The polymer quartz piezoelectric crystal sensor based e-nose^[1] is a machine olfaction device to identify and monitor a variety of chemical gases of interest. It can imitate human senses by using polymer quartz piezoelectric crystal sensor arrays and pattern recognition system. Recently the developments of the polymer quartz piezoelectric crystal sensor, a quality sensitivity sensor outputting vibration frequency signal values, significantly promoted the improvement of the e-nose technology.^[2–4] However, the e-nose is still in the early stages of application.

An effective pattern recognition system in the e-nose is needed to distinguish among different kinds of frequency signal values of testing materials. In fact, researchers have proposed many pattern recognition systems to improve the identification performance of the e-nose in market. For example, Chang and Freeman,^[5] and Chang et al.^[6] modeled a mammalian neural olfactory system that was characterized anatomically by modifiable synaptic connections, and physiologically by the spatiotemporal interactions among neural ensembles under external inputs. Ekachai and Nagle^[7] dealt with selecting sensors for e-noses to make a portable device with fast response time, small size and reduced cost. Zhang et al.^[8] presented an on-line counteraction of unwanted odor interference based on pattern recognition for the first time. Zhang et al.^[9] studied the concentration estimations of indoor contaminants for air quality monitoring in dwellings by using chaos based on an optimal artificial neural network integrated into their self-designed portable e-nose instrument. Pawel and Wojciech^[10] presented two innovative systems based on the feed-forward and recurrent neural network used for qualitatively analyzing the specimens of different fruit teas. Kizil et al.^[11] developed an artificial neural network (ANN) model to identify the Salmonella enterica in poultry manure, and the accuracy is 94%. Although many successful pattern recognition systems have been designed for the e-nose, they cannot be used for the polymer quartz piezoelectric crystal sensor based e-nose directly. This is mainly because the signals, captured by the polymer quartz piezoelectric crystal sensor array, are nonlinear, and the sensors have cross sensitivity to the testing materials. The existing pattern recognition systems designed for the e-nose cannot extract the characteristic information about the polymer quartz piezoelectric crystal sensor effectively, and they must be updated and improved.

In our previous work, Gu and Li^[12] presented a new pattern recognition system based on moving average and linear discriminant analysis (LDA), which could be used to process the original signal values of the new polymer quartz piezoelectric crystal air-sensitive sensor system. But only three kinds of Chinese spirits could be effectively identified and the adaptability of the pattern recognition system needs to be improved. In this paper, we propose another pattern recognition system for the polymer quartz piezoelectric crystal sensor based e-nose. For the vibration frequency signal values of six kinds of Chinese spirits captured by the e-nose, the proposed new pattern recognition system could distinguish from them effectively. The new polymer quartz piezoelectric crystal sensor based e-nose, assembled with the new pattern recognition system, provides us with a new method to distinguish among Chinese spirits.

To identify different kinds of Chinese spirits, a new pattern recognition system for the polymer quartz piezoelectric crystal sensor based e-nose is presented. First, the vibration frequency signal values are obtained by the polymer quartz piezoelectric crystal sensor array. Then for each set of vibration frequency signal values, seven characteristic values are extracted and the principal components analysis (PCA) method is used to reduce the dimension of the characteristic values. Finally the back propagation (BP) artificial neural network method is used to identify different kinds of Chinese spirits.

2.1. Signal acquisition of Chinese spirits

In this paper, we use the novel polymer quartz piezoelectric crystal sensor array to obtain the signal data of Chinese spirits, which is composed of different kinds of affinity-type polymer quartz piezoelectric sensors and very sensitive to the liquor product material. As shown in Fig. 1(a), the sensor is composed of polymer coating, quartz piezoelectric crystal plate, and polymer coating. The sensor is a quality-sensitive sensor, and its resonance will be disturbed by adding or removing a small mass of the polymer coating due to the adsorption or desorption to the testing materials, respectively. The vibration frequency values will be captured due to piezoelectric property of quartz crystal plate. The photo of the polymer quartz piezoelectric crystal sensor is shown in Fig. 1(b). We can prepare different kinds of sensors via adhering different kinds of polymer coatings.

Fig. 1.

	Figure Option ViewDownloadNew Window
	Fig. 1. (a) Photo and (b) structure of polymer quartz piezoelectric crystal sensor.

After studying the components of six kinds of Chinese spirits which are shown in Table 1, we select eight kinds of polymers to prepare eight kinds of polymer quartz piezoelectric crystal sensors. The selected polymers are shown in Table 2. As shown in Fig. 2, these sensors are assembled into an 8-channel sensor array. The vibration frequency signal values of the sensor array are captured 100 times per min.

Table 1.

Table 1.

Table 1. Six kinds of Chinese spirits. . No Chinese spirit Proof Production date [year] Origin 1 Shanxi spirits 106 2011 Xinghua village in Shanxi province 2 GuoJiao1573 106 2011 Luzhou in Sichuan province 3 Jiannanchun 106 2011 Mianyang in Sichuan province 4 Niu Lanshan 106 2011 Beijing in China 5 Wuliangye (spirit) 106 2011 Yibin city in Sichuan province 6 Jiugui Liquor 106 2011 Xiangxi autonomous prefecture, Hunan province

Table 1. Six kinds of Chinese spirits. .

Table 2.

Table 2.

Table 2. Selected polymers for constructing quartz piezoelectric crystal sensors. . No Sensor 1 PVC 2 polyamide 3 polyethylene + AgCl 4 polytef 5 AgCl 6 azithromycin 7 CuCl2 + PE 8 CuCl2 + AgCl + PE

Table 2. Selected polymers for constructing quartz piezoelectric crystal sensors. .

Fig. 2.

	Figure Option ViewDownloadNew Window
	Fig. 2. Photo of the 8-channel sensor array.

In this paper, six kinds of Chinese spirits are used to testify the performance of the designed new pattern recognition system in the e-nose. First, the 8-channel polymer quartz piezoelectric crystal sensor array is used for signal acquisition. For each kind of Chinese spirit, we capture vibration frequency signals for 2 min at room temperature in a clean room and 200 vibration frequency signal values are acquired. Such a manipulation is repeated 10 times. The signal values from 1st to 5th manipulation are used for training and the signal values from 6th to 10th manipulation are used for testing. Then six kinds of vibration frequency signal values of Chinese spirits can be acquired using the above method. Because of the limitation of length, a set of vibration frequency signal values of six selected kinds of Chinese spirits by sensor-1 are listed in Table 3, and their change tendencies are shown in Figs. 3–8. From Figs. 3 to 8, we can see that the changing trends of six kinds of Chinese spirits are similar.

Table 3.

Table 3.

Table 3. Frequency values of six selected kinds of Chinese spirits, detected by sensor-1. . No Shanxi spirits GuoJiao1573 Jiannanchun Niu Lanshan Wuliangye (spirit) Jiugui Liquor Chinese spirit 1 9996217 9996208 9996201 9995406 9995497 9996203 2 9996216 9996210 9996201 9995407 9995498 9996202 3 9996212 9996206 9996203 9995406 9995498 9996205 4 9996205 9996205 9996195 9995406 9995499 9996200 5 9996198 9996205 9996188 9995400 9995498 9996194 6 9996188 9996200 9996185 9995398 9995499 9996187 7 9996181 9996195 9996178 9995396 9995497 9996182 8 9996176 9996190 9996176 9995397 9995494 9996178 9 9996173 9996187 9996172 9995396 9995493 9996180 … … … … … … … 200 9996184 9996178 9996179 9995391 9995491 9996186

Table 3. Frequency values of six selected kinds of Chinese spirits, detected by sensor-1. .

Fig. 3.

	Figure Option ViewDownloadNew Window
	Fig. 3. Frequency values of Shanxi spirits.

Fig. 4.

	Figure Option ViewDownloadNew Window
	Fig. 4. Frequency values of GuoJiao1573.

Fig. 5.

	Figure Option ViewDownloadNew Window
	Fig. 5. Frequency values of Jiannanchun.

Fig. 6.

	Figure Option ViewDownloadNew Window
	Fig. 6. Frequency values of Niu Lanshan.

Fig. 7.

	Figure Option ViewDownloadNew Window
	Fig. 7. Frequency values of Wuliangye (spirit).

Fig. 8.

	Figure Option ViewDownloadNew Window
	Fig. 8. Frequency values of Jiugui Liquor.

For each set of vibration frequency signal values, seven characteristic values of six kinds of Chinese spirits are computed by using the signal values from 1st to 5th manipulation. Averaging each kind of characteristic value, we could obtain the characteristic values of six kinds of Chinese spirits. Taking the sensor-1 for example, the average characteristic values are shown in Table 4.

Table 4.

Table 4.

Table 4. Average characteristic values of six kinds of Chinese spirits (sensor-1). . Chinese spirit A/Hz RMS/Hz Sf Cf If CLf Kv Shanxi spirits 9996200 9996200 1.0000 1.0000 1.0000 1.0004×10−7 7.8691×10−24 GuoJiao1573 9996200 9996200 1.0000 1.0000 1.0000 1.0004×10−7 4.0899×10−24 Jiannanchun 9996200 9996200 1.0000 1.0000 1.0000 1.0004×10−7 4.1867×10−25 Niu Lanshan 9995400 9995400 1.0000 1.0000 1.0000 1.0005×10−7 5.6810×10−26 Wuliangye (spirit) 9995500 9995500 1.0000 1.0000 1.0000 1.0005×10−7 2.0544×10−26 Jiugui Liquor 9996200 9996200 1.0000 1.0000 1.0000 1.0004×10−7 2.1594×10−24

Table 4. Average characteristic values of six kinds of Chinese spirits (sensor-1). .

As shown in Table 4, S_f, C_f, I_f, and CL_f have similar values, but A, RMS, and K_v have different values. These characteristic values indicate that the characters of the six kinds of Chinese spirits are similar, so we use the PCA method to extract more distinctive features. With the PCA method, we can obtain the first principal component of each kind of the seven characteristic values of Chinese spirits, and the results are shown in Table 5.

Table 5.

Table 5.

Table 5. Principal components of seven characteristic values. . Chinese spirit Sensor-1 Sensor-2 Sensor-3 Sensor-4 Sensor-5 Sensor-6 Sensor-7 Sensor-8 Shanxi spirits 341.7683 –40.9658 –63.5901 –482.2617 119.7976 –302.9743 –14.7917 604.6131 GuoJiao1573 350.0796 21.5269 –36.6451 –503.0959 115.4177 –311.8485 –11.0455 594.4576 Jiannanchun 357.2256 –34.5113 –54.6155 –494.3843 127.4046 –324.9568 –25.2810 594.9893 Niu Lanshan –761.2023 48.3277 82.5689 1007.1188 –250.7080 630.1365 43.2490 –1230.4733 Wuliangye (spirit) –633.8962 36.34222 128.9848 960.9800 –231.1791 618.8752 16.3323 –1164.3036 Jiugui Liquor 346.0251 –30.71978 –56.7029 –488.3569 119.2672 –309.2322 –8.4631 600.7169

Table 5. Principal components of seven characteristic values. .

As shown in Fig. 9, we insert the feature values of six kinds of Chinese spirits into a radar chart where each coordinate represents one kind of sensor. We can know that there are differences between different kinds of Chinese spirits. So we can use the principal components to represent the distinctive feature of each kind of Chinese spirit.

Fig. 9.

	Figure Option ViewDownloadNew Window
	Fig. 9. Radar chart of six kinds of Chinese spirits’ feature values of Chinese spirits.

Finally, the BP neutral network is trained in the Matlab software. The number of hidden layer nodes is set to be 6, the maximum training time is set to be 1000, the minimum mean square error is set to be 1 × 10⁻⁸, and the training goal is set to a 6×6 unit matrix. The training performance of the BP neutral network is shown in Fig. 10. The best training performance is reached at epoch 15. After training, the network weight and threshold of the BP neutral network can be computed, and then the structure of the BP neutral network can be determined. The trained BP neutral network can be viewed as a classifier, and we use it to identify different kinds of Chinese spirits. In this paper, there are five testing samples for each kind of Chinese spirit, and the total number of testing samples is 30. The recognition rates are shown in Table 6. For Shanxi spirits, Jiannanchun, Niu Lanshan and Jiugui Liquor, the recognition rate of the new pattern recognition system is 100%. For GuoJiao1573 and Wuliangye, the recognition rate of the new pattern recognition system is 80%. The recognition of six kinds of spirits is 93.33%. These experimental results indicate that our proposed pattern recognition can distinguish among the six kinds of Chinese spirits effectively.

Fig. 10.

	Figure Option ViewDownloadNew Window
	Fig. 10. Training performance of the BP neutral network.

Table 6.

Table 6.

Table 6. Recognition results. . Chinese spirit Shanxi spirits GuoJiao1573 Jiannanchun Niu Lanshan Wuliangye (spirit) Jiugui Liquor Total Recognition rates 100% 80% 100% 100% 80% 100% 93.33%

Table 6. Recognition results. .

This paper presents a new pattern recognition system to distinguish among six kinds of Chinese spirits by using a polymer quartz piezoelectric crystal sensor e-nose. First, seven characteristic values of six kinds of vibration frequency signal values of Chinese spirits, i.e., A, RMS, S_f, C_f, I_f, CL_f, and K_v are extracted. Then the PCA method is used to reduce the dimension of the characteristic values. Finally, the first principal components are used to train the BP neutral network and the trained BP neutral network is used to recognize Chinese spirits. The recognition results show that the recognition rates of six kinds of Chinese spirits are 93.33%. Our proposed new pattern recognition system provides us with a highly innovative method to recognize Chinese spirits.