Metal–oxide–semiconductor (MOS) gas sensors are widely used for gas detection and monitoring. However, MOS gas sensors have always suffered from instability in the link between gas sensor data and the measured gas concentration. In this paper, we propose a novel deep learning approach that combines the surface state model and a Gated Recurrent Unit (GRU)-based regression to enhance the analysis of gas sensor data. The surface state model provides valuable insights into the microscopic surface processes underlying the conductivity response to pulse heating, while the GRU model effectively captures the temporal dependencies present in time-series data. The experimental results demonstrate that the theory guided model GRU+β outperforms the elementary GRU algorithm in terms of accuracy and astringent speed. The incorporation of the surface state model and the parameter rate enhances the model’s accuracy and provides valuable information for learning pulse-heated regression tasks with better generalization. This research exhibits superiority of integrating domain knowledge and deep learning techniques in the field of gas sensor data analysis. The proposed approach offers a practical framework for improving the understanding and prediction of gas concentrations, facilitating better decision-making in various practical applications.

Topics
Chemical reaction dynamics, Surface states, Metal-oxide-semiconductor, Microsensor, Deep learning, Machine learning, Reaction rate constants, Regression analysis

Detecting toxic and harmful gas molecules is essential for various fields, including environmental monitoring, food quality assessment, industrial safety, explosive detection, and healthcare. Metal–oxide–semiconductor (MOS) gas sensors have gained significant attention and have proven to be useful in Internet of Things (IoT) applications due to their low manufacturing costs, small size, long lifetime, and compatibility with different systems.1–7 However, MOS gas sensors are prone to long-term drift due to aging and contamination.8 To mitigate this issue, it is to use a calibration model that corrects for signal variations.1,9 The calibration process involves exposing the sensor to a known concentration of the target gas, measuring the sensor’s response, and generating a calibration curve using linear regression or other mathematical methods. This allows for accurate gas concentration measurements.10 However, the lack of a solid theoretical foundation for this calibration method leads to a highly variable correlation between measurement information and concentration levels. In addition, the calibration process requires MOS sensors to operate continuously at high temperatures between 300 and 400 °C to maintain a stable equilibrium state, resulting in high power consumption and unreliable measurement methods.11–13 

Recently, researchers have explored alternative methods to improve the reliability and efficiency of gas sensing technologies. The pulse heating (PH) mode detects transient characteristics of multivariable response arising from gas diffusion and gas–solid surface reactions.14–22 Compared to the conventional continuous heating (CH) mode, the PH mode provides better stability, sensitivity, and selectivity, making it a promising candidate for establishing a relationship between conductance information and gas concentration.23–30 

The surface state trapping model proposed by Ding et al.31 in the pulse mode reveals surface processes connected with the conductance response and provides the possibility to establish a regression relationship. However, in practical applications, computing all the parameters in the model is nearly impossible. To address this limitation, the model has been simplified through approximate calculations, making specific physical parameters more intuitive, as demonstrated by Fort et al.32 Schultealbert et al.33 further simplified the model by proposing the use of a rate constant parameter for sensor calibration. Miquel-Ibarz et al.34 also utilized the PH mode for calibration and achieved better results. However, using a calibration model to describe the relationship between measurement information and gas concentration cannot blindly exhaust all regression possibilities. Therefore, researchers have widely employed machine learning to simulate regression relationships, treating them as black boxes that can be predicted when similar inputs are obtained.35–37 Although this approach offers a simple and straightforward method for establishing regression relationships as data-oriented tools, it may sacrifice accuracy when the conductance of a sensor is susceptible to fluctuations.38–46 

To overcome the limitations of traditional machine learning techniques and enhance reliability and consistency, feature fusion deep learning has emerged as a promising approach. This approach integrates scientific knowledge into data science models, improving their effectiveness and generalizability. Furthermore, this method can improve the accuracy of the model, accelerate the training speed of the model, and have better interpretability.47–49 

In this study, we present a novel regression model that combines the strengths of physical fusion machine learning and pulsed heating response characteristics. We developed a pulse-heated gas sensing test system to gather pulse data and extracted the feature parameter β using the surface state model. Subsequently, we incorporated this parameter as a covariate during the training of a machine learning network, specifically the Gated Recurrent Unit (GRU). To assess the effectiveness of our proposed regression model, we conducted tests on a set of gas test data that was not utilized in the training phase. By integrating characteristic parameter from a physical model, our regression model distinguishes itself from traditional machine learning approaches and follows a straightforward process from measurement information to gas concentration, demonstrating improved regression performance. These findings indicate the potential of our model in establishing more stable and scientifically robust regression relationships between measurement information and gas concentrations. The integration of pulse heating with deep learning in testing methods holds significant potential across a wide array of scenarios characterized by diverse applications, rigorous power consumption requirements, and stringent stability prerequisites. These areas include but are not limited to environmental monitoring, Internet of Things (IoT) applications, and more. This combined approach offers the prospect of enhancing efficiency and accuracy in these domains, making it a compelling solution for addressing various real-world challenges.

The test system used in this study consists of a dynamic gas distribution instrument (DGL-II, Beijing Elite Tech Co., Ltd., China), sensor arrays (GM-302B, Zhengzhou Winsen Technology Co., Ltd.), and data acquisition equipment (NI-USB-6002), as shown in Fig. 1. The DGL-II employs a peristaltic pump to accurately inject liquid, which is then evaporated on a heating platform to create a gaseous state. The mixed air with the target gas is precisely controlled using a Mass Flow Controller (MFC) to achieve a stable atmosphere in the gas chamber.

FIG. 1.
Test system schematic and pulse heating strategy.
View largeDownload slide

Test system schematic and pulse heating strategy.

FIG. 1.
Test system schematic and pulse heating strategy.
View largeDownload slide

Test system schematic and pulse heating strategy.

Close modal

A customized printed circuit board (PCB) with an integrated power amplifier circuit was used for concentration testing with the sensor array. The real-time humidity inside the test chamber was monitored using an AM-1011A sensor (Guangzhou Aosong Electronic Co., Ltd.). Voltage information was sampled and transferred to the LabVIEW software on a PC through a USB-6002 data acquisition card. The test circuit utilized the standard series voltage divider method to measure the conductance of each sensor.

In this study, a pulse heating strategy was employed, where a pulse voltage (Vh) was applied to the heating electrode of the gas sensor to create a temporary high-temperature working environment. This allowed for the measurement of the pulse response output Vn of each sensor. The pulse parameters used in the experiments included an operating voltage (Vh) of 0.5–3 V and a pulse width of 300 ms. Ethanol, being one of the most prevalent Volatile Organic Compounds (VOCs), assumes a pivotal role in various domains, such as food safety assessment, drunk driving monitoring, and public safety diagnostics.50,51 Consequently, ethanol was selected as the target gas for subsequent experiments.

During these experiments, ethanol gas concentrations were systematically varied within the range of 10–100 parts per million (ppm), with an increment of 10 ppm for each step. Each concentration was tested over a duration of 1 h. In order to simulate the environmental conditions of real world as much as possible, the random gas mixtures method is a good choice.52 However, it requires complex atmosphere control operations and long-time consumption for a test. Therefore, we choose mixed air as the background gas for sequential testing, a simpler and more efficient testing method to avoid the interference of other environmental factors caused by long-term testing. The conductance of each sensor, denoted as Gs, can be calculated using the following equation:
(1)
where VCC is the supply voltage of the test circuit and Rf is the load resistance, as shown in the test circuit of Fig. 1.

Data analysis involved determining the optimal duty cycle for pulsed heating by testing different periods with a fixed pulse width of 300 ms. Test signals of the gas sensors at 5, 10, and 20 s periods were analyzed to determine the optimized stabilization time. Figure 2(a) in the original description shows the timing test signals of the gas sensors at different periods, while Fig. 2(b) illustrates the conductance response of the sensor at various concentrations. As depicted in Fig. 2(a), a run-in effect is observed in gas sensor testing; the run-in effect refers to the phenomenon where the sensor’s readings stabilize over time when it is initially used or after a prolonged period of non-use, and the run-in effect in pulse heating is closely related to the amplitude and duty cycle of the pulse. In Fig. 2(a), through a simple comparison experiment, the 10 s pulse heating method was selected for subsequent tests. In Fig. S1 of the supplementary material, we can see that there is no dependence between the different concentration datasets tested due to the run-in effect. A comparison of power consumption between the continuous heating (CH) mode and the pulse heating (PH) mode was made, revealing that the PH mode can save more than 95% of power consumption when reaching a suitable operating temperature. See Fig. S2 of the supplementary material for details. This energy-saving advantage is beneficial for applications in scenarios with strict energy consumption requirements.

FIG. 2.
(a) Results of pulsed heating at different cycles. (b) Conductance at different gas concentrations under pulsed heating. (c) Pulse heating and continuous heating transient curves toward 10, 20, 30, 40, and 50 ppm ethanol.
View largeDownload slide

(a) Results of pulsed heating at different cycles. (b) Conductance at different gas concentrations under pulsed heating. (c) Pulse heating and continuous heating transient curves toward 10, 20, 30, 40, and 50 ppm ethanol.

FIG. 2.
(a) Results of pulsed heating at different cycles. (b) Conductance at different gas concentrations under pulsed heating. (c) Pulse heating and continuous heating transient curves toward 10, 20, 30, 40, and 50 ppm ethanol.
View largeDownload slide

(a) Results of pulsed heating at different cycles. (b) Conductance at different gas concentrations under pulsed heating. (c) Pulse heating and continuous heating transient curves toward 10, 20, 30, 40, and 50 ppm ethanol.

Close modal

Figure 2(c) displays ten consecutive test data samples and the envelope of the PH mode test data; the conductance response of the sensor at various concentrations was analyzed; and it was observed that the conductance information enables the distinction of gas concentrations under pulsed non-equilibrium conditions. In addition, the stability and sensitivity of the PH mode test data were demonstrated by analyzing ten consecutive test data samples ranging from 10 to 50 ppm and comparing them with the results of the CH mode test. It is important to note that the deep learning dataset is derived from a subset of the original data, specifically from the 1/6 to 5/6 portion. This selection ensures that the data incorporated into the deep learning dataset remain nearly impervious to the run-in effect, preserving the integrity of the results.

While traditional machine learning is based on the data shown in Fig. 2, this paper is based on a deep learning approach incorporating the surface state trapping model. The surface state trapping model, described in Ref. 31, explained the mechanism for the microscopic surface processes of the conductivity response to pulse heating.

The surface state trapping model consists of four stages in sequence, as shown in Fig. 3. In the first stage, the rapid rise in heating voltage causes a temperature increase in the sensitive material, leading to an increase in conductivity due to the jumping of charges from the valence band to the conduction band. This temperature rise occurs within a short time, typically around 10 ms.

FIG. 3.
(a) Conductance change during pulsed heating. (b) Surface trapping model micro-processes.
View largeDownload slide

(a) Conductance change during pulsed heating. (b) Surface trapping model micro-processes.

FIG. 3.
(a) Conductance change during pulsed heating. (b) Surface trapping model micro-processes.
View largeDownload slide

(a) Conductance change during pulsed heating. (b) Surface trapping model micro-processes.

Close modal

The second stage works at a constant temperature. The high temperature disrupts the charge balance between the occupied surface state and the conduction band, resulting in charges crossing the surface potential barrier (qVs) and being captured by oxygen chemisorbed on the surface, forming an occupied surface state (O−). This stage is characterized by the consumption of charges in the conduction band and the transition from a non-equilibrium state to an equilibrium state.

In the third stage, the falling edge of the pulse induces a rapid thermal-induced relaxation of conduction band electrons to the valence band, causing a sudden drop in conductance. At this point, equilibrium may not yet be attained from the second stage, resulting in the system transition to a new non-equilibrium state.

Finally, in the fourth stage, the conductance gradually increases due to the low operating temperature. This model provides a comprehensive explanation of the microscopic surface processes underlying the conductivity response to pulse heating.

To account for the changing width of the space charge region during conduction band and surface state charge transfer, the surface potential barrier is considered a variable in the process. The conductance (G) can be expressed as a function of (Vs) using the potential barrier model, as shown in the following equation:
(2)
where q is the magnitude of electronic charge, K is the Boltzmann constant, T is the thermodynamic temperature, G0 = gqμsNd, g is a constant determined by the semiconductor geometry, μs is the mobility of the electrons, and Nd is the density of donors. Taking into account the baseline drift, Gc is added to the equation. In general, G0 is often regarded as a constant.
The surface potential (Vs) is related to the density of occupied surface states (Ns) and can be expressed using the following equation:
(3)
where ɛ0ɛr is the electrical permittivity of the semiconductor.
The charge transfer between the surface state and the conduction band can be interpreted as a first-order charge transfer between energy bands. The captured charge by the surface state and the released charge from the surface state are expressed in Eqs. (4) and (5), respectively,
(4)
where Ns denotes the density of the surface already occupied by charge and Nt denotes the total density of surface states, considered here as a constant. k1 is the rate constant at which charge is trapped. The charge that is released from the surface state can be expressed as
(5)
where k−1 denotes the rate constant for the decomposition of electrons from the surface state.
The reactions on the surface of sensitive materials during the second stage are described as follows:
(6)
The corresponding relationship here is that the chemisorbed oxygen is represented as the unoccupied surface state, the number of charges e is ns, and the density of O− is Ns.
The change rate of Ns is given by Eq. (7), considering the reaction with the reducing gas, as shown in Eqs. (8) and (9),
(7)
(8)
(9)
where [R] is the concentration of reducing gas, k2 is the reaction rate of reducing gas, and k−2 is its corresponding reverse reaction rate. When reducing gases are taken into account, Eq. (7) is modified as follows:
(10)
where k 1 ′ and k − 1 ′ are k1 and k−1 modified by the reducing gas term. Simplifying the model according to Fort et al.,53 we obtain
(11)
where A, B, and C are constants when the temperature is constant, θ = k2[R]. The Taylor expansion of the above-mentioned equation leads to Eq. (12), and after merging like terms, Eq. (13) is derived,
(12)
(13)
where α = A and β = B + C + θ. Then, the general solution of Eq. (13) is N s = C n s e − β t + α β ⁠. Bringing the general solution into Eq. (3) and then into Eq. (2), we obtain
(14)
The general solution of the function is given by Eq. (14), where λ = − q 2 2 K T ε 0 ε r N d , C 1 = C n s 2 ⁠, C 2 = 2 C n s α β ⁠, and C 3 = α 2 β 2 ⁠. By fitting this equation to the test data, the parameter rate β can be extracted. In stage 2, the reduction in conductance is primarily due to the reaction of the occupied surface state, and the rate of conductance reduction decreases with increasing concentration of the reducing gas. This is reflected in a decrease in the value of β, which tends to approach 0. This finding is consistent with the experimental results, as shown in Fig. 4, where the parameter rate β decreases as the concentration increases.
FIG. 4.
Distribution of characteristic parameter rate (β) at different gas concentrations.
View largeDownload slide

Distribution of characteristic parameter rate (β) at different gas concentrations.

FIG. 4.
Distribution of characteristic parameter rate (β) at different gas concentrations.
View largeDownload slide

Distribution of characteristic parameter rate (β) at different gas concentrations.

Close modal

Since the pulse test data conform to the characteristics of time-series data, a time-series model is preferred. In addition, the GRU has fewer parameters that make it better adapted to edge devices, such as gas sensors, so the GRU-based model is chosen in this paper, as shown in Fig. 5. The model architecture consists of a GRU model with four layers and two predictor layers. The output of each GRU layer serves as the input for the next layer, and the output of the last layer passes through the predictor layers to obtain the concentration prediction. The GRU model is configured with a hidden dimension of 8 and an input dimension of 1. It has three gates, and the Netron visualization tool is used to obtain the parameter statistics; see Fig. S3 of the supplementary material for details. The predictors were used to project the high-dimensional features extracted by the GRU onto a single value. This projection is carried out in two steps.

FIG. 5.
Schematic of the GRU-based model.
View largeDownload slide

Schematic of the GRU-based model.

FIG. 5.
Schematic of the GRU-based model.
View largeDownload slide

Schematic of the GRU-based model.

Close modal

First, the input steps are projected to 1, and then, the hidden dimensions are projected to 1. Finally, we get the output of the GRU model for the concentration prediction task.

To enhance the performance of the model, we have incorporated domain knowledge by introducing a parameter rate for effective training. The new structure, shown in Fig. 6, implements this idea. The conductance information is pre-processed to generate a series of 32@70 × 1 samples, where 32 represents the batch size and 70 represents the data points captured at a sampling rate of 100 Hz over a period of 700 ms. This period includes a 200 ms pre-pulse, a 300 ms pulse, and a 200 ms post-pulse. The pre-processed data pass through the GRU module, resulting in a 32@70 × 8 output, where 70 represents the output of the 70 GRU units with a hidden dimension of 8. This output is then fed into the two predictors to obtain a 32@1 × 1 output. In addition, the parameters of the test information are extracted using Eq. (14) and multiplied by a learnable variable as a bias to add to the output of the predictor, resulting in a new output. The range of learnable variable is set between 1 and 10. Deep learning will find the optimal solution within its own range, but this may lead to data distribution that is deviated from the actual scenario. We add that this parameter serves as a priori knowledge, which can very effectively correct the prediction interval of the model, which will help the model avoid out of distribution. The product of the learnable variable and the parameter β are used as a whole as a bias to achieve fine-tuning of the regression task. Compared to other pulse parameters such as the maximum value and other static parameters, gas response process is not easy to change with the aging of the gas sensor, baseline drift, and other problems. β is a dynamically based parameter derived from the gas response process. The parameters derived from the dynamic response process exhibit a degree of regularity and stability. This regularity and stability can positively impact the accuracy and generalization of the gas concentration prediction model. It is important to note that the current level of regularity and stability is solely based on data analysis. To gain a deeper understanding, further research into the microscopic model of the metal oxide gas sensor pulse heating process is essential for enhancing the performance of our concentration prediction model.

FIG. 6.
Schematic diagram of our proposed network (GRU+β). The number before and after “@” refers to the number of feature maps and the dimension of feature map in the layer.
View largeDownload slide

Schematic diagram of our proposed network (GRU+β). The number before and after “@” refers to the number of feature maps and the dimension of feature map in the layer.

FIG. 6.
Schematic diagram of our proposed network (GRU+β). The number before and after “@” refers to the number of feature maps and the dimension of feature map in the layer.
View largeDownload slide

Schematic diagram of our proposed network (GRU+β). The number before and after “@” refers to the number of feature maps and the dimension of feature map in the layer.

Close modal

We use Mean Squared Error (MSE) as the loss function to calculate the difference between output and label. We compared the results obtained from the GRU+β model with those from the elementary GRU algorithm. Figure 7(a) illustrates the results, showing that the GRU+β model achieves a higher accuracy and faster astringent compared to the GRU model. The loss curves [Figs. 7(b)–7(d)] demonstrate the MSE losses for the training set, validation set, and test set. The GRU model achieves MSE losses of 8.91, 2.94, and 3.01 for the training, validation, and test sets, respectively. In contrast, the GRU+β model achieves lower losses of 1.032, 0.874, and 2.1, respectively. Moreover, Fig. 7(e) shows the prediction results from the dataset in Fig. 7(f) by the two models. GRU+β has a better accuracy and generalization when a certain type of data is missing from the training set. It is worth mentioning that for scenarios extending beyond the prediction of unknown gas concentrations, such as complex mixtures of background gases and high humidity conditions, the achievement of similarly accurate regression results necessitates the construction of more comprehensive and mature datasets to enhance our existing model. Other three sensors’ results are mentioned in Fig. S4 of the supplementary material. These results show the superiority of the GRU+β model and illustrate the feasibility of feature fusion deep learning for gas detection.

FIG. 7.
GRU and GRU+β comparison results. (a) Accuracy and astringent speed results. (b)–(d) Test loss, train loss, and validation (val) loss with epoch. (e) The regression results of GRU and GRU+β under the dataset of (f).
View largeDownload slide

GRU and GRU+β comparison results. (a) Accuracy and astringent speed results. (b)–(d) Test loss, train loss, and validation (val) loss with epoch. (e) The regression results of GRU and GRU+β under the dataset of (f).

FIG. 7.
GRU and GRU+β comparison results. (a) Accuracy and astringent speed results. (b)–(d) Test loss, train loss, and validation (val) loss with epoch. (e) The regression results of GRU and GRU+β under the dataset of (f).
View largeDownload slide

GRU and GRU+β comparison results. (a) Accuracy and astringent speed results. (b)–(d) Test loss, train loss, and validation (val) loss with epoch. (e) The regression results of GRU and GRU+β under the dataset of (f).

Close modal

In this paper, we have presented an accurate and reliable regression model for gas sensor data using a deep learning approach that combines the surface state model and a Gated Recurrent Unit (GRU). Our findings highlight the effectiveness of incorporating domain knowledge and temporal dependencies in improving the accuracy and performance of gas sensor data analysis. By leveraging the surface state model, we have provided sophisticated understanding of the microscopic surface processes underlying the conductivity response to pulse heating, with valuable insights into the charge transfer mechanisms and reactions occurring on the sensor surface, enhancing the interpretability of the data. The integration of the GRU-based regression model allows us to capture and leverage the temporal dependencies present in the time-series data. The GRU model efficiently learns the sequential patterns and long-term dependencies, enabling accurate predictions of gas concentrations. The experimental results demonstrate that in datasets where partial data are missing, the GRU+β model has a higher accuracy and better generalization than GRU, which means that, in practice, the GRU+β model will be a better way to predict gas concentrations.

This research contributes to the advancement of gas sensor data predicted by providing a practical framework that combines domain knowledge with deep learning techniques. The approach presented here exhibits significant promise for gas monitoring in a variety of distinct scenarios. Furthermore, extending this methodology to encompass different types of gas sensors and assessing its suitability in real-time applications would constitute valuable avenues for further research. While the performance of this parameter β has yet to be validated in more intricate testing environments, such as gas detection involving multiple gas types, it is conceivable that by identifying parameters capable of distinguishing between various gas types, combining these with the parameter β could yield more sophisticated gas detection capabilities. This potential for complex gas detection is an exciting prospect worth exploring. There is no doubt that humidity is an important parameter in the use of MOS gas sensors, and subsequent research on the effect of humidity on pulse heating will be promising in the future. Overall, this research showcases the importance of integrating domain knowledge and deep learning techniques for accurate and reliable of gas sensor data regression. It paves the way for advancements in gas sensing technologies and their applications in diverse industries.

See the supplementary material for information on power consumption comparisons and concentration tests in this study.

This work was financially supported, in part, by the National Key R & D Program of China under Grant No. 2020YFB2008604 and, in part, by the National Natural Science Foundation of China under Grant No. 62174077. The authors acknowledge the technical support from SUSTech CRF.

The authors have no conflicts to disclose.

Yi Zhuang: Conceptualization (equal); Data curation (equal); Funding acquisition (equal); Investigation (equal); Methodology (equal); Project administration (equal); Resources (equal); Software (equal); Supervision (equal); Writing – original draft (equal); Writing – review & editing (equal). Du Yin: Conceptualization (equal); Data curation (equal); Investigation (equal); Methodology (equal); Software (equal); Writing – original draft (equal). Lang Wu: Data curation (equal); Investigation (equal); Software (equal). Gaoqiang Niu: Conceptualization (equal); Data curation (equal); Investigation (equal); Methodology (equal). Fei Wang: Conceptualization (equal); Data curation (equal); Funding acquisition (equal); Methodology (equal); Project administration (equal); Resources (equal); Supervision (equal); Writing – review & editing (equal).

The data that support the findings of this study are available within the article and its supplementary material. The data and codes that support the findings of this study are openly available in GitHub at https://github.com/HWare-magic/GRU_regression_gas-sensor.54 

1.
A.
Dey
, “
Semiconductor metal oxide gas sensors: A review
,”
Mater. Sci. Eng. B
229
,
206
–
217
(
2018
).
https://doi.org/10.1016/j.mseb.2017.12.036
Google Scholar
Crossref
Search ADS
 
2.
S.
Yuan
 and
S.
Zhang
, “
Recent progress on gas sensors based on graphene-like 2D/2D nanocomposites
,”
J. Semicond.
40
,
111608
(
2019
).
https://doi.org/10.1088/1674-4926/40/11/111608
Google Scholar
Crossref
Search ADS
 
3.
A.
Kolmakov
,
D. O.
Klenov
,
Y.
Lilach
,
S.
Stemmer
, and
M.
Moskovits
, “
Enhanced gas sensing by individual SnO2 nanowires and nanobelts functionalized with Pd catalyst particles
,”
Nano Lett.
5
,
667
–
673
(
2005
).
https://doi.org/10.1021/nl050082v
Google Scholar
Crossref
Search ADS
PubMed
 
4.
G.
Niu
 and
F.
Wang
, “
A review of MEMS-based metal oxide semiconductors gas sensor in Mainland China
,”
J. Micromech. Microeng.
32
,
054003
(
2022
).
https://doi.org/10.1088/1361-6439/ac5b98
Google Scholar
Crossref
Search ADS
 
5.
B.
Wang
,
A.
Thukral
,
Z.
Xie
,
L.
Liu
,
X.
Zhang
,
W.
Huang
,
X.
Yu
,
C.
Yu
,
T. J.
Marks
, and
A.
Facchetti
, “
Flexible and stretchable metal oxide nanofiber networks for multimodal and monolithically integrated wearable electronics
,”
Nat. Commun.
11
,
2405
(
2020
).
https://doi.org/10.1038/s41467-020-16268-8
Google Scholar
Crossref
Search ADS
PubMed
 
6.
Z.
Song
,
W.
Ye
,
Z.
Chen
,
Z.
Chen
,
M.
Li
,
W.
Tang
,
C.
Wang
,
Z.
Wan
,
S.
Poddar
,
X.
Wen
,
X.
Pan
,
Y.
Lin
,
Q.
Zhou
, and
Z.
Fan
, “
Wireless self-powered high-performance integrated nanostructured-gas-sensor network for future smart homes
,”
ACS Nano
15
,
7659
–
7667
(
2021
).
https://doi.org/10.1021/acsnano.1c01256
Google Scholar
Crossref
Search ADS
PubMed
 
7.
Z.
Idrees
,
Z.
Zou
, and
L.
Zheng
, “
Edge computing based IoT architecture for low cost air pollution monitoring systems: A comprehensive system analysis, design considerations and development
,”
Sensors
18
,
3021
(
2018
).
https://doi.org/10.3390/s18093021
Google Scholar
Crossref
Search ADS
PubMed
 
8.
H.
Chai
,
Z.
Zheng
,
K.
Liu
,
J.
Xu
,
K.
Wu
,
Y.
Luo
,
H.
Liao
,
M.
Debliquy
, and
C.
Zhang
, “
Stability of metal oxide semiconductor gas sensors: A review
,”
IEEE Sens. J.
22
,
5470
–
5481
(
2022
).
https://doi.org/10.1109/jsen.2022.3148264
Google Scholar
Crossref
Search ADS
 
9.
N.
Dennler
,
S.
Rastogi
,
J.
Fonollosa
,
A.
van Schaik
, and
M.
Schmuker
, “
Drift in a popular metal oxide sensor dataset reveals limitations for gas classification benchmarks
,”
Sens. Actuators, B
361
,
131668
(
2022
).
https://doi.org/10.1016/j.snb.2022.131668
Google Scholar
Crossref
Search ADS
 
10.
S.
Xue
,
S.
Cao
,
Z.
Huang
,
D.
Yang
, and
G.
Zhang
, “
Improving gas-sensing performance based on MOS nanomaterials: A review
,”
Materials
14
,
4263
(
2021
).
https://doi.org/10.3390/ma14154263
Google Scholar
Crossref
Search ADS
PubMed
 
11.
D.
Dorcea
,
M.
Hnatiuc
, and
I.
Lazar
, “
Acquisition and calibration interface for gas sensors
,” in
2018 IEEE 24th International Symposium for Design and Technology in Electronic Packaging (SIITME)
(
IEEE
,
Iasi, Romania
,
2018
), pp.
120
–
123
.
Google Scholar
Crossref
Search ADS
 
12.
X.
Liu
,
S.
Cheng
,
H.
Liu
,
S.
Hu
,
D.
Zhang
, and
H.
Ning
, “
A survey on gas sensing technology
,”
Sensors
12
,
9635
–
9665
(
2012
).
https://doi.org/10.3390/s120709635
Google Scholar
Crossref
Search ADS
PubMed
 
13.
Y.
Robin
,
J.
Amann
,
P.
Goodarzi
,
A.
Schutze
, and
C.
Bur
, “
Transfer learning to significantly reduce the calibration time of MOS gas sensors
,” in
ISOCS/IEEE International Symposium on Olfaction and Electronic Nose (ISOEN)
(
IEEE
,
Aveiro, Portugal
,
2022
), pp.
1
–
4
.
Google Scholar
Crossref
Search ADS
 
14.
T.
Dai
,
G.
Meng
,
Z.
Deng
,
Y.
Chen
,
H.
Liu
,
L.
Li
,
S.
Wang
,
J.
Chang
,
P.
Xu
,
X.
Li
, and
X.
Fang
, “
Generic approach to boost the sensitivity of metal oxide sensors by decoupling the surface charge exchange and resistance reading process
,”
ACS Appl. Mater. Interfaces
12
,
37295
–
37304
(
2020
).
https://doi.org/10.1021/acsami.0c07626
Google Scholar
Crossref
Search ADS
PubMed
 
15.
X.
Liang
,
T.-H.
Kim
,
J.-W.
Yoon
,
C.-H.
Kwak
, and
J.-H.
Lee
, “
Ultrasensitive and ultraselective detection of H2S using electrospun CuO-loaded In2O3 nanofiber sensors assisted by pulse heating
,”
Sens. Actuators, B
209
,
934
–
942
(
2015
).
https://doi.org/10.1016/j.snb.2014.11.130
Google Scholar
Crossref
Search ADS
 
16.
K.
Suematsu
,
Y.
Hiroyama
,
W.
Harano
,
W.
Mizukami
,
K.
Watanabe
, and
K.
Shimanoe
, “
Double-step modulation of the pulse-driven mode for a high-performance SnO2 micro gas sensor: Designing the particle surface via a rapid preheating process
,”
ACS Sens.
5
,
3449
–
3456
(
2020
).
https://doi.org/10.1021/acssensors.0c01365
Google Scholar
Crossref
Search ADS
PubMed
 
17.
K.
Suematsu
,
T.
Oyama
,
W.
Mizukami
,
Y.
Hiroyama
,
K.
Watanabe
, and
K.
Shimanoe
, “
Selective detection of toluene using pulse-driven SnO2 micro gas sensors
,”
ACS Appl. Electron. Mater.
2
,
2913
–
2920
(
2020
).
https://doi.org/10.1021/acsaelm.0c00547
Google Scholar
Crossref
Search ADS
 
18.
J.-H.
Suh
,
I.
Cho
,
K.
Kang
,
S.-J.
Kweon
,
M.
Lee
,
H.-J.
Yoo
, and
I.
Park
, “
Fully integrated and portable semiconductor-type multi-gas sensing module for IoT applications
,”
Sens. Actuators, B
265
,
660
–
667
(
2018
).
https://doi.org/10.1016/j.snb.2018.03.099
Google Scholar
Crossref
Search ADS
 
19.
J.
Geist
,
M.
Afridi
,
E. M.
Secula
,
D. G.
Seiler
,
R. P.
Khosla
,
D.
Herr
,
C.
Michael Garner
,
R.
McDonald
, and
A. C.
Diebold
, “
Temperature-programmed gas-sensing with microhotplates: An opportunity to enhance microelectronic gas sensor metrology
,” in
AIP Conference Proceedings
(
AIP
,
Albany, New York
,
2009
), pp.
207
–
211
.
Google Scholar
Crossref
Search ADS
 
20.
H.
Ji
,
Z.
Yuan
,
H.
Zhu
,
W.
Qin
,
H.
Wang
, and
F.
Meng
, “
Dynamic temperature modulation measurement of VOC gases based on SnO2 gas sensor
,”
IEEE Sens. J.
22
,
14708
–
14716
(
2022
).
https://doi.org/10.1109/jsen.2022.3184511
Google Scholar
Crossref
Search ADS
 
21.
T. A.
Kunt
,
T. J.
McAvoy
,
R. E.
Cavicchi
, and
S.
Semancik
, “
Optimization of temperature programmed sensing for gas identification using micro-hotplate sensors
,”
Sens. Actuators, B
53
,
24
–
43
(
1998
).
https://doi.org/10.1016/s0925-4005(98)00244-5
Google Scholar
Crossref
Search ADS
 
22.
G.-Y.
Miao
,
S.-S.
Chen
,
Y.-J.
Wang
,
Z.
Guo
, and
X.-J.
Huang
, “
SnO2 nanostructures exposed with various crystal facets for temperature-modulated sensing of volatile organic compounds
,”
ACS Appl. Nano Mater.
5
,
10636
–
10644
(
2022
).
https://doi.org/10.1021/acsanm.2c01963
Google Scholar
Crossref
Search ADS
 
23.
A.
Ortega
,
S.
Marco
,
A.
Perera
,
T.
Šundic
,
A.
Pardo
, and
J.
Samitier
, “
An intelligent detector based on temperature modulation of a gas sensor with a digital signal processor
,”
Sens. Actuators, B
78
,
32
–
39
(
2001
).
https://doi.org/10.1016/s0925-4005(01)00788-2
Google Scholar
Crossref
Search ADS
 
24.
T.
Baur
,
J.
Amann
,
C.
Schultealbert
, and
A.
Schütze
, “
Field study of metal oxide semiconductor gas sensors in temperature cycled operation for selective VOC monitoring in indoor air
,”
Atmosphere
12
,
647
(
2021
).
https://doi.org/10.3390/atmos12050647
Google Scholar
Crossref
Search ADS
 
25.
W.
Tang
,
Z.
Chen
,
Z.
Song
,
C.
Wang
,
Z.
Wan
,
C. L. J.
Chan
,
Z.
Chen
,
W.
Ye
, and
Z.
Fan
, “
Microheater integrated nanotube array gas sensor for parts-per-trillion level gas detection and single sensor-based gas discrimination
,”
ACS Nano
16
,
10968
–
10978
(
2022
).
https://doi.org/10.1021/acsnano.2c03372
Google Scholar
Crossref
Search ADS
PubMed
 
26.
T.
Baur
,
A.
Schütze
, and
T.
Sauerwald
, “
Optimierung des temperaturzyklischen Betriebs von Halbleitergassensoren
,”
Tech. Mess.
82
,
187
–
195
(
2015
).
https://doi.org/10.1515/teme-2014-0007
Google Scholar
Crossref
Search ADS
 
27.
Y.
Robin
,
P.
Goodarzi
,
T.
Baur
,
C.
Schultealbert
,
A.
Schütze
, and
T.
Schneider
, “
Machine learning based calibration time reduction for gas sensors in temperature cycled operation
,” in
2021 IEEE International Instrumentation and Measurement Technology Conference (I2MTC)
(
IEEE
,
2021
), pp.
1
–
6
.
Google Scholar
Crossref
Search ADS
 
28.
M.
Jaegle
,
J.
Wöllenstein
,
T.
Meisinger
,
H.
Böttner
,
G.
Müller
,
T.
Becker
, and
C.
Bosch-vBraunmühl
, “
Micromachined thin film SnO2 gas sensors in temperature-pulsed operation mode
,”
Sens. Actuators, B
57
,
130
–
134
(
1999
).
https://doi.org/10.1016/s0925-4005(99)00074-x
Google Scholar
Crossref
Search ADS
 
29.
F.
Parret
,
P.
Ménini
,
A.
Martinez
,
K.
Soulantica
,
A.
Maisonnat
, and
B.
Chaudret
, “
Improvement of micromachined SnO2 gas sensors selectivity by optimised dynamic temperature operating mode
,”
Sens. Actuators, B
118
,
276
–
282
(
2006
).
https://doi.org/10.1016/j.snb.2006.04.055
Google Scholar
Crossref
Search ADS
 
30.
Z.
Wu
,
H.
Zhang
,
H.
Ji
,
Z.
Yuan
, and
F.
Meng
, “
Novel combined waveform temperature modulation method of NiO-In2O3 based gas sensor for measuring and identifying VOC gases
,”
J. Alloys Compd.
918
,
165510
(
2022
).
https://doi.org/10.1016/j.jallcom.2022.165510
Google Scholar
Crossref
Search ADS
 
31.
J.
Ding
,
T. J.
McAvoy
,
R. E.
Cavicchi
, and
S.
Semancik
, “
Surface state trapping models for SnO2-based microhotplate sensors
,”
Sens. Actuators, B
77
,
597
–
613
(
2001
).
https://doi.org/10.1016/s0925-4005(01)00765-1
Google Scholar
Crossref
Search ADS
 
32.
A.
Fort
,
M.
Mugnaini
,
S.
Rocchi
,
M. B.
Serrano-Santos
,
V.
Vignoli
, and
R.
Spinicci
, “
Simplified models for SnO2 sensors during chemical and thermal transients in mixtures of inert, oxidizing and reducing gases
,”
Sens. Actuators, B
124
,
245
–
259
(
2007
).
https://doi.org/10.1016/j.snb.2006.12.030
Google Scholar
Crossref
Search ADS
 
33.
C.
Schultealbert
,
T.
Baur
,
A.
Schütze
,
S.
Böttcher
, and
T.
Sauerwald
, “
A novel approach towards calibrated measurement of trace gases using metal oxide semiconductor sensors
,”
Sens. Actuators, B
239
,
390
–
396
(
2017
).
https://doi.org/10.1016/j.snb.2016.08.002
Google Scholar
Crossref
Search ADS
 
34.
A.
Miquel-Ibarz
,
J.
Burgués
, and
S.
Marco
, “
Global calibration models for temperature-modulated metal oxide gas sensors: A strategy to reduce calibration costs
,”
Sens. Actuators, B
350
,
130769
(
2022
).
https://doi.org/10.1016/j.snb.2021.130769
Google Scholar
Crossref
Search ADS
 
35.
Z.-H.
Zhou
, “
A brief introduction to weakly supervised learning
,”
Natl. Sci. Rev.
5
,
44
–
53
(
2018
).
https://doi.org/10.1093/nsr/nwx106
Google Scholar
Crossref
Search ADS
 
36.
Y.
Robin
,
J.
Amann
,
P.
Goodarzi
,
T.
Schneider
,
A.
Schütze
, and
C.
Bur
, “
Deep learning based calibration time reduction for MOS gas sensors with transfer learning
,”
Atmosphere
13
,
1614
(
2022
).
https://doi.org/10.3390/atmos13101614
Google Scholar
Crossref
Search ADS
 
37.
L.
Li
,
J. C.
Snyder
,
I. M.
Pelaschier
,
J.
Huang
,
U.-N.
Niranjan
,
P.
Duncan
,
M.
Rupp
,
K.-R.
Müller
, and
K.
Burke
, “
Understanding machine-learned density functionals: Understanding machine-learned density functionals
,”
Int. J. Quantum Chem.
116
,
819
–
833
(
2016
).
https://doi.org/10.1002/qua.25040
Google Scholar
Crossref
Search ADS
 
38.
J.
Chu
,
W.
Li
,
X.
Yang
,
Y.
Wu
,
D.
Wang
,
A.
Yang
,
H.
Yuan
,
X.
Wang
,
Y.
Li
, and
M.
Rong
, “
Identification of gas mixtures via sensor array combining with neural networks
,”
Sens. Actuators, B
329
,
129090
(
2021
).
https://doi.org/10.1016/j.snb.2020.129090
Google Scholar
Crossref
Search ADS
 
39.
H.-R.
Hou
,
A. J.
Lilienthal
, and
Q.-H.
Meng
, “
Gas source declaration with tetrahedral sensing geometries and median value filtering extreme learning machine
,”
IEEE Access
8
,
7227
–
7235
(
2020
).
https://doi.org/10.1109/access.2019.2963059
Google Scholar
Crossref
Search ADS
 
40.
P.
Jia
,
H.
Liu
,
S.
Wang
, and
P.
Wang
, “
Research on a mine gas concentration forecasting model based on a GRU network
,”
IEEE Access
8
,
38023
–
38031
(
2020
).
https://doi.org/10.1109/access.2020.2975257
Google Scholar
Crossref
Search ADS
 
41.
D.
Kwon
,
G.
Jung
,
W.
Shin
,
Y.
Jeong
,
S.
Hong
,
S.
Oh
,
J.-H.
Bae
,
B.-G.
Park
, and
J.-H.
Lee
, “
Low-power and reliable gas sensing system based on recurrent neural networks
,”
Sens. Actuators, B
340
,
129258
(
2021
).
https://doi.org/10.1016/j.snb.2020.129258
Google Scholar
Crossref
Search ADS
 
42.
D.
Ma
,
J.
Gao
,
Z.
Zhang
, and
H.
Zhao
, “
Gas recognition method based on the deep learning model of sensor array response map
,”
Sens. Actuators, B
330
,
129349
(
2021
).
https://doi.org/10.1016/j.snb.2020.129349
Google Scholar
Crossref
Search ADS
 
43.
X.
Pan
,
Z.
Zhang
,
H.
Zhang
,
Z.
Wen
,
W.
Ye
,
Y.
Yang
,
J.
Ma
, and
X.
Zhao
, “
A fast and robust mixture gases identification and concentration detection algorithm based on attention mechanism equipped recurrent neural network with double loss function
,”
Sens. Actuators, B
342
,
129982
(
2021
).
https://doi.org/10.1016/j.snb.2021.129982
Google Scholar
Crossref
Search ADS
 
44.
X.
Wu
,
Z.
Zhao
, and
L.
Wang
, “
Deep belief network based coal mine methane sensor data classification
,”
J. Phys.: Conf. Ser.
1302
,
032013
(
2019
).
https://doi.org/10.1088/1742-6596/1302/3/032013
Google Scholar
Crossref
Search ADS
 
45.
M.
Schweizer-Berberich
,
M.
Zdralek
,
U.
Weimar
,
W.
Göpel
,
T.
Viard
,
D.
Martinez
,
A.
Seube
, and
A.
Peyre-Lavigne
, “
Pulsed mode of operation and artificial neural network evaluation for improving the CO selectivity of SnO2 gas sensors
,”
Sens. Actuators, B
65
,
91
–
93
(
2000
).
https://doi.org/10.1016/s0925-4005(99)00333-0
Google Scholar
Crossref
Search ADS
 
46.
Y.
Hu
,
Y.
Tian
,
Y.
Zhuang
,
C.
Zhao
, and
F.
Wang
, “
Rapid gas sensing based on pulse heating and deep learning
,” in
2021 IEEE 34th International Conference on Micro Electro Mechanical Systems (MEMS)
(
IEEE
,
Gainesville, FL, USA
,
2021
), pp.
438
–
441
.
Google Scholar
Crossref
Search ADS
 
47.
H. V.
Dang
,
H.
Tran-Ngoc
,
T. V.
Nguyen
,
T.
Bui-Tien
,
G.
De Roeck
, and
H. X.
Nguyen
, “
Data-driven structural health monitoring using feature fusion and hybrid deep learning
,”
IEEE Trans. Autom. Sci. Eng.
18
,
2087
–
2103
(
2021
).
https://doi.org/10.1109/tase.2020.3034401
Google Scholar
Crossref
Search ADS
 
48.
Y.
Du
,
Y.
Huang
,
G.
Wan
, and
P.
He
, “
Deep learning-based cyber–physical feature fusion for anomaly detection in industrial control systems
,”
Mathematics
10
,
4373
(
2022
).
https://doi.org/10.3390/math10224373
Google Scholar
Crossref
Search ADS
 
49.
Y.
Du
,
W.
Song
,
Q.
He
,
D.
Huang
,
A.
Liotta
, and
C.
Su
, “
Deep learning with multi-scale feature fusion in remote sensing for automatic oceanic eddy detection
,”
Inf. Fusion
49
,
89
–
99
(
2019
).
https://doi.org/10.1016/j.inffus.2018.09.006
Google Scholar
Crossref
Search ADS
 
50.
J.
Liu
,
Y.
Luo
,
L.
Ge
,
W.
Zeng
,
Z.
Rao
, and
X.
Xiao
, “
An intelligent online drunk driving detection system based on multi-sensor fusion technology
,”
Sensors
22
,
8460
(
2022
).
https://doi.org/10.3390/s22218460
Google Scholar
Crossref
Search ADS
PubMed
 
51.
N.
Saito
,
K.
Watanabe
,
H.
Haneda
,
I.
Sakaguchi
, and
K.
Shimanoe
, “
Highly sensitive ethanol gas sensor using pyramid-shaped ZnO particles with (0001) basal plane
,”
J. Phys. Chem. C
122
,
7353
–
7360
(
2018
).
https://doi.org/10.1021/acs.jpcc.8b01936
Google Scholar
Crossref
Search ADS
 
52.
T.
Baur
,
M.
Bastuck
,
C.
Schultealbert
,
T.
Sauerwald
, and
A.
Schütze
, “
Random gas mixtures for efficient gas sensor calibration
,”
J. Sens. Sens. Syst.
9
,
411
–
424
(
2020
).
https://doi.org/10.5194/jsss-9-411-2020
Google Scholar
Crossref
Search ADS
 
53.
A.
Fort
,
S.
Rocchi
,
M. B.
Serrano-Santos
,
R.
Spinicci
, and
V.
Vignoli
, “
Surface state model for conductance responses during thermal-modulation of SnO2-based thick film sensors: Part I—Model derivation
,”
IEEE Trans. Instrum. Meas.
55
,
2102
–
2106
(
2006
).
https://doi.org/10.1109/tim.2006.887118
Google Scholar
Crossref
Search ADS
 
54.
Y.
Zhuang
 (
2023
). “
A deep learning approach for gas sensor data regression: Incorporating surface state model and GRU-based model
,” Dataset. https://github.com/HWare-magic/GRU_regression_gas-sensor
Google Scholar
 
© 2024 Author(s). All article content, except where otherwise noted, is licensed under a Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

Supplementary Material

Supplementary materials- zip file