VOC sensor

Overview

In our project, we are exploring a new way to sense the health condition of crops which will further send this information (healthy or not) to both human and our T.atroviride. In this way, we can take actions on time and achieve the goal of prevention some phytopathogens in advance. Meanwhile, this method should be conveniently used in variety range of plants.

We finally build up an e-nose (electronic nose) system to sense the VOC (volatile organic compounds) emitted by the plants. Afterwards, using machine learning method, we achieved a high accuracy rate, which is over 85%, when sensing the tobaccos infected by P.nicotianae. That is to say: our device successfully distinguishes whether the tobaccos are ill by "smelling them".

Background

Plants will release a variety of volatile organic compounds (VOC) to resist some infected phytopathogens, which can lead to a change of VOC composition around the plants[1]. Firstly, we verified the VOC differences between healthy tobaccos and tobaccos infected by P.nicotianae by GC-MS.

Fig.1 GC-MS analysis of VOCs emitted by healthy tobaccos and tobaccos infected by P.nicotianae(Thanks Hunan Tobaccos Bureau for providing this data).

We decide to use ten highly sensitive CMOS(Complementary Metal Oxide Semiconductor) gas detector to catch the VOC information, each detector is sensitive to one kind or type of VOCs. E-nose can’t sense a specific VOC like mass spectrometers did. instead, it catches the overall characteristics of VOCs as "fingerprint".

Fig.2 A photo of our gas sensor chamber.

Fig.3 The response curve of gas sensors (from ill tobacco)and the stability curve (40min in air)

We use full automatic headspace sampling method to detect the VOCs: Choose six basins of tobaccos infected by P.nicotianae for 5 days and another six basins of healthy tobaccos, put them into two boxes and sealed them. After 30 minutes, we start the measurement.

Our device is controlled by Arduino Single Chip Microcomputer(SCM) in order to achieve automatic measurement, and we standardize the testing steps to ensure the data between every groups is comparable.

Fig.4 An illustration of our standardized measuring steps

Click here to see more information about our device.

We build median filter algorithm in SCM to remove some outliers. Then, we preprocess the raw data after getting them: First, we identify and remove the base line value; Then, we read the response value on 1min, 2 min, the integral average value and maximum value of the response curve.

We have measured 17 groups of healthy tobaccos and 18 groups of infected ones. Click here to download our raw data and preprocessed data(这里链接到两个excel文件）

Fig.5 The integral average value of 35 response curve.

After preprocessing the data, we apply some algorithms to do the classification work. We employ Decision Tree, Multi-Layer Perception algorithm, and Leaner Model. Finally, we achieve more than 85% accuracy rate of distinguishing whether the tobaccos are infected by P.nicotianae.

Moreover, based on the result of the modeling, 4 CMOS detectors are enough to make a judgement for the health condition of tobaccos, which means we can further reduce the cost of our device.

Click here to see the details of our modeling process.

Conclusions

Our method achieves more than 85% accuracy rate on distinguish the health state of tobaccos, and we find we only need four detectors for our devices to make such judgments. Moreover, the VOC data can be sent to PC port, we also build a web app to achieve real time monitoring. In the future, we plan to make the web app possible to show the healthy condition of plants in real time.

On the other hand, the sensitivity of our device is still limited when facing real world conditions. It is hard to fit different conditions in different field. Therefore, we come up with a plan to solve this problem, Click here to see our improvements.