Survey on Applications of Electronic Nose

: Food plays a vital role in our daily life. Providing good quality food to consumers is essential. Food quality can be accessed using the Electronic nose. Electronic nose (E-nose) is an instrument for odor analysis. E-nose mimics the human olfaction system. It is widely used in predicting the quality of foodstuffs and detecting the contamination in foods. E-nose can also be used in outdoor monitoring such as air quality monitoring and detect the hazardous odors emitted wastewater treatment plants. Application of E-nose is increasing day by day. In this paper, we consolidated the previous works on E-nose. They had applied different machine learning algorithms to construct a model. In most of the works, for the classification of data, they used Support Vector Machine and Linear Discriminant Analysis, which shows higher accuracy when comparing to other algorithms.


Introduction
Traditionally, human tester involved detecting the contamination level of fruits, vegetables, dairy products, animal foods based on the appearance, colour and aroma. It is an inefficient method. E-nose is a non-destructive method to test the quality of products. In 1982, Dodd and Persuade introduced E-nose system. It can detect the volatile compounds emitted from various source. E-nose consists of an array of sensors. Sensors are selected based on applications. E-nose responses are collected using microcontrollers like Arduino, Raspberry pi. PCA (Principal Component Analysis) is one of the feature extraction techniques to choose the important feature in Enose data. Machine learning algorithms such as Naive Bayes (NB), Support Vector Machine (SVM), Linear Regression (LR), Logistic Regression, Linear Discriminant Analysis (LDA), Support Vector Regression (SVR), Partial Least Square Regression (PLSR), Artificial Neural Networks (ANN) and K-Nearest Neighbour (KNN) are applied to the datasets and performance are analysed.
The electronic nose has a significant impact on outdoor monitoring. Humans can't work in specific odor analysis like detecting the gases emitted from the wastewater treatment plant and detecting toxic gases in the air. But E-nose provides an efficient approach to outdoor monitoring. Electronic nose was used in predicting the ripening stage of fruit and also detect the quality of fruit. Nowadays, Human disease is detected using E-nose from the breath sample (Goor et al., 2018). E-nose applications are outlined in the Table 1 and hierarchical chart is shown in Fig. 1. The performances of machine learning algorithms are given in Table 2.

Electronic Nose in Outdoor Monitoring
The electronic nose can be employed in the area where humans can't detect the odor. For example, wastewater treatment plant emits malodorous, which cause serious health issues on humans. Blanco-Rodríguez et al. (2018) Suggested a method for characterizing the hazardous gas emitted from the wastewater treatment plant using electronic-nose. In their experiment, odor samples collected from the six stages of the plant. They performed signal filtering, normalization and feature extraction with the dataset. They established the correlation between Enose response and olfactometry analysis by using Partial Least Square Regression (PLSR).
Nowadays, Air pollution is a significant concern in this world, which has a severe impact on human health. Jasinski et al. (2018) suggested a method for predicting the toxic gases present in the air. They used three types of electronic nose system depending upon the type of sensor used semiconductor sensor, amperometric sensor and third one combination of both sensors. They had collected the data in 1 minute from each electronic nose and applied PLS regression and SVM. They compared the performance of all three systems. They measured the concentration of the four gases carbon monoxide (CO), Nitrogen dioxide (NO 2 ), Sulphur dioxide (SO 2 ) and ozone (O 3 ) Among the three types of Electronic nose, a combination of both sensors provides better results.    Feature extraction techniques were applied to the datasets to select critical features. Most commonly used feature extraction is PCA. Kong et al. (2019) Proposed a new procedure for feature extraction Weighted Summation (WS). Gaseous pollutants emitted from the pig farm can affect the environment and also a severe impact on the health of humans. They had collected data from pig farm data using E-nose. And they had applied Weighted Summation to datasets. They compared the results with existing feature extraction algorithms and weighted summation showed higher accuracy. Herrero et al. (2016), had proposed classification of water pollutants using wireless portable electronic noses.

Disease Detection Using Electronic-Nose
Detection of soft rot disease in potato (Rutolo et al., 2018). In this work, they used WOLF 4.1 (Warwick OLFaction), electronic nose for predicting the contamination of Pectobacterium carotovorum in potato. Data analyzed using algorithms such as LDA, SVM, Naive Bayes, ensemble methods. Jia et al. (2019) suggested PEN3 was used to detect the level of contamination of moldy apple inoculated with Penicillium expansum and Aspergillus niger. Dataset collected from both the apple inoculated with and without bacteria. Four machine learning algorithms were used to analyze the data such as Back Propagation Neural Network (BPNN), SVM and radial basis function neural network (RBFNN), Linear Discriminant Analysis (LDA). They found that the BPNN shows higher accuracy among all the algorithms.
Electronic nose not only detects the human disease but also recognizes the condition in plant and animal (Wilson, 2018). Electronic nose detects the disease based on the Volatile Compound (VOC) emitted from the sample.

Discrimination of Substance Using Electronic Nose
Zhan (2018) suggested a method for discriminating 12 different categories of Chinese herbal medicine using electronic nose. In this work, the electronic nose consists of 16 TGS (Taguchi gas sensors) made in Japan. Data acquired from 600 samples one by one. And they preprocessed the dataset and applied machine learning algorithm such as Support Vector Machine (SVM), Decision Tree (DT), Naive Bayes (NB), Conformal Prediction K-Nearest Neighbour (CP-KNN), Artificial Neural Network (ANN), Linear Discriminant Analysis (LDA). Among them, SVM and LDA show higher accuracy with 98.94% and 98.33%. But the KNN (CP-1NN and CP-3NN) provides the prediction reliability. Centonze et al. (2019) in their paper, they used electronic nose to discriminate the different varieties of oranges belongs to three regions Italy, South Africa and Spain. They applied multivariate statistical models to the E-nose response. And the LDA provides better prediction accuracy.

Fruits and Vegetables Quality Prediction Using Electronic Nose
The freshness of cherry tomato was evaluated in their work (Feng et al., 2018). They divided tomatoes into two groups. Two groups of cherry tomato treated with and without high-pressure Argon in the ratio of 0.4, 0.8 and 1.2 Mpa. They used Argon gas as a preservative. And data collected from both groups of tomato. After data acquisition, they used Partial Least Square Regression (PLSR) and Single Layer Feed Forward Neural Network. And they made the comparative study between the two algorithms. Based on the E-nose data, they classified the cherry tomato into four groups.
Jia et al. (2019) they inoculated Golden delicious apples with Penicillium expansum and Aspergillus niger. The PEN3 was used to identify the fresh and moldy apples (apples inoculated with Penicillium expansum and Aspergillus Niger). Gas emitted from apple matches with sensors available in PEN3. E-nose response was analyzed using LDA, BPNN, SVM, Radial Basis Function Neural Network (RBFNN). BPNN provided the best accuracy among all methods. E-nose identifies not only fresh apple and but also discriminated moldy apples inoculated with Penicillium expansum and Aspergillus niger. Brezmes et al. (2000) suggested a method for monitoring fruit ripeness using E-nose. For this purpose, they used Peach, Apple and Pear. They observed the fruit from the day of harvest until it became overripe. They used the neural network to classify the fruits into different stages of green, ripe and overripe. Peach and Pear shows higher accuracy than Apple. Rayappan et al. (2018) proposed a method for predicting the quality of fresh and contaminated Royal delicious apple. E-nose consists of six readymade sensors and integrated into a single circuit. E-nose values recorded at the sample time of 1s. They applied PCA and the ward's method of cluster analysis to data. This method found a correlation between the different stages of apple. Wen et al. (2019) sweeping E-nose was used to identify and detect the presence of Bactrocera dorsali in citrus fruits. They applied PCA and LDA to the E-nose data. They found that LDA provides better performance with an accuracy of 98.21% and discriminate the different stages of incubation and invasion. Xu et al. (2016) used PEN3 E-nose predicts to litch quality. Litch quality was detected in different environments (normal temperature, refrigerator, controlled condition). The hardness of the litch sample was obtained from E-nose. LDA, Canonical Correlation Analysis (CCA), BPNN and BPNN-Partial Least Squares Regression (BPNN-PLSR), were employed to sensor data. They found that BPNN-PLSR effectively predicted hardness of litchi under refrigerator storage conditions and a controlled-atmosphere environment, but it was poor in normal storage. Liu et al. (2018) proposed method for identifying fungal contamination in peaches using a PEN3. Peaches were inoculated with spoilage fungi such as Botrytis cinerea, Monilinia fructicola and Rhizopus stolonifer and then stored for long periods. E-nose was used to analyze volatile compounds generated in the fungi-inoculated peaches. Data pre-processing was done by Standard Normal Variate(SNV) to eliminate the signal drift. PLSR was applied to classify the fungi species.
They successfully discriminated Peach sample inoculated with fungi after 48hours of storage. The statistical results showed Volatile compounds of peach was affected by the total count and species of fungi. Sanaeifar et al. (2016) suggested E-nose for predicting the properties of banana. They made the comparison between E-nose response data and quality indices of banana was applying by Partial Least Square, Multiple Linear Regression and Support Vector Regression. They found pH and Titratable Acidity of quality indices of banana showed poor correlation with E-nose response.
They found that the quality indices of banana predicted using SVR were better than other algorithms. They discovered that E-nose was reliable to predict the properties of banana.

Quality Prediction of Beverages
Electronic nose provides an efficient method for checking the quality of beverages. (Rodriguez et al. 2019) Suggested design for identifying wine quality is analyzed using the electronic nose. E-nose response of wine collected at the sampling frequency of 18.5 Hz during 180 seconds. And they classified the dataset into high quality, average quality and low quality. They identified the threshold of wine quality. Xu et al. (2018) Proposed a method for predicting the quality of Tea by E-nose and Computer Vision System (CVS). E-nose was used to categorize the quality of Tea based on the aroma. CVS analyzed the appearance of the Tea; it captures the image of Tea and extracts information such as size and color. They made a comparison between E-nose signals and CVS signals. They developed the data fusion strategy combining both the methods of E-nose and CVS to predict the quality of the E-nose. Yasuo et al. (2019) proposed a method to analyze the coffee sample of six types using PEN2 (seven MOS sensors). Common dimension analysis was used to reduce the large datasets and LDA was applied to classify the samples. This method is efficient to classify the coffee samples. Omatu and Yano (2016) Designed the E-nose (14 sensors) system to discriminate Tea or coffee based on aroma emitted from the samples of different concentrations. They used the Learning Vector Quantization neural network to analyze the data. After reducing the E-nose noise, obtain the maximum value of odor. Normalize the datasets; values were affected due to different concentration level. E-nose, along with learning vector quantization neural network was efficient to discriminate between Tea or coffee.

Electronic Nose in Animal Food Analysis
Animal foods are highly perishable. Electronic nose used to detect the quality and identify the adulteration of animal food. In their study, (Wang, 2019) suggested a method for detecting adulteration of mutton with duck meat using E-nose. They performed Multivariate data analysis by using linear regression, Fisher Linear Discriminant Analysis (FLDA) and Multilayer Perceptron Neural Networks analysis (MLPN) on E-nose signals. Güney and Atasoy (2015) designed the E-nose (8 sensors) to discriminate between different species of fish. After data acquisition, data pre-processing was done by signal pre-processing, normalization and feature extraction. The proposed Hybrid algorithm shows higher accuracy when compared to all methods KNN, NB, LDA.

Electronic Nose for Edible Oil
Nowadays, E-nose can also be used in detecting the quality of the oil. (Majchrzak et al., 2017) proposed method for determination of the product's geographical origin and further in the detection of adulteration as well as deterioration caused by external factors. E-nose used to discriminate between non-oxidized and oxidized oils. They used Cluster Analysis (CA), PCA and LDA to E-nose data. LDA produced better results than CA and PCA in discriminating between oxidized oil and non-oxidized oil. Upadhyay et al. (2017) designed the E-nose (18 Metal Oxide Semiconductor sensors) used for monitoring the disposal time of deep-fried sunflower oil stabilized with natural oxidants.
Rapeseed is one of the sources of edible oil (Gancarz et al., 2017) agrinose used for detecting the quality of rapeseed. Agrinose(eight MOS sensors), sensors were selected based on lower power consumption, low susceptibility to humidity and temperature. The quality of rapeseed was detected using Enose during 31 days of storage was studied. Agrinose used for the examination of Colony Forming Unit, Ergosterol content, Fourier Transform Infrared Spectroscopy and Volatile Organic Compounds. Agrinose monitored the microbiological count of rapeseed during the first twelve days of storage. PCA had shown a correlation between Ergosterol content, sensor response, Colony Forming Unit and the type of microflora.

Electronic Nose in Dairy Product
E-nose provides the best quality assessment of the dairy product. It is used to detect the adulteration in milk and classify the cheese according to manufacturing techniques. Cevoli et al. (2011) proposed E-nose to classify the pecorino cheese. They used ANN and E-nose data feed as input to the ANN. After feature extraction using PCA, feed the features to the ANN. They made a comparison between the before and after feature extraction. Before feature extraction, it showed higher accuracy. Feature extraction was not efficient in this method. Tohidi et al. (2017) suggested a technique using the electronic nose to detect adulteration in raw milk. After data acquisition, data pre-processing involves steps such as baseline correction, compression and normalization. PCA was used to reduce the dimensionality of data. They used multivariate data analysis such as LDA, SVM to analyze E-nose. SVM showed higher accuracy. This method found the adulteration and percentage of adulteration by using chemometrics.

Conclusion
In this review, we have given a summary of E-nose applications in various fields, finding the adulteration in mutton, milk and predicting the quality of fruits. It can also classify the food based on the aroma emitted from food. Most of the classification techniques provide more than 90% accuracy. Among them, SVM and LDA provide 100% prediction rate and Support Vector Regression provides lower RMSE and higher R 2 . BPNN delivers the desired performance in most of the cases.