Highly selective multiple gases detection using a thermal- conductivity-based MEMS resonator and machine learning

(1)

conductivity-based MEMS resonator and machine learning

Item Type Article

Authors Yaqoob, Usman;Lenz, Wagner B.;Alcheikh, Nouha;Jaber, Nizar;Younis, Mohammad I.

Citation Yaqoob, U., Lenz, W. B., Alcheikh, N., Jaber, N., & Younis, M. I.

(2022). Highly selective multiple gases detection using a thermal- conductivity-based MEMS resonator and machine learning. IEEE Sensors Journal, 1–1. https://doi.org/10.1109/jsen.2022.3203816 Eprint version Post-print

DOI 10.1109/jsen.2022.3203816

Publisher Institute of Electrical and Electronics Engineers (IEEE)

Journal IEEE Sensors Journal

Rights (c) 2022 IEEE. Personal use of this material is permitted.

Permission from IEEE must be obtained for all other users, including reprinting/ republishing this material for advertising or promotional purposes, creating new collective works for resale or redistribution to servers or lists, or reuse of any copyrighted components of this work in other works.

Download date 2023-12-04 18:02:03

Link to Item http://hdl.handle.net/10754/681237

(2)

Abstract—There is an urgent need to develop innovative and highly selective gas sensors for environmental, residential, and industrial applications. Here, we propose a highly selective multiple gases detection system using an electro-thermally heated silicon micro-resonator and machine learning algorithms. The device is based on the cooling/heating effect of gases on the thermal stresses of a resonating structure. As a case study, we demonstrate sensitive and selective responses towards He, Ar, and CO2. To generate the unique signature markers for each

gas, multiple datasets are collected at three different concentration levels (4%, 10%, and 25%). A principal component analysis (PCA) is conducted using seven customized highly significant and unique signature markers. The markers are obtained through the data processing from the response of all gases. Quadratic discriminant and medium Gaussian support vector machine (SVM) are used for training based on the seven customized features, which yield a high classification accuracy of 100 %. The proposed gas sensing approach can be promising for the development of intelligent and highly selective compact size sensing platforms without the need for special materials for coating and functionalization.

Index Terms—Smart gas sensor, selectivity, MEMS resonator, data processing, machine learning

I. Introduction

ECENTLY, machine learning has been increasingly implemented to realize smart sensors for various applications, including medical, industrial, and environmental monitoring [1-3]. Despite the rapid developments in the various kinds of gas sensors, for instance the chemo-resistive metal oxide semiconductor (MOS) sensors, they still suffer

This publication is based upon work supported by King Abdullah University of Science and Technology (KAUST). Nizar Jaber acknowledges the support of King Fahd University of Petroleum and Minerals (KFUPM)(usman Yaqoob and Wagner B. Lenz contributed equally to this work) (corresponding author: Mohammad I. Younis).

Usman Yaqoob, Wagner B. Lenz, and NouhaAlcheikh are with physical Science and Engineering Division, King Abdullah University of Science and Technology (KAUST), Thuwal 23955-6900, Saudi Arabia.

Nizar Jaber is with department of Mechanical Engineering and the Center for Communication Systems and Sensing, King Fahd University of Petroleum and Minerals, Dhahran, 31261, Saudi Arabia.

Mohammad I. Younis is with physical Science and Engineering Division, King Abdullah University of Science and Technology (KAUST), Thuwal 23955-6900, Saudi Arabia, and Department of Mechanical Engineering, State University of New York, Binghamton, NY 13902, USA.

from considerable cross-sensitivity [4, 5]. In the past few decades, electronic nose (E-nose) systems, which consist of arrays of sensors, have been successfully utilized to generate unique signature markers for a particular analyte through different data processing methods [6-8]. Jaeschke et al. [9]

developed an E-nose system using a set of analog and digital metal oxide sensors to detect and classify different concentrations of acetone and ethanol in dry and humid environments. The distinguished signature markers for each species were obtained through the principal component analysis (PCA) technique. Then a linear discriminant analysis (LDA) model was trained and tested to determine the accuracy. Schroeder et al. [10] demonstrated the classification and identification of several complex odors using an E-nose system based on an array of functionalized carbon nanotubes sensors. The PCA was used to perform accurate classification.

In addition, with the random forest model (RF) trained on extracted features, they could differentiate among food samples with high accuracy up to 91%. Similarly, Salhi et al.

[11] developed a hardware setup for gas leakage detection in smart homes, and their model demonstrated almost 100%

accuracy. Although high-performance E-nose systems have

Highly selective multiple gases detection using a thermal-conductivity-based MEMS resonator

and machine learning

Usman Yaqoob, Wagner B. Lenz, Nouha Alcheikh, Nizar Jaber, and Mohammad I. Younis, Member, IEEE

R

(3)

been realized using machine learning, they require the utilization of many sensors in a single system, which increases the complexity of the overall platform.

Resonator-based micro/nano-electromechanical systems (M/NEMS) have shown significant potential for sensing multiple physical, chemical, biological, and environmental parameters due to their high sensing performances, such as increased sensitivity, small size, and compatibility with CMOS circuits [12-14]. Their principle of operation is based on functionalizing the resonator surface with a material that has an affinity toward a targeted gas/analytes. A wide range of sensing materials has been developed and investigated for sensing different gases/chemicals such as graphene oxide, metal-organic frameworks (MOFs), metal oxide, and polymer composites [15-21]. For instance, Xu et al. [22] has recently developed a highly sensitive piezo-resistive cantilever resonator for the trace level detection of NO2. ZnO nanorods functionalized with self-assembled monolayer (SAM) of 3- aminopropyl)-trimethoxysilane was grown over the Si nanopilars to enhance the surface area, sensitivity, and selectivity of the sensing material. The developed resonator showed strong response to NO2 with great stability and a limit of detection of 2 ppb. Zhong et al. [23] proposed a ZnO-based CO2 sensor using a micro-cantilever resonator. The results showed good sensitivity for CO2 at high temperature with cross-sensitivity to H2 and CO. Most recently, Perelló-Roig et al. [24] developed CMOS-MEMS plate micro-resonator coated with polymer via inkjet printing for acetone sensing.

The proposed volatile organic piezo-resistive microcantilevers for humidity sensing compounds (VOCs) sensor shows a high sensitivity of 0.012 ppm/Hz to acetone with a cross-sensitivity to butane. They demonstrated that the coated sensor system had proven to increase the acetone detection response by six times compared to the uncoated platform. Therefore, the selection of a proper sensing material is a key factor in improving the sensitivity and selectivity toward a specific analyte. However, a precise and homogenous coating of the sensing material over tiny MEMS resonators is highly challenging and may require sophisticated tools. Also, the gas- sensing performance and sensitivity of these devices are inherently related to the coated surface, which limits the micro-resonator surface.

Gas sensors that utilize heated MEMS micro-resonators have become excellent candidates for developing high-performance gas sensing systems. The sensing mechanism of these devices is based on thermal energy dissipation (cooling or heating) in the presence of a target gas species [25, 26]. By taking advantage of the low stiffness of a heated buckled doubly- clamped micro-beam, high-performance resonant gas sensors have been realized with high sensitivity and good linearity [25]. However, the proposed method does not lead to a fully selective detection technique. Moreover, in a previous work based on a similar device structure, we demonstrated a highly selective gas sensor for He and Humidity detection by exploiting two different sensing mechanisms; absorption and thermal conductivity [19, 21]. The moveable mass surface (T- shaped cantilever) coated with graphene oxide (GO) is utilized for humidity detection, while the thermally heated flexural beams are used for He detection. However, in case of a mixture with three gases or more, this technique does not lead

to fully selective detection technique, and thus, it is not possible to discriminate the nature of the tested gases. Hence, the proposed work incorporate machine learning to further extend the capability of thermal-conductivity-based MEMS resonators. The machine learning algorithm is used to distinguish different gases and to estimate their concentration.

The accuracy is demonstrated for sensing three gases with different thermal conductivity (He (156.7 mW/mK), Ar (17.9 mW/mK), and CO2 (16.8 mW/mK) at 300 K). To generate the unique signature markers, multiple repeated gas sensing responses of each gas is recorded at different concentration levels of 4%, 10%, and 25%. The seven most significant customized features are extracted and then utilized to generate a well-scattered PCA graph, which helps in performing accurate gases classification. Three trained models, including Quadratic discriminant, Medium Gaussian support vector machine (SVM), and fine K- nearest neighbors (KNN) have been tested for high-performance classification. It is found that the Quadratic discriminant and SVM models lead to maximum classification accuracy (100%).

II. DEVICE DESIGN AND EXPERIMENTAL SETUP A schematic of the proposed device is illustrated in Fig. 1a.

It is mainly composed of a clamped-clamped (c-c) flexural beam (length (l) = 1000 µm, width (w) = 10 µm, and height (h) = 25 µm) attached to a T-shaped cantilever (l = 1000 µm, w = 25 µm, and h = 25 µm). In addition, to avoid the high rotational compliance of the structure in the y–z plane, the T- shaped structure is supported with an in-plane clamped-guided arch micro-beam. The highly conductive n-doped Si device is fabricated using a silicon-on-insulator (SOI) wafer by MEMSCAP. To vibrate the device, AC (Vac) voltage is applied at the anchors of the flexural micro-beam. As seen from Fig. 1a, the electro-thermal DC voltage Vdc (Vth) will generate an input current (iX) along the flexural micro-beam.

In addition, an external magnetic field is applied parallel to the sensor (BY~100 mT), which leads to an out-of-plane Lorentz force (FZ) normal to the sensor. Hence, Fz allows activating an out-of-plane motion of the flexural beam, so-called flexural mode (FM) [27, 28]. Fig. 1b shows an optical image of the proposed gas sensor. The inset shows an enlarged view of the T-shaped structure where the vibrometer laser spot position is focused at the center of the flexural beam. As shown in Fig.

1c, the gas setup consists of two gas supplies (Nitrogen and target analyte) and two mass flow controllers (Alicat scientific: flow vision), which maintain a constant flow rate of 2 SLPM (standard liters per minute) of the mixture gases during measurements. Flow Vision SC software is used to automate the on-off characteristics of the gas supply in the chamber. The device is placed into a chamber in which different gas concentrations are supplied through the inlet port.

The flexural beam resonator is electro-thermally driven using a data acquisition card (DAQ) at ambient temperature and atmospheric pressure. The resonance frequencies are tracked using a laser Doppler vibrometer (MSA-500) from Polytec and recorded using a data acquisition card (DAQ), see Fig. 1c.

The following expression is used to determine the gas concentration inside the chamber:

(4)

𝑇𝑎𝑟𝑔𝑒𝑡 𝐺𝑎𝑠 𝑐𝑜𝑛𝑐𝑒𝑛𝑡𝑟𝑎𝑡𝑖𝑜𝑛 (𝑃𝑃𝑀) =

𝑓𝑙𝑜𝑤 𝑟𝑎𝑡𝑒 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑎𝑟𝑔𝑒𝑡 𝑔𝑎𝑠

𝑓𝑙𝑜𝑤 𝑟𝑎𝑡𝑒 𝑜𝑓 𝑡ℎ𝑒 𝑡𝑎𝑟𝑔𝑒𝑡 𝑔𝑎𝑠 +𝑓𝑙𝑜𝑤 𝑟𝑎𝑡𝑒 𝑜𝑓 𝑁₂× 10⁶ (1)

Fig. 1. (a) Schematic diagram of the device. (b) an optical image of the fabricated sensor. Inset: an enlarged view of the T-shaped cantilever showing the laser spot position. (c) Schematic of the gas detection setup.

III. RESULTS AND DISCUSSIONS

Fig. 2 shows the simulated and experimental results of the variation of the resonance frequency of the flexural micro- beam with electrothermal voltages (Vth). At Vth = 0V, the white noise measurement shows an out-of-plane resonance frequency of the flexural beam mode at 𝑓_𝑜𝑓= 162.14 kHz, Fig. 2a. The inset displays its corresponding mode shape. Fig.

2b shows the flexural mode frequency response of the micro- beam by sweeping near fof for different Vac values while Vth is fixed at 0.5 V. A maximum linear amplitude of ~0.9 µm is achieved at Vac = 3 V. At higher AC values, the response shows a hardening behavior, as shown in Fig. 2b. The inset of Fig. 2b shows an example of such nonlinear behavior at Vac = 3.5 V. Moreover, in a previous work, we exploited the sensitivity of the resonance frequency near the buckling zone of a heated clamped-clamped micro-beam to demonstrate highly sensitive gas sensors [25]. Here; we choose to operate the device near the buckling zone of the out-of-plane vibration mode to maximize sensitivity. In addition, the presence of a magnetic field (BY) generates a Lorentz-force (FZ) that is normal to the micro-beam. This makes the beam biased (imperfect), and hence it will always buckle in one direction [29]. Hence, the direction of the buckled beam depends only on the direction of BY. Thus, any other imperfections in the beam will not affect its performance.

Fig. 2c shows the measured linear frequency sweep response of the micro-beam for different Vth values at Vac= 3 V. It is clear that increasing Vth reduces the resonance frequency. To identify the in-plane buckling zone, a multi-

physics finite element study (FEM) has been conducted using the COMSOL software, Fig. 2d.

It can be seen that the experimental data and the simulation results are in good agreement. As shown, increasing Vth, the resonance frequency decreases until reaching a minimum value of f = 147.205 kHz, which corresponds to the buckling point (Vth = 9V). In addition, the variation of the micro-beam temperature (T) with Vth is calculated, inset of Fig. 2d. The results show that for all the possible operating voltages, the temperature is below the silicon melting point (1410 °C).

Therefore, the sensing characterizations are carried out near the buckling zone at Vth = 7V, Vac = 3V, and f = 156.41 kHz.

Fig. 2. (a) The white noise response showing the FM resonance frequency fof value is at around 162.14 kHz, (b) linear frequency sweep response of the micro-beam for different Vac values while Vth set to 0.5 V and (c) frequency sweep at different Vth values Vac= 3 V, and (d) experimental data and simulation results of the variation of the out–of- plane flexural mode (FM) resonance frequency at different Vth. The inset shows the simulated results using a finite-element model for the variation of the temperature of the device versus Vth.

In previous work [19, 21], we showed performance comparisons of the proposed sensor for Helium detection with previously published MEMS based gas sensors. We demonstrated that our proposed device demonstrates much stronger response to He when compared to the previously published results. It should be noted that the environmental temperature might alter the micro-resonator temperature, and thus its resonance frequency may change. Hence, in our previous work we simulated the variation of the resonance frequency of the flexural beam mode versus temperature range from 10 to 55 °C. The simulated thermal coefficient of frequency (TCF) of the device is found to be −32.7 ppm/°C.

This result indicates that the proposed silicon MEMS sensor has low sensitivity to temperature variation. It is worth mentioning that the sensitivity of the proposed sensor is strongly dependent on the microbeam stiffness (resonance frequency of the resonator and its mass). Therefore, different geometries of resonator will affect the sensitivity of the device. We demonstrated that a big mass surface will enhance the sensitivity of the sensor. Another shortcoming is for the

(5)

low quality factor (Q ≈ 680 at Vac = 3 V and Vth = 7 V). Future work should be performed to optimize the resonator dimensions to enhance the quality factor, which can consequently improve sensing resolution.

A. Gas sensing mechanism

We first generate frequency response curves by exciting the micro-beam with a harmonic signal (Vth = 7V and Vac = 3V) near its flexural mode resonance frequency (FM), Fig. 3a. As described above that our resonator possess low quality factor, and hence designing phase locked loop (PLL) circuit for such devices become complicated. Therefore, the micro-beam is actuated near the resonance and the change in amplitude is monitored in presence of a target gas. Nevertheless, one should note that the amplitude change and the frequency shift are correlated through the frequency response curve [30].

Initially, the chamber is constantly flushed with dry Nitrogen (N2) for 30 min to generate a synthetic air environment. To investigate the sensing performance of the heated flexural beam, we expose the resonator to gases having higher/lower thermal conductivity, He (156.7 mW/mK), Ar (17.9 mW/mK), and CO2 (16.8 mW/mK), compared to N2 (26 mW/mK). Fig.

3b shows the frequency response of the micro-beam under 20% concentration of gases (He, Ar, and CO2) at room temperature and in air. As shown in [25], exposing a heated micro-beam to gases with a thermal conductivity higher than air, such as He, reduces the beam temperature, thereby decreasing the axial stress and increasing the resonance frequency. On the other hand, gases with lower thermal conductivity, such as Ar or CO2, heat the beam, and hence increase the axial stress and decrease the resonance frequency.

The concept of the proposed sensor is based on the cooling/heating effect between the device and the surrounding gas molecules, to enable the capturing of real-time data for the amplitude evolution during gas exposures, we excite the structure into out-of-plane vibration using magnets. In this case, the amplitude can be captured using laser Doppler vibrometers. Moreover, the gas sensitivity is related to the surface to volume ratio. Hence, the attachment of the T-shaped cantilever to the clamped–clamped micro-beam leads to an increase of the cooling/heating effect, and thus its sensitivity.

Fig. 4 shows schematic diagrams illustrating the interaction between the electro-thermally heated micro-beam and gas particles. Fig. 4a shows the heated micro-beam in synthetic air. It is well known that at higher operating temperature different O2 ionic species get adsorbed on metal oxide semiconductor materials and form multiple depletion regions [31-33]. Therefore, with the increase in Vth the heat dissipation in the beam increases, which causes the formation of depletion regions on the beam surface as illustrated in Fig. 4a. Figs. 4b, c and d present schematics of the micro-beam after exposing it to He, Ar, and CO2 gases, showing the thermal interaction between the heated micro-beam and the surrounding gas atoms/molecules. During He exposure, He atoms strike the micro-beam surface, which leads to increase in the heat transfer energy leading to cooler spots on the beam surface.

Hence, the effective thermal conductivity of He-N2 increases, and the micro-beam’s temperature decreases, which increases the stiffness.

On the other hand, Ar and CO2 with lower thermal conductivities compared to air show less heat energy transfer and dissipate more energy on the beam surface (Figs. 4c, b).

This decreases the beam stiffness and hence the frequency.

However, their sensing mechanism might be different from each other depending on their physical/chemical properties. Ar is an inert gas; therefore, it does not interact with the micro- beam surface, but rather generates a passivation layer around the beam, which does not allow heat to conduct through. As a result, it decreases the micro-beam stiffness (Fig. 4c). On the contrary, CO2 is a reactive gas, and its sensing mechanism can be better explained through reaction kinetics with n-Si microbeam surface [34, 35]. It is believed that at higher temperatures, CO2 dissociates and reacts with the beam surface resulting in the formation of additional depletion regions, thereby decreasing the microbeam’ stiffness, as shown in Fig. 4d.

Fig. 3.Frequency response curves at the flexural mode (FM). (a) the response showing the selected operating frequency for gas sensing, and (b) response curves for three gases with different thermal conductivity.

(6)

Fig. 4. Schematic diagrams illustrating the interaction between electrothermally heated micro-beams and gas particles: (a) synthetic air, (b) He, (c) Ar, and (d) CO2. (a) O2 adsorbents and formation of depletion regions at the higher temperature, (b) illustration of beam cooling during He exposure, (c) formation of passivation layer in the presence of Ar gas, (d) formation of extra depletion regions on the beam surface during CO2 exposure.

B. Collection of sensing data

The input data for the proposed machine learning algorithms consist of the amplitude sensing response due to three gases (He, Ar, and CO2) with different concentration levels at an operating frequency f of 156.44 kHz. Fig. 5 presents the sensing response to He, Ar, and CO2 at three concentration levels (4%, 10%, and 25%). The injection time of the target gas and dry N2 is automated through the Flow Vision SC software to obtain the multiple response cycles for data processing and machine learning. Fig. 5a shows the multiple response cycles of the micro-beam for He sensing.

The response/recovery time for He response measurements is set to 150/300s. Also, a similar dataset is collected for Ar and CO2 sensing with 300s/300s, as shown in Figures 5b, c. The results clearly show that both Ar and CO2 gases have almost similar response. During the Ar and CO2 exposure, a decrease in the resonance amplitude is observed, which is attributed to the decrease in stiffness, and the frequency shifts to the left.

However, for the He response, two extra jumps are noted, one is observed just after the injection of He, and the other is found during the recovery near the baseline. As explained earlier, the actuating frequency is on the right side of the frequency response curve (Fig. 3a). Therefore, after the He injection due to the beam cooling, the operating frequency immediately jumps to the left side, then the amplitude is reduced linearly with the He exposure time until it reaches saturation (response time). On the other hand, during recovery, when the amplitude approaches the resonance peak from the left side it again jumps to the right side (initial point), leaving another jump in the response. At higher concentration levels (10% and 25%) the micro-beam sensor shows good repeatability and stability for each gas. However, it shows significant variance in the response when exposed to the 4%

of Ar and CO2 as shown in Figures 5b, c. It is well known that data processing techniques can significantly reduce drift issues. Therefore, PCA is used to tackle the drift issue and to generate the unique signature markers which can successfully represent target class without overlapping [2, 3, 36, and 37].

Fig. 5d shows the combined results for all gases at all concentration levels.

Fig. 5.Gas sensing responses (a) He, (b) Ar, (c) CO2 with three different concentration levels 4%, 10%, and 25%, and (d) combined response of all three gases with their concentration levels.

C. Data processing and machine learning

The selection and classification of the most influential features are essential to accurately train a model with 100%

accuracy rate. 32 customized and 720 tsfresh extracted features using Python are processed and reduced to the 7 most significant and unique features for efficient and accurate model training. Fig. 6 shows a graph for a single cycle gas response labeled with the seven most significant and unique customized features processed through the Fscchi2 function and heatmap using the Matlab software. Selected features are namely represented as D1 (difference between baseline and saturation), 𝑎1− 𝑟𝑒𝑠 and 𝑅²res (coefficient of determination), which are acquired from the curve fitting of response time,𝑎2− 𝑟𝑒, 𝑎3− 𝑟𝑒𝑐 and 𝑅²rec, which are deduced from the curve fitting of the recovery time. Finally, t250 is the time at 250 sec near the saturation of the response time curve. The response 𝐹𝑟𝑒𝑠 and recovery time 𝐹𝑟𝑒𝑐curve-fitting equations are shown below as equations (2) and (3), respectively: These customized features are used to train different models for accurate classification of gases and their concentrations.

𝐹_𝑟𝑒𝑠(𝑥) = 𝑎₁cos(𝑎₂𝑥) + 𝑎₃ (2) 𝐹_𝑟𝑒𝑐(𝑥) = ^𝑎¹

(1+𝑒𝑎2𝑥−𝑎3) (3)

(7)

Fig. 6. Single-cycle of a gas response labeled with all 7 customized features. The selected parameters from the response and recovery time curves are highlighted in red.

Fig. 7a shows all the customized features with their importance score calculated by performing the chi-square (X²) test through Fscchi2 function in the Matlab software. This test determines the level of difference between the observed and expected values and ranks each feature on the basis of the observed p-value. The chi-square test can be conducted using equation (4)

𝜒𝑐2= ∑ ^(𝑂^𝑖^−𝐸^𝑖⁾²

𝐸_𝑖 𝑛

𝑖 (4)

where C is degrees of freedom, O is the observed value, and E is the expected value. The calculated p-value will be compared with the standard chi-square table to determine the significance of each feature.

After features ranking, it is essential to eliminate features, which are carrying repetitive information. Therefore, to eliminate the redundancy and correlation among the variables, the heatmap correlation matrix is developed as shown in Fig.

7b. The correlation matrix is obtained using the following equations:

𝜌_𝑖𝑗 =^{𝑐𝑜𝑣(𝑥,𝑦)}

𝜎_𝑥𝜎_𝑦 (5)

where 𝜌_𝑖𝑗 is the correlation coefficient, cov is the covariance, and 𝜎𝑛 is the standard deviation. Further, the covariance (cov) matrix can be calculated as:

𝑐𝑜𝑣(𝑥, 𝑦) =^{∑ (𝑥}^𝑛^𝑖 ^𝑖^{−𝑥̅)(𝑦}^𝑖^−𝑦̅)

𝑛 (6)

where 𝑥𝑖 and𝑦𝑖 are variables. Each variable will be compared pair wise with itself as well as with all other variables to determine their correlation score. 𝑥̅ and 𝑦̅ are the mean values. Finally, correlation matrix (𝑹) can be constructed as below:

𝑹 = (𝜌_𝑖𝑖 𝜌_𝑖𝑗

𝜌_𝑖𝑗 𝜌_𝑖𝑖) (7)

Fig. 7. (a) Feature importance scores for 32 customized features, (b) heatmap correlation matrix for 32 customized features, (c) feature importance scores for the 7 most significant features after the implementation of redundancy eliminating algorithm, and (d) heatmap correlation matrix for the 7 most significant and independent features.

From Fig. 7b, the boxes with the red color represent the maximum absolute correlation among the variables. To reduce the redundant information, we eliminate all the variables having a correlation score above 0.8, i.e., the red coded areas of the heatmap matrix. After eliminating the redundancy, the customized features are reduced to the seven most significant and unique variables. Figures 7c, d display the 7 most significant and unique features of the dataset with their importance score and heatmap correlation matrix, respectively. From Fig. 7d, after elimination, the only red coded areas that remain are in the diagonal of the matrix, which indicates the maximum variable correlation with itself only. Thus from the feature importance and heatmap graphs, it can be concluded that each selected variable carries important and unique information to implement into the machine learning algorithms for accurate model training. Furthermore, a large number of automatically extracted features are also generated using tsfresh. Tsfresh is a package in Python, which combines the multiple time series characterization methods for timely extraction of 794 features on the basis of automatically configured hypothesis tests [38]. Fig. 8 indicates the important feature scores and heatmap graphs for all 720 features obtained from the tsfresh using Python (a,b) and when they are reduced to 19 variables (c,d) after applying the redundancy elimination algorithm.

Fig. 8.(a) Feature importance scores for 720 tsfresh extracted features, (b) heatmap correlation matrix for 720 tsfresh extracted features, (c) feature importance scores for the 19 most significant features after the implementation of redundancy eliminating algorithm, and (d) heatmap correlation matrix for the 19 most significant and independent features

The 3D principal component analysis (PCA) graphs are generated from the seven customized as well as from the first seven tsfresh features. PCA further reduces these seven features into a unique single signature marker through diagonalization of the correlation matrix of the input dataset for efficient model training with low power consumption and

(8)

high processing speed. More details on PCA can be found in [2, 39]. PCA results can be obtained through the product sum of the normalized score (z-scores standardization) and eigenvectors as shown in equation (8)-(10):

𝑧_𝑖 =^𝑥^𝑖^−𝑥̅

𝜎 (8)

where 𝑧𝑖 is standardization scores for the ith feature, 𝑥𝑖 is the feature value for each response cycle, 𝑥̅ is the average of all the response cycles, and 𝜎 is the standard deviation.

Eigenvectors can be obtained using equation (9).

(𝑨 − 𝜆_𝑖𝑬)𝑽_𝒊= 0 (9)

where 𝑨 is the correlation matrix, 𝜆_𝑖 is eigenvalue, 𝑬 is identity matrix, and 𝑽𝒊 is the eigenvector. Finally, the PCA can be calculated according to equation (10) below.

𝑃𝐶𝐴 = ∑𝑛 𝑍𝑖

𝑖=1 𝑽𝑖 (10)

In the PCA graphs, no overlapping among the data points is observed and it is clear that using these customized and tsfresh features, the algorithm can easily and efficiently classify each gas species with their concentration levels.

Fig. 9. PCA results generated from the (a) customized, and (b) tsfresh features.

Finally, using seven customized features, three different classifiers, including quadratic discriminant, medium Gaussian SVM, and fine KNN are trained and tested. The confusion matrix obtained through seven customized features for all three classifiers are shown in Fig.10. To avoid overfitting and for stable prediction accuracy rates, it is vital to train and test

the model using cross validation (CV) techniques. Therefore, 8-fold CV test was conducted where the entire dataset is converted into 8 subsets to ensure that each data point takes part in training and testing of the model. Furthermore, to determine the accuracy, all the models are also validated with unknown datasets and the results are shown in the form of confusion matrix. The diagonal numbers in the confusion matrix represent the total examples for each gas used to train and test the model. Quadratic discriminant and medium Gaussian SVM reveal a 100% accuracy rate. However, the acceptable KNN model demonstrates 98% accuracy when the validation of trained model was tested with unknown dataset.

As shown in Fig. 10 (d), during KNN validation it confused 4% Ar with 4% CO2. On the other hand, the models, which are trained using first seven tsfresh extracted features, reveal a 100% accuracy rate during their validation with unknown datasets. All the results and extracted features have been summarized in Table 1. With the small datasets and k-fold validation technique our trained classifier model showed excellent performance. However, for more complex problems such as prediction of unknown gas concentration could be more effectively achieved with deep learning which requires bigger datasets for model training.

TABLEI

SUMMARY OF THE EXTRACTED FEATURES AND RESULTS

aThe explanation of each feature can be found on their website“https://tsfresh.readthedocs.io/en/latest/text/quick_start.html”

Extracted features Models Accuracy

Customized

D1 Quadratic

Discriminant 100%

a3-Rec

a1-Res Medium

Gaussian SVM 100%

R²res

250s

Fine KNN 98%

a2-Rec R²rec

aTsfresh

x_index_mass_quantile

_q_0.2 Quadratic

Discriminant 100%

x_index_mass_quantile _q_0.9

x_index_mass_quantile _q_0.1

Medium

Gaussian SVM 100%

x_agg_linear_trend_attr _stderr"_hunk_len_50_f _agg_"min""

x_fft_coefficient_attr_re al"_coeff_1"

Fine KNN 100%

x_longest_strike_below _mean

x_kurtosis

(9)

Fig. 10. Customized features results, confusion matrix for all three classifier models, (a) quadratic discriminant training/testing, (b) quadratic discriminant validation with unknown data, (c)l medium Gaussian SVM training/testing, (d) medium Gaussian SVM validation with unknown data, (e) fine KNN training/testing, and (f) fine KNN validation with unknown data.

IV. CONCLUSIONS

We successfully utilized the first out-of-plane mode of a flexural micro-beam resonator for selective and accurate detection of multiple gases with their different concentration levels using machine learning algorithms. The device is based on the axial load variation of a flexural electrothermally heated micro-beam in the presence of different gases. He, Ar, and CO2 gases with different concentration levels are tested and classified using machine learning. Seven highly significant and unique features are achieved through Fscchi2 function and heatmap correlation matrix. Well-scattered signature markers for each gas with all concentrations are realized using PCA. Then, they are used to train three different classifier models, including quadratic discriminant, medium Gaussian SVM, and fine KNN. Customized feature trained quadratic discriminant and SVM reveal a 100% accuracy rate for the classification of an unknown dataset. Finally, we believe that our proposed Si microbeam resonator-based multiple gases detection system can be a promising approach for developing highly selective and smart sensors. Moreover, in case of mixtures of four or more gases, the proposed technique and the absorption mechanism (T-shaped cantilever) may lead to even more highly selective gas sensors.

V. REFERENCES

[1] W. W. Hu, L. T. Wan, Y. Y. Jian, C. Ren, K. Jin, X. H. Su, X. X. Bai, H. Haick, M. S. Yao, and W. W. Wu, “Electronic Noses: From Advanced Materials to Sensors Aided with Data Processing,” Adv. Mater. Technol., vol. 4, no. 2, pp. 1800488- 38, Dec. 2019.

[2] U. Yaqoob, and M. I. Younis, “Chemical gas sensors:

recent developments, challenges, and the potential of machine

learning-A Review,” Sensors, vol. 21, no. 8, pp. 2877-40, Apr.

2021.

[3] S. B. Feng, F. Farha, Q. J. Li, Y. L. Wan, Y. Xu, and T.

Zhang, “Review on smart gas sensing technology,” Sensors, vol. 19, no. 17, pp. 3760-82, Aug. 2019.

[4] A. Dey, “Semiconductor metal oxide gas sensors: A review,” Mater. Sci. Eng. B-Adv., vol. 229, pp. 206-217, Mar.

2018.

[5] H.C. Ji, W. Zeng, and Y.Q. Li, “Gas sensing mechanisms of metal oxide semiconductors: a focus review,” Nanoscale, vol. 11 no. 47, pp. 22664-84, Oct. 2019.

[6] J. Yan, X. Z. Guo, S. K. Duan, P. F. Jia, L. D. Wang, C.

Peng, and S. L. Zhang, “Electronic Nose Feature Extraction Methods: A Review,” Sensors-Basel, vol. 15, no. 11, pp.

27804-27831, Nov. 2015,

[7] F. Rock, N. Barsan, and U. Weimar, “Electronic nose:

Current status and future trends,” Chem. Rev., vol. 108, no. 2, pp. 705-725, Jan. 2008.

[8] Q. H. Liu, X. Li, M. Ye, S. S. Ge, and X. Du, “Drift Compensation for Electronic Nose by Semi-Supervised Domain Adaption,” IEEE Sens. J.,vol. 14, no. 3 pp. 657-665, Mar. 2014.

[9] C. Jaeschke, J. Glockler, O. El Azizi, O. Gonzalez, M.

Padilla, J. Mitrovics, and B. Mizaikoff, “An Innovative Modular eNose System Based on a Unique Combination of Analog and Digital Metal Oxide Sensors,” ACS Sens., vol. 4, no. 9, pp. 2277-2281, Aug. 2019.

[10] V. Schroeder, E. D. Evans, Y. C. M. Wu, C. C. A. Voll, B. R. McDonald, S. Savagatrup, and T. M. Swager,

“Chemiresistive Sensor Array and Machine Learning Classification of Food,” ACS Sens., vol. 4, no. 8, pp. 2101- 2108, Jul. 2019.

[11] L. Salhi, T. Silverston, T. Yamazaki, and T. Miyoshi,

“Early Detection System for Gas Leakage and Fire in Smart Home Using Machine Learning,” I. Symp. Consum. Electr (ICCE)., Las Vegas, NV, USA, Jan. 2019, pp. 1–6.

[12] B. Lassagne, D. Garcia-Sanchez, A. Aguasca, and A.

Bachtold, “Ultrasensitive mass sensing with a nanotube electromechanical resonator,” Nano Lett., vol. 8, no. 11, pp.

3735-8, Oct. 2008.

[13] N. Jaber, S. Ilyas, O. Shekhah, M. Eddaoudi, and M.I.

Younis, “Resonant gas sensor and switch operating in air with metal-organic frameworks coating,” J. Microelectromech.

Sys., vol. 27, no. 2, pp. 156-63, Apr. 2018

[14] C. Stampfer, T. Helbling, D. Obergfell, B. Schoberle, M.

K. Tripp, and A. Jungen, “Fabrication of single-walled carbon-nanotube-based pressure sensors,” Nano Lett., vol. 6, no. 2, pp. 233-7, Jan. 2006.

[15] W. P. Xuan, M. Cole, J. W. Gardner, S. Thomas, F. H.

Villa-Lopez, and X. Z. Wang, “A film bulk acoustic resonator oscillator based humidity sensor with graphene oxide as the sensitive layer,” J. Micromech. Microeng., vol. 27, no. 5, pp.

055017-24, Mar. 2017.

[16] L. Kosuru, A. Bouchaala, N. Jaber, and M. I. Younis,

“Humidity detection using metal organic framework coated on QCM,” J. Sensors, vol. 2016, pp. 01-08, May 2016.

[17] J. S. Xu, M. Bertke, H. S. Wasisto, and E. Peiner,

“Piezoresistive microcantilevers for humidity sensing,” J.

Micromech. Microeng., vol. 29, pp. 053003-32, Apr. 2019.

(10)

[18] J. Xu, M. Bertkea, X. Li, H. Mu, H. Zhou, F. Yu, G.

Hamdan, A. Schmidt, H. Bremers, and E. Peiner, “Fabrication of ZnO nanorods and Chitosan@ZnO nanorods on MEMS piezoresistive self-actuating silicon microcantilever for humidity sensing,” Sens. Actuat. B-Chem., vol. 273 no. pp.

276-87, Nov. 2018.

[19] U. Yaqoob, N. Jaber, N. Alcheikh, M. I. Younis, “Mode- Dependent Selective Detection of Humidity and Helium Using Electromagnetically Actuated Clamped Guided MEMS Resonators,” Proc. of the ASME IDETC., virtual, Aug. 2021, pp. 1–5

[20] H. Yu, P. Xu, D. W. Lee, and X. Li, “Porous-layered stack of functionalized AuNP–rGO (gold nanoparticles–

reduced graphene oxide) nanosheets as a sensing material for the micro-gravimetric detection of chemical vapor,” J. Mater.

Chem. A, vol. 1, no. 14, pp. 4444-4450, Jan. 2013.

[21] U. Yaqoob, N. Jaber, N. Alcheikh, and M. I. Younis,

“Selective multiple analyte detection using multi-mode excitation of a MEMS resonator,” Sci. Rep., vol. 12, pp.

5297-11, Mar. 2022.

[22] J. Xu, A. Setiono, and E. Peiner, “Piezoresistive

Microcantilever with SAM-Modified ZnO-

Nanorods@Silicon-Nanopillars for Room-Temperature Parts- per-Billion NO2 Detection” ACS Appl. Nano Mater. Vol. 3 no.

7, pp. 6609-6620 Jun. 2020.

[23] C. Zhang, W. Jiang, A. Ghosh, G. Wang, F. Wu, and H.

Zhang, “Miniaturized langasite MEMS microcantilever beam structured resonator for high temperature gas sensing,” Smart.

Mater. Struct., vol. 29 no. 5, pp. 055002-14, Mar. 2020.

[24] R. Perelló-Roig, J. Verd, S. Bota, B. Soberats, A. Costa, and J. Segura, “CMOS–MEMS VOC sensors functionalized via inkjet polymer deposition for high-sensitivity acetone detection,” Lab Chip, vol. 21, no. 17, pp. 3307–3315, Jul.

2021.

[25] A.Z. Hajjaj, N. Jaber, N. Alcheikh, and M.I. Younis, “A resonant gas sensor based on multimode excitation of a buckled microbeam,” IEEE Sens. J., vol. 20, no. 4, pp. 1778- 85, Feb. 2020.

[26] A. Mahdavifar, M. Navaei, P.J. Hesketh, M. Findlay, J.R.

Stetter, and G.W. Hunter, “Transient thermal response of micro-thermal conductivity detector (mu TCD) for the identification of gas mixtures: An ultra-fast and low power method,” Microsyst. Nanoeng., vol. 1, pp. 15025-32, Oct.

2015.

[27] D. S. Greywall, “Micromechanical RF filters excited by the lorentz force,” J. Micromech. Microeng., vol. 9, no. 1, pp.

78 -84, Sep. 1999.

[28] T. Kemp, and M. Ward, “Tunable response nanomechanical beam resonator,” Sens. and Actuator. A.

Phys., vol. 123, pp. 281 -284, Sep. 2005.

[29] D. Roodenburg, J. W. Spronck, H. S. J. van der Zant, and W. J. Venstra, “Buckling beam micromechanical memory with on-chip readout”, Appl. Phys. Lett. vol. 94, pp. 183501, May 2009.

[30] M. T. Boudjiet, J. Bertrand, C. Pellet, and IDufour, "New characterization methods for monitoring small resonant frequency variation: Experimental tests in the case of hydrogen detection with uncoated silicon microcantilever- based sensors," Sens. and Actuat. B: Chemical, vol. 199, pp.

269-276. Aug. 2014.

[31] D. R. Miller, S.A. Akbar, and P.A. Morris, “Nanoscale metal oxide-based heterojunctions for gas sensing: A review,”

Sens. Actuat. B Chem., vol. 204, pp. 250–272, Dec. 2014.

[32] U. Yaqoob, D. T. Phan, A. S. M. I. Uddin, and G. S.

Chung, “Highly flexible room temperature NO2 sensor based on MWCNTs-WO3 nanoparticles hybrid on a PET substrate,”

Sens. Actuat. B-Chem. Vol. 221, pp. 760-08, Jul. 2015.

[33] 4. M. Al-Hashem, S. Akbar, and P. Morris, “Role of oxygen vacancies in nanostructured metal-oxide gas sensors:

A review,” Sens. Actuat. B-Chem. Vol. 301, pp. 126845-69, Dec. 2019.

[34] G. F. Fine, L. M. Cavanagh, A. Afonja, and R. Binions,

“Metal oxide semi-conductor gas sensors in environmental monitoring,” Sensors, vol. 10, no. 6, pp. 5469-5502, May 2010.

[35] B. Ashford , Y. Wang, C-K. Poh, L. Chen, and X. Tua, Plasma-catalytic conversion of CO2 to CO over binary metal oxide catalysts at low temperatures,” Appl. Catal. B., vol.

276, pp. 119110-20, Nov. 2020.

[36] A. Ziyatdinov, S. Marco, A. Chaudry, K. Persaud, P.

Caminal, and A. Perera, “Drift compensation of gas sensor array data by common principal component analysis,” Sens.

Actuat. B-Chem, vol. 146, no. 2, pp. 460-465, Apr. 2010.

[37]R. Gutierrez-Osuna, “Pattern Analysis for Machine Olfaction: A Review,” IEEE Sens. J., vol. 2, no. 3, pp.189- 202, Jun. 2002.

[38] M. Christ, N. Braun, J. Neuffer, and A. W. Kempa-Liehr,

“Time Series Feature Extraction on basis of Scalable Hypothesis tests (tsfresh – A Python package),”

Neurocomputing, Vol. 307, pp. 72-77, Sep. 2018.

[39] T. I. T. Akamatsu, A. Tsuruta, W. Shin, T. Itoh, and T.

Akamatsu, “Selective Detection of Target Volatile Organic Compounds in Contaminated Humid Air Using a Sensor Array with Principal Component Analysis,” Sensors, vol. 17, pp. 1662, Jul. 2017.