Because the sensor response is dependent on its operating temperature, modulated temperature operation is usually applied in gas sensors for the identification of different gases. In this paper, the modulated operating temperature of microhotplate gas sensors combined with a feature extraction method based on Short-Time Fourier Transform (STFT) is introduced. Because the gas concentration in the ambient air usually has high fluctuation, STFT is applied to extract transient features from time-frequency domain, and the relationship between the STFT spectrum and sensor response is further explored. Because of the low thermal time constant, the sufficient discriminatory information of different gases is preserved in the envelope of the response curve. Feature information tends to be contained in the lower frequencies, but not at higher frequencies. Therefore, features are extracted from the STFT amplitude values at the frequencies ranging from 0 Hz to the fundamental frequency to accomplish the identification task. These lower frequency features are extracted and further processed by decision tree-based pattern recognition. The proposed method shows high classification capability by the analysis of different concentration of carbon monoxide, methane, and ethanol.
An electronic nose is an instrument that imitates the functionalities of biological olfactory, which typically consists of an array of chemical sensors (usually gas sensors) and a pattern recognition system [
The selectivity and sensitivity of most gas sensors are dramatically dependent on the operating temperature, since the reaction rate of different analytes and the stability of surface-adsorbed oxygen species are a function of temperature. Operating temperature modes of sensors can be divided into two categories: (i) the constant operating temperature (e.g., the heating voltage is set to 5 V) and (ii) temperature modulation (e.g., the sensors are driven by self-adapted or periodic heating voltages) [
In addition, gas molecules are carried by air flow and distributed by turbulence [
In this paper, the STFT is calculated over the responses of SnO2 microhotplate gas sensors, and the relationship between the STFT spectrum and sensor response is further explored. The optimal features tend to be contained in the lower frequency responses of the sensor close to direct current (DC). Therefore, features are extracted from the STFT amplitude values at the frequency ranging from 0 Hz to the fundamental frequency to fulfill the identification task. The gas identification effect of the proposed method is demonstrated by the analysis of carbon monoxide, methane, and ethanol with different concentrations. Various waveform types and periods of the heating voltage are studied, and results show that it is possible to find the optimal features to obtain excellent identification of the gases studied regardless of their concentrations. The structure of the paper is as follows. Section
The four-element microhotplate gas sensor arrays are fabricated using complementary metal-oxide semiconductor (CMOS) and post-CMOS technology [
The photos of microhotplate gas sensor arrays. (a) SEM photo of a four-element gas sensor array. Each sensor element is a square freestanding membrane supported by four arms. A tungsten thin film resister is embedded in the membrane acting as both heater and thermometer. The SnO2 sensitive film is sputtered on the top of the microhotplate. (b) Microhotplate gas sensors packaged with DIP16 or TO5.
The main body of the sensor is a square freestanding membrane supported by four bridge arms. In the center membrane, a tungsten thin film resister that has the snake shape is designed to monitor the temperature of the microhotplate as well as to heat up the membrane. The SnO2 sensitive film is sputtered on the top of the microhotplate, and Pt catalyst with a thickness of 1 nm is sputtered for improving its selectivity.
The thermal efficiency of the microhotplate is subsequently acquired by electrically heating with a digital source meter (Keithley 2400) and measuring the resistance values. The measurement results show that the thermal impedance of the microhotplate with electrodes and sensitive materials is about 16°C/mW. The thermal response of the coated membranes is near 8 ms (10% to 90% rise time) when working at 300°C [
Experimental setup and the measurement circuit are showed in Figure
Experimental setup for data acquisition. (a) The gas mixture is injected to the testing chamber at a constant flow rate, with gas components and concentrations controlled by the MFCs. The gas sensors are placed in the testing chamber, and the data is recorded by computer. (b) The measurement circuit. Each gas sensor has two resisters.
The sensor microarray is used with dynamic modulation of the heating voltage that is generated by a programmable DC voltage source (HP6626A) to classify and identify three reducing gases: Methane: 1000 ppm, 2000 ppm, 3000 ppm, and 4000 ppm. Carbon monoxide: 50 ppm, 100 ppm, 150 ppm, and 200 ppm. Ethanol: 30 ppm, 40 ppm, 50 ppm, and 60 ppm.
Dynamic modulation waveforms of the heating voltage are sinusoidal, rectangular, sawtooth, and triangular waveform. Each heating waveform has 8 modulation periods, 4 s, 10 s, 20 s, 30 s, 40 s, 50 s, 60 s, and 80 s, and the operating temperature ranges from 200 to 300°C. The test procedure is described as follows.
Dry air at a constant flow rate of 330 sccm by the flow system is circulated through the testing chamber for 1200 s to measure the baseline steady-state sensor response.
The gas with the desired concentration is injected into the testing chamber for 600 s.
The testing chamber is cleaned with dry air for 900 s. Then, the measurement steps are replicated for subsequent measurements.
Table
The dataset in detail.
Heating waveform and period (s) | Number of samples | |||
---|---|---|---|---|
Methane | Carbon monoxide | Ethanol | Total | |
Sinusoid | 128 | 128 | 128 | 384 |
4, 10, 20, 30, 40, 50, 60, and 80 | ||||
|
||||
Rectangle | 128 | 128 | 128 | 384 |
4, 10, 20, 30, 40, 50, 60, and 80 | ||||
|
||||
Sawtooth | 128 | 128 | 128 | 384 |
4, 10, 20, 30, 40, 50, 60, and 80 | ||||
|
||||
Triangle | 128 | 128 | 128 | 384 |
4, 10, 20, 30, 40, 50, 60, and 80 |
The typical output voltages and heating voltage. (a) The output voltages when the sensor is exposed in three analytes: 50 ppm carbon monoxide, 2000 ppm methane, and 30 ppm ethanol, respectively. (b) The heating voltage is sawtooth modulation waveform at 40 s periods.
Short-Time Fourier Transform (STFT) is a method that FFT transform is applied after the signal is cut out by the window function. For an arbitrary signal
Then,
In order to reduce the drift of the sensors and the background noise, the sensor responses need to be preprocessed by
Figure
The STFT amplitude spectrums of sensor responses to three analytes modulated with the rectangular modulation waveform at 4 s period. Hann window is used as the window function. The length of the window function is 375. (a) 150 ppm carbon monoxide; (b) 3000 ppm methane; (c) 50 ppm ethanol.
In order to extract the optimal features, the frequency distribution of the sensor responses should be clearly analyzed. As seen from Figure
The frequency distributions of sensor responses (150 ppm carbon monoxide) are mainly composed of a fundamental wave and some harmonic waves.
Suppose
In fact, because of the low thermal time constant, the sufficient discriminatory information is preserved in the envelope of the response curve, but not at higher frequencies. The optimal features tend to be contained in the lower frequency responses of the sensor close to DC. Therefore, features are extracted from the STFT amplitude values at the frequency ranging from 0 Hz to the fundamental frequency
The
The feature vectors for three analytes (150 ppm carbon monoxide, 3000 ppm methane, and 50 ppm ethanol) with sawtooth modulated operating temperature (
Decision tree is one of the most well-known methods used for extracting classification rules from data [
There are many window functions with different shapes. Short-Time Fourier Transform is also defined as Gabor Transform if Gaussian window function is selected. Window function makes the STFT observe the features of the sensor responses from time-frequency domain. Aimed at high time resolution, a narrow window function should be selected. Aimed at high frequency resolution, a wide window function should be selected. Hence, it is very important to select the shape and length of the window function. In this work,
The effects of 8 window functions are tested. The window functions are Boxcar window, Hamming window, Hann window, Gaussian window, Taylor window, Blackman window, Tukey window, and Triangle window.
The specific process of the algorithm for each heating waveform is summarized as follows: (1) select a desired window function (e.g., the hamming window function), and calculate the optimal feature vectors
Meanwhile, the classifier uses 12-fold cross-validation method to test the robustness of the algorithm. The original sample is randomly partitioned into 12 equal sized subsamples. Of the 12 subsamples, a single subsample is retained as the validation data for testing the model, and the remaining 11 subsamples are used as training data. The cross-validation process is then repeated 12 times, with each of the 12 subsamples used exactly once as the validation data. The 12 results from the folds can then be averaged to produce a single estimation.
Tables
The recognition accuracy for the rectangle waveform (decision tree classifier and 12-fold cross-validation are applied to the identification of gases) (%).
Frequency selected (mHz) | Feature selected | Window function | |||||||
---|---|---|---|---|---|---|---|---|---|
Boxcar | Triangle | Blackman | Taylor | Tukey | Hann | Hamming | Gaussian | ||
0.97 |
|
96.88 | 96.88 | 94.79 | 93.75 | 96.88 | 94.79 | 90.63 | 95.83 |
1.94 |
|
97.92 | 93.75 | 95.83 | 96.88 | 97.92 | 92.71 | 95.83 | 96.88 |
2.91 |
|
97.92 | 95.83 | 95.83 | 97.92 | 97.92 | 94.79 | 97.92 | 94.79 |
3.88 |
|
95.83 | 97.92 | 88.54 | 97.92 | 96.88 | 90.63 | 96.88 | 93.75 |
4.85 |
|
94.79 | 95.83 | 92.71 | 97.92 | 92.71 | 98.96 | 94.79 | 96.88 |
5.82 |
|
93.75 | 95.83 | 98.96 | 100 | 95.83 | 97.92 | 100 | 96.88 |
6.79 |
|
92.71 | 94.79 | 97.92 | 92.71 | 91.67 | 96.88 | 90.63 | 96.88 |
7.76 |
|
96.88 | 85.42 | 100 | 92.71 | 86.46 | 89.58 | 94.79 | 96.88 |
8.73 |
|
83.33 | 86.46 | 95.83 | 84.38 | 87.5 | 95.83 | 85.42 | 94.79 |
9.7 |
|
87.5 | 90.63 | 91.67 | 84.38 | 81.25 | 86.46 | 94.79 | 88.54 |
10.67 |
|
79.17 | 90.63 | 88.54 | 76.04 | 84.38 | 84.38 | 87.5 | 84.38 |
The recognition accuracy for the sawtooth waveform (decision tree classifier and 12-fold cross-validation are applied to the identification of gases) (%).
Frequency selected (mHz) | Feature selected | Window function | |||||||
---|---|---|---|---|---|---|---|---|---|
Boxcar | Triangle | Blackman | Taylor | Tukey | Hann | Hamming | Gaussian | ||
0.97 |
|
93.75 | 96.88 | 93.75 | 94.79 | 94.79 | 94.79 | 94.79 | 96.88 |
1.94 |
|
97.92 | 95.83 | 95.83 | 93.75 | 90.63 | 92.71 | 94.79 | 94.79 |
2.91 |
|
96.88 | 96.88 | 95.83 | 92.71 | 93.75 | 93.75 | 97.92 | 95.83 |
3.88 |
|
94.79 | 96.88 | 90.63 | 100 | 97.92 | 93.75 | 94.79 | 94.79 |
4.85 |
|
97.92 | 98.96 | 92.71 | 100 | 97.92 | 94.79 | 96.88 | 95.83 |
5.82 |
|
97.92 | 95.83 | 97.92 | 100 | 90.63 | 100 | 100 | 96.88 |
6.79 |
|
91.67 | 91.67 | 93.75 | 100 | 97.92 | 96.88 | 95.83 | 96.88 |
7.76 |
|
98.96 | 93.75 | 96.88 | 100 | 98.96 | 97.92 | 97.92 | 100 |
8.73 |
|
96.88 | 91.67 | 97.92 | 96.88 | 91.67 | 97.92 | 94.79 | 100 |
9.7 |
|
88.54 | 98.96 | 96.88 | 97.92 | 98.96 | 91.67 | 95.83 | 96.88 |
10.67 |
|
97.92 | 98.96 | 92.71 | 94.79 | 91.67 | 84.38 | 88.54 | 85.42 |
The recognition accuracy for the sinusoidal waveform (decision tree classifier and 12-fold cross-validation are applied to the identification of gases) (%).
Frequency selected (mHz) | Feature selected | Window function | |||||||
---|---|---|---|---|---|---|---|---|---|
Boxcar | Triangle | Blackman | Taylor | Tukey | Hann | Hamming | Gaussian | ||
0.97 |
|
89.58 | 89.58 | 89.58 | 90.63 | 87.5 | 89.58 | 92.71 | 93.75 |
1.94 |
|
93.75 | 87.50 | 90.63 | 86.46 | 88.54 | 91.67 | 90.63 | 92.71 |
2.91 |
|
98.96 | 91.67 | 85.42 | 97.92 | 96.88 | 92.71 | 91.67 | 85.42 |
3.88 |
|
94.79 | 93.75 | 92.71 | 97.92 | 100 | 100 | 96.88 | 96.88 |
4.85 |
|
86.46 | 95.83 | 94.79 | 98.96 | 91.67 | 100 | 97.92 | 92.71 |
5.82 |
|
97.92 | 92.71 | 95.83 | 93.75 | 94.79 | 92.71 | 97.92 | 100 |
6.79 |
|
90.63 | 96.88 | 97.92 | 89.52 | 92.71 | 97.92 | 94.79 | 100 |
7.76 |
|
84.38 | 92.71 | 93.75 | 94.79 | 95.83 | 95.83 | 95.83 | 100 |
8.73 |
|
82.29 | 88.54 | 95.83 | 89.58 | 86.46 | 96.88 | 92.71 | 95.83 |
9.7 |
|
85.42 | 95.83 | 93.75 | 90.63 | 88.54 | 91.67 | 94.79 | 91.67 |
10.67 |
|
90.63 | 98.96 | 92.71 | 92.71 | 87.5 | 96.88 | 96.88 | 96.88 |
The recognition accuracy for the triangular waveform (decision tree classifier and 12-fold cross-validation are applied to the identification of gases) (%).
Frequency selected (mHz) | Feature selected | Window function | |||||||
---|---|---|---|---|---|---|---|---|---|
Boxcar | Triangle | Blackman | Taylor | Tukey | Hann | Hamming | Gaussian | ||
0.97 |
|
84.38 | 96.88 | 95.83 | 91.67 | 87.5 | 92.71 | 96.88 | 94.79 |
1.94 |
|
87.50 | 89.58 | 96.88 | 92.71 | 91.67 | 96.88 | 94.79 | 96.88 |
2.91 |
|
95.83 | 88.54 | 90.63 | 82.29 | 90.63 | 91.67 | 93.75 | 91.67 |
3.88 |
|
92.71 | 91.67 | 84.38 | 88.54 | 97.92 | 85.42 | 88.54 | 85.42 |
4.85 |
|
86.46 | 94.79 | 87.5 | 91.67 | 87.5 | 89.58 | 93.75 | 86.46 |
5.82 |
|
89.58 | 97.92 | 89.58 | 92.71 | 88.54 | 87.50 | 95.83 | 93.75 |
6.79 |
|
93.75 | 91.67 | 85.42 | 89.58 | 93.75 | 90.63 | 82.29 | 86.46 |
7.76 |
|
88.54 | 88.54 | 81.25 | 86.46 | 92.71 | 93.75 | 82.29 | 86.46 |
8.73 |
|
91.67 | 86.46 | 83.33 | 87.5 | 93.75 | 79.17 | 76.04 | 76.04 |
9.7 |
|
89.58 | 75 | 70.83 | 70.83 | 81.25 | 84.38 | 93.75 | 88.54 |
10.67 |
|
92.71 | 77.08 | 68.75 | 76.04 | 71.88 | 78.13 | 79.13 | 83.33 |
In order to analyze the effect of operating temperature waveform, the selected frequencies are extended to the first 50 frequencies that are
The recognition accuracy rates of four modulation waveforms. Decision tree classifier and 12-fold cross-validation are applied for the pattern recognition system.
A feature combination that describes the original data perfectly can make the classifier work more efficiently. In this work, the optimal combinations of 11 features are optimised by Genetic Algorithm (GA). GA is adaptive heuristic search algorithm premised on the evolutionary ideas of natural selection and genetics [
Randomly generate initial population
Compute and save the fitness
Selection, crossover, and mutation are set to roulette wheel selection, single point crossover, and single point mutation, respectively.
Generate
Repeat Step 2 until satisfying solution is obtained.
Tables
The optimal feature combination selected by GA and the recognition accuracy for the rectangle waveform (%).
Window function | Feature selected | Accuracy (%) |
---|---|---|
Boxcar |
|
100 |
|
100 | |
|
||
Triangle |
|
100 |
|
100 | |
|
||
Hamming |
|
100 |
|
100 | |
|
||
Hann |
|
100 |
|
||
Blackman |
|
100 |
|
100 | |
|
||
Taylor |
|
100 |
|
||
Gaussian |
|
100 |
|
100 | |
|
||
Tukey |
|
100 |
|
100 | |
|
100 |
The optimal feature combination selected by GA and the recognition accuracy for the sawtooth waveform (%).
Window function | Feature selected | Accuracy (%) |
---|---|---|
Boxcar |
|
100 |
|
||
Triangle |
|
100 |
|
100 | |
|
||
Hamming |
|
100 |
|
100 | |
|
||
Hann |
|
100 |
|
||
Blackman |
|
100 |
|
100 | |
|
100 | |
|
||
Taylor |
|
100 |
|
100 | |
|
100 | |
|
100 | |
|
||
Gaussian |
|
100 |
|
100 | |
|
100 | |
|
100 | |
|
||
Tukey |
|
100 |
The optimal feature combination selected by GA and the recognition accuracy for the sinusoidal waveform (%).
Window function | Feature selected | Accuracy (%) |
---|---|---|
Boxcar |
|
100 |
|
||
Triangle |
|
100 |
|
100 | |
|
100 | |
|
||
Hamming |
|
100 |
|
100 | |
|
100 | |
|
||
Hann |
|
100 |
|
100 | |
|
||
Blackman |
|
100 |
|
100 | |
|
||
Taylor |
|
100 |
|
100 | |
|
||
Gaussian |
|
100 |
|
100 | |
|
100 | |
|
||
Tukey |
|
100 |
|
100 | |
|
100 |
The optimal feature combination selected by GA and the recognition accuracy for the triangular waveform (%).
Window function | Feature selected | Accuracy (%) |
---|---|---|
Boxcar |
|
100 |
|
||
Triangle |
|
100 |
|
100 | |
|
100 | |
|
||
Hamming |
|
100 |
|
100 | |
|
100 | |
|
||
Hann |
|
100 |
|
100 | |
|
||
Blackman |
|
100 |
|
100 | |
|
||
Taylor |
|
100 |
|
||
Gaussian |
|
100 |
|
100 | |
|
100 | |
|
||
Tukey |
|
100 |
The recognition accuracy of the presented method is compared against the standard Fast Fourier Transform (FFT) and the Discrete Wavelet Transform (DWT). For FFT method, a 1024-point FFT is computed and the absolute values of the eight harmonics corresponding to the modulating frequencies are extracted. Eight harmonic frequencies are 250 mHz, 100 mHz, 50 mHz, 33 mHz, 25 mHz, 20 mHz, 17 mHz, and 12.5 mHz. Put 8 features into the decision tree classifier for pattern recognition. When analyzing Discrete Wavelet Transform, 3-level decomposition and the fourth-order Daubechies (db4) are selected. The third-level decomposition coefficients are extracted as features. PCA is used to reduce the dimensionality of features and keep only the first 10 principal components. Decision tree classifier and 12-fold cross-validation are applied for all of the pattern recognition systems. The recognition accuracy of three reducing gases is showed in Table
The recognition accuracy when FFT and DWT approaches are used (%).
Sawtooth | Triangle | Rectangle | Sinusoid | |
---|---|---|---|---|
FFT | 77.5 | 52.08 | 79.17 | 58.33 |
DWT | 93.75 | 94.79 | 93.75 | 93.75 |
As seen from Table
This paper introduces a novel method to extract optimal features of microhotplate gas sensors that modulated with different frequency operating temperature. The lower frequency amplitudes are extracted by STFT method, and the optimal feature combinations are selected by GA, since gas information tends to be contained in the lower frequencies, but not at higher frequencies.
We then evaluate the performance of our method by using the decision tree classifier and obtain high classification capability. Moreover, it is found that the proposed method is robust against not only dynamical heating frequency changes, but also different concentration levels. Therefore, we conclude that the proposed method could improve the recognition performance of temperature modulated microhotplate gas sensors.
The authors declare that there is no conflict of interests regarding the publication of this paper.
This work was financially supported by the Natural Science Foundation of China (Project nos. 61174007; 61274076; 61131004; 61307041).