learning classifier, the SVM obtained the highest average accuracy of 0.7778 ± 0.0393 and an F1-score of 0.7702 ± 0.0433. The results revealed that the classification using the MLP model with ReLU activation function and three hidden layers with 100 hidden nodes obtained the highest average accuracy of 0.8750 ± 0.0241 and F1-score of 0.8703 ± 0.0246. Based on these experiments, it is noteworthy that the MLP model using the ReLU activation function is better than the Sigmoid activation function while increasing the number of hidden layers and hidden nodes does not always enhance its performance. Overall, the MLP model produces better performance evaluation results than traditional machine learning classifiers.

The portable electronic nose developed in this research used 6 sensors and limited tea sample data consisting of 3 tea aroma qualities with each sample taken 30 times. Future research can be improved by adding sensors to the portable e-nose and increasing the number of datasets used. Apart from that, the model classifiers used to detect tea quality are still limited to traditional machine learning classifiers. In the future, this research can be explored by using ensemble learning to combine several machine learning models or using deep learning models based on recurrent neural network (RNN) architecture.