Gas detection method using transient current fluctuation

Abstract

The present invention relates to a gas detection method using transient current fluctuation data that can dramatically improve the detection characteristics of individual gases and mixed gases by extracting transient current fluctuation (ΔI) data from current value data over time measured by a metal oxide-based gas sensor and applying it to an artificial intelligence model. The gas detection method using transient current fluctuation data according to the present invention is characterized by extracting transient current fluctuation (ΔI) data from 'current value over time' data measured by a gas sensor, detecting a specific gas adsorbed to the gas sensor from the transient current fluctuation (ΔI) data, and the transient current fluctuation (ΔI) data being the 'amount of change in current value per unit time' converted from the 'current value over time' (see Equation 1 below). (Equation 1) ΔIn = In + 1 - In (n represents a specific time in the time series flow)

Description

Technology Field

[0001] The present invention relates to a gas detection method using transient current fluctuation data, and more specifically, to a gas detection method using transient current fluctuation data that can dramatically improve the detection characteristics of individual gases and mixed gases by extracting transient current fluctuation (ΔI) data from current value data over time measured by a metal oxide-based gas sensor and applying it to an artificial intelligence model. Background Technology

[0002] The demand for gas sensors is growing due to rapid climate change and the increase in industrial pollutants such as fine dust and toxic gases. Gas sensors play a crucial role in monitoring air quality by measuring and identifying the types and concentrations of gases and particles in the atmosphere.

[0003] Gas sensors are classified in various ways based on their detection principles; among them, semiconductor-metal oxide (SMO) based gas sensors, which detect gases by measuring changes in electrical conductivity, are gaining attention. SMO-based gas sensors have the advantages of a simple structure, low manufacturing costs, and excellent durability and safety. Furthermore, they exhibit high sensitivity to ultra-trace gas molecules.

[0004] However, SMO-based gas sensors do not exhibit excellent gas detection selectivity in situations where various gases are mixed. To improve this, active research is being conducted on gas sensors capable of detecting multi-species gases, as well as on various nanostructures, functionalizations, and post-processing techniques.

[0005] For example, Non-patent Literature 1 presents a technology for fabricating a gas sensor by individually growing various types of SMO thin films (SnO2, In2O3, WO3, CuO) and then integrating multiple SMO thin films onto a single substrate. However, this method has the disadvantage that the heating method is complex and a separate compensation circuit is required because each SMO thin film operates at a different temperature and has different resistances.

[0006] As another example, Non-patent Literature 2 presents a gas sensor that uses the same SMO, namely In2O3, but functionalizes In2O3 with Au and Ag. This method has the advantage of reducing the time and cost required for sensor fabrication and consuming less power. However, it has the disadvantage of low gas selectivity because there is not a significant difference in characteristics between the individual functionalized sensors, and additional post-processing or screening techniques are required to compensate for this.

[0007] Meanwhile, the recent rapid advancement of artificial intelligence (AI) is revolutionizing sensor data analysis. These AI technologies facilitate rapid data analysis and enable the prediction of molecular types and concentrations by identifying patterns. In particular, AI models such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) demonstrate high accuracy in identifying gases with limited samples and short response times, without the need for initial data processing (see Patent Documents 1 to 3).

[0008] However, while Recurrent Neural Networks (RNNs) demonstrate excellent pattern recognition capabilities in time-series data, RNN models with diverse functions simultaneously suffer from the disadvantages of redundancy and complexity. Prior art literature

[0009] Korean Published Patent Application No. 2023-0134238 (Published Sep. 21, 2023) Korean Published Patent Application No. 2022-0142030 (Published Oct. 21, 2022) Korean Published Patent Application No. 2021-0109400 (Published Sep. 6, 2021) Korean Published Patent Application No. 2023-0068820 (Published May 18, 2023)

[0010] M. Kang, I. Cho, J. Park, J. Jeong, K. Lee, B. Lee, D. Del Orbe Henriquez, K. Yoon, I. Park, High Accuracy Real-Time Multi-Gas Identification by a Batch-Uniform Gas Sensor Array and Deep Learning Algorithm, ACS Sens. 7 (2022) 430??440. .K. Lee, I. Cho, M. Kang, J. Jeong, M. Choi, K. Y. Woo, K.-J. Yoon, Y.-H. Cho, I. Park, Ultra-Low-Power E-Nose System Based on Multi-Micro-LED-Integrated, Nanostructured Gas Sensors and Deep Learning, ACS Nano 17 (2023) 539??551. . The problem to be solved

[0011] The present invention was devised to solve the above-mentioned problems, and aims to provide a gas detection method using transient current fluctuation data that can accurately detect not only individual gases but also mixed gases in an environment where various gases coexist.

[0012] In addition, another objective of the present invention is to provide a technology for effectively manufacturing various types of SMO unit sensors through a very simple method in manufacturing a metal oxide-based gas sensor having a plurality of SMO unit sensors on a single substrate.

[0013] In addition, the present invention has another objective of presenting a technology that can dramatically improve the detection characteristics of individual gases and mixed gases by extracting transient current fluctuation (ΔI) data from time-dependent current value data measured by a metal oxide-based gas sensor and applying it to an artificial intelligence model. means of solving the problem

[0014] A gas detection method using transient current fluctuation data according to the present invention for achieving the above objective is characterized by extracting transient current fluctuation (ΔI) data from 'current value over time' data measured by a gas sensor, detecting a specific gas adsorbed to the gas sensor from the transient current fluctuation (ΔI) data, and the transient current fluctuation (ΔI) data being the 'amount of change in current value per unit time' converted from the 'current value over time' (see Equation 1 below).

[0015] (Equation 1) ΔI n = I n+1 - I n

[0016] ( n (refers to a specific time in a time series flow)

[0017] By inputting transient current fluctuation (ΔI) data into an artificial intelligence model, specific gases adsorbed to the gas sensor can be detected.

[0018] The artificial intelligence model is a Sparse Recurrent Neural Networks (SRNN) model. In addition, the artificial intelligence model can utilize Recurrent Neural Networks (RNN) models.

[0019] The gas sensor has a structure in which multiple SMO unit sensors are equipped on a single substrate.

[0020] Each of the multiple SMO unit sensors is composed of a different material, and the multiple SMO unit sensors are manufactured through multiple shadow masks with different aperture arrangements and multiple sputtering processes.

[0021] The deposition material for each sputtering process is different, the aperture region corresponds to the SMO unit sensor region, and the deposition material is a metal oxide (SMO) or catalyst layer material having semiconductor properties.

[0022] In addition to varying the deposition material during each sputtering process, shadow masks with different aperture arrangements are used during each sputtering process, or the same shadow mask is used but the substrate or shadow mask is rotated at a certain angle, thereby allowing multiple SMO unit sensors formed on the substrate to be composed of different materials. Effects of the invention

[0023] The gas detection method using transient current fluctuation data according to the present invention has the following effects.

[0024] By incorporating a new concept called transient current fluctuation (ΔI) data into an artificial intelligence model, the accuracy of gas detection and the efficiency of analysis can be improved. Brief explanation of the drawing

[0025] FIG. 1 is a schematic diagram of a shadow mask for manufacturing a gas sensor according to one embodiment of the present invention. FIG. 2 is a schematic diagram illustrating a method for manufacturing a gas sensor according to one embodiment of the present invention. FIGS. 3 and FIGS. 4 are schematic diagrams of a gas sensor manufactured according to one embodiment of the present invention. Figure 5 is a schematic diagram of the experimental apparatus applied in Experimental Example 2. Figure 6 shows the conduction band (E) of the SMO thin film. c ) and valence band (E v This is a schematic diagram showing the energy levels of ). Figure 7(a) is data of 'current value over time' measured by a gas sensor, and Figure 7(b) is the 'current value over time' of Figure 7(a) converted into 'change in current value per unit time', i.e., transient current fluctuation (ΔI) data. Figures 8a and 8b are schematic diagrams illustrating that the current value of the SMO thin film decreases or increases depending on the oxidizing or reducing properties of the gas to be detected. Figure 9 shows the standard deviation for the 'current value over time' data of Figure 7(a) and the standard deviation for the transient current fluctuation (ΔI) data of Figure 7(b). Figure 10 shows the t-SNE plot of the SRNN model, where (a) is the case where the model was trained using transient current fluctuation (ΔI) data, and (b) is the case where the model was trained using only 'current value over time' data. Figure 11 shows the feature weights for the 'current value over time' (response), ΔI, and absolute ΔI ([ΔI]) of the RNN and SRNN. Specific details for implementing the invention

[0026] The present invention presents a technology that can dramatically improve the selectivity and accuracy of gas detection by combining a novel concept called transient current fluctuation (ΔI) with artificial intelligence.

[0027] Transient current fluctuation (ΔI)

[0028] In the present invention, transient current fluctuation (ΔI) refers to the 'amount of change in current value per unit time' (see Equation 1 below), and transient current fluctuation (ΔI) is data converted from the 'current value over time' measured by a gas sensor.

[0029] (Equation 1) ΔI n = I n+1 - I n

[0030] ( n(refers to a specific time in a time series flow)

[0031] Resistance-changing gas sensors, such as metal oxide (SMO)-based gas sensors, detect gas through changes in current values—that is, current values ​​over time. However, the reliability of gas detection based on 'current values ​​over time' using such resistance-changing gas sensors is limited to a certain level due to the presence of current values ​​similar to those of the specific gas to be detected, as well as various current values ​​corresponding to noise. Accordingly, as described in the 'Description of Background of the Invention' above, a method has been proposed to improve the accuracy of gas detection by integrating the data measured by the gas sensor ('current values ​​over time') with an artificial intelligence model. However, even when integrated with AI models such as Recurrent Neural Networks (RNNs), unnecessary neuron weights are assigned to the aforementioned similar current values ​​and noise current values, resulting in a decrease in both the accuracy of gas detection and the efficiency of analysis.

[0032] Therefore, to enhance the accuracy of gas detection by integrating with an artificial intelligence (AI) model, it is necessary to input data with minimized neuron weights into the model; furthermore, the data itself, with its minimized neuron weights, must possess high reliability. In other words, by inputting highly reliable data with minimized neuron weights into the AI ​​model, it is possible to secure not only the accuracy of gas detection but also the efficiency of analysis.

[0033] A new concept that can satisfy the minimization of these neuron weights and high reliability is transient current fluctuation (ΔI) data.

[0034] The 'current value over time', which is data measured by a metal oxide-based gas sensor, is data that continuously represents the measured current value over time, whereas the transient current fluctuation (ΔI), which is the 'change in current value per unit time' converted from the 'current value over time' data, is data that simplifies the 'current value over time' into unit time intervals.

[0035] In terms of gas detection, when observing the adsorption and desorption of the target gas on the gas sensor through 'current value over time' data, a high current value is observed at a specific time when the target gas is adsorbed. Subsequently, a low current value is observed when the adsorbed target gas is desorbed from the gas sensor. Here, although the current value increases or decreases during adsorption depending on the type of target gas, for the sake of convenience, the explanation will use the example where the current value increases during adsorption.

[0036] However, since 'current values ​​over time' data represents current values ​​continuously, it is expressed as if the current values ​​change in a continuous manner during the process of adsorption and desorption. In other words, it is expressed as the current values ​​changing in a time series between the current values ​​during adsorption and desorption. To put it hyperbolically, if we assume the current value during adsorption is 1 and the current value during desorption is 0, it is represented as existing current values ​​in the range of 0 to 1 between adsorption and desorption. Although only the current value of 1 (during adsorption) and the current value of 0 (during desorption) are required for gas detection, the accuracy of gas detection inevitably decreases because current values ​​in the 0 to 1 range exist in the 'current values ​​over time' data. Furthermore, even if this 'current values ​​over time' data is input into an artificial intelligence model such as a Recurrent Neural Network (RNN), neuron weights are assigned to current values ​​in the 0 to 1 range in addition to the current value of 1 (during adsorption) and the current value of 0 (during desorption), which not only limits the accuracy of gas detection but also reduces the efficiency of analysis.

[0037] On the other hand, the transient current fluctuation (ΔI) data, which converts the 'current value over time' into the 'change in current value per unit time,' is used to highlight the data during adsorption.

[0038] The current values ​​measured by the gas sensor can be broadly classified into current values ​​during adsorption, current values ​​during desorption, current values ​​while desorption is maintained, current values ​​while adsorption is maintained, and the current values ​​between the aforementioned adsorption and desorption.

[0039] When viewing these current values ​​in terms of the 'change in current value per unit time,' the current value during adsorption shows a significant change because it fluctuates substantially compared to the current value in the previous state.

[0040] In contrast, in all states except adsorption, the current value remains constant or increases or decreases at a constant slope. Here, the constant slope refers to the slope of the current value over time. Maintaining a constant value or increasing or decreasing at a constant slope implies that there is almost no fluctuation in terms of the 'change in current value per unit time.' When the current value remains constant, there is no change in the current value over time, so the 'change in current value per unit time' naturally remains unchanged. Similarly, when the current value increases or decreases at a constant slope, it implies that the change in the current value over time is constant, so there appears to be almost no fluctuation in terms of the 'change in current value per unit time.'

[0041] In this way, when analyzing the measurement data of the gas sensor using the 'change in current value per unit time,' that is, the transient current fluctuation (ΔI) data, only the data during adsorption is highlighted, and the data for the remaining states is considered to have almost no change.

[0042] The above description is examined in conjunction with the experimental examples of the present invention as follows. Fig. 7(a) shows data of 'current values ​​over time' measured by a gas sensor, and Fig. 7(b) shows the 'current values ​​over time' of Fig. 7(a) converted into 'change in current value per unit time,' i.e., transient current fluctuation (ΔI) data. Furthermore, in Figs. 7(a) and (b), the top graph relates to NO2 gas, the middle graph relates to HCHO gas, and the bottom graph relates to a mixed gas of NO2 and HCHO. In addition, in the experiments of Figs. 7(a) and (b), the target gas (NO2, HCHO, NO2+HCHO) was injected at 2 to 4 minutes and 6 to 8 minutes, and an inert gas (N2) was injected during the remaining time periods.

[0043] For example, in the graph of the mixed gas of NO2 and HCHO in Fig. 7, the current value data over time in Fig. 7(a) is expressed in a continuous form during the process of adsorption and desorption of the mixed gas. Specifically, in the 2–4 minute and 6–8 minute time periods when the mixed gas is supplied, the current value initially decreases sharply and then remains constant, while in the 4–6 minute and 8–12 minute time periods when N2 is supplied, the current value increases gradually. As such, the graph in Fig. 7(a) is expressed in a continuous form, with various types of current values ​​existing, such as sections where the current value decreases sharply, sections where the current value increases gradually, and sections where the current value remains constant.

[0044] In contrast, looking at the graph in Fig. 7(b), the transient current fluctuation (ΔI) decreases sharply only during the initial period when the mixed gas is supplied, while the remaining conditions show a transient current fluctuation (ΔI) of almost the same value. Furthermore, there is a significant quantitative difference between the transient current fluctuation (ΔI) during the mixed gas supply and the transient current fluctuation (ΔI) in the remaining conditions. Therefore, the adsorption of the target gas can be easily determined through the transient current fluctuation (ΔI) data in Fig. 7(b). In addition to the mixed gas of NO2 and HCHO, the NO2 and HCHO in Figs. 7(a) and (b) can be interpreted in the same way.

[0045] Previously, it was stated that in order to improve the accuracy of gas detection by integrating with an artificial intelligence model, data with minimized neuron weights must be input into the AI ​​model, and the reliability of the data with minimized neuron weights must be high. The 'change in current value per unit time,' that is, transient current fluctuation (ΔI) data, fits this description.

[0046] Since the transient current fluctuation (ΔI) data highlights the adsorption data, the reliability of the data is high, and as neuron weighting can be omitted for the remaining data excluding the adsorption data, the accuracy of gas detection is increased and the efficiency of analysis can be secured.

[0047] In addition, the experimental examples described below present NO2, HCHO, and a mixture of NO2 and HCHO as examples of gases to be detected, but the technical concept of the present invention is characterized by extracting transient current fluctuation (ΔI) data from 'current value over time' data generated during the adsorption and desorption of the gases to be detected, and detecting the gases to be detected from the transient current fluctuation (ΔI) data, so it can be applied regardless of the type of gases to be detected.

[0048] In the present invention, the artificial intelligence model combined with transient current fluctuation (ΔI) data is not particularly limited. However, as neuron weighting is minimized, the accuracy of gas detection can be ensured and the efficiency of analysis improved by applying a sparse recurrent neural network (SRNN) rather than a recurrent neural network (RNN), which requires noise removal to be performed beforehand. This is described in detail in Experimental Example 3 below.

[0049] Gas sensor manufacturing method

[0050] Meanwhile, as described above, transient current fluctuation (ΔI) data is obtained by converting the 'current value over time' into the 'amount of change in current value per unit time.' In order for the reliability of transient current fluctuation (ΔI) data to be high, the reliability of the 'current value over time' measured by the gas sensor must be high.

[0051] In order to increase the reliability of the 'current value over time' and the detection reliability of mixed gases, various methods can be devised as mentioned in the 'Technology Forming the Background of the Invention'; however, integrating SMO unit sensors composed of different materials onto a single substrate can be considered as a priority. That is, by mounting multiple SMO unit sensors on a substrate and utilizing gas sensors in which the constituent materials of each SMO unit sensor are different, highly reliable 'current values ​​over time' can be generated.

[0052] The present invention provides a technology that enables the manufacture of a gas sensor equipped with a plurality of SMO unit sensors through a very simple method.

[0053] The method for manufacturing a gas sensor according to the present invention utilizes a shadow mask and a sputtering process. Multiple shadow masks are applied, and the sputtering process is performed multiple times.

[0054] Each shadow mask is equipped with multiple apertures, and the area of ​​the apertures corresponds to the area of ​​a single SMO unit sensor. In addition, the arrangement of the apertures differs among the shadow masks.

[0055] The sputtering process is performed with a shadow mask mounted above the substrate, and the deposition material detached from the target passes through the openings of the shadow mask and is deposited on the substrate. During the process of performing the sputtering process multiple times, the material deposited through each sputtering process is different.

[0056] Since each shadow mask has a different aperture arrangement shape and the material deposited during each sputtering process is different, it is possible to apply different shadow masks during the process of performing the sputtering process multiple times, or to use the same shadow mask but rotate the shadow mask or substrate by a certain angle (e.g., 90 degrees or 180 degrees) to make it possible to configure multiple SMO unit sensors formed on the substrate with different materials.

[0057] For example, when forming eight types of SMO unit sensors composed of different materials on a substrate, this can be completed through two shadow masks and four sputtering processes. Specifically, the following examples are based on the formation of a total of 16 SMO unit sensors composed of eight different materials.

[0058] A substrate is prepared, and a first sputtering is performed using a first shadow mask.

[0059] The first shadow mask and the second shadow mask described below both have an area in which 16 openings can be formed, and the 16 openings are arranged in 4 rows (1 to 4 rows in the top-to-bottom direction) and 4 columns (1 to 4 columns in the left-to-right direction) and are spaced apart at regular intervals in the up-down-left-right directions (see FIG. 1).

[0060] As illustrated in FIG. 1, the first shadow mask (Shadow mask 1) is provided with 8 openings, and the 8 openings are a combination of 1 row and 2 rows. The second shadow mask (Shadow mask 2) is provided with 8 openings, and the 8 openings are a combination of 2 rows and 3 rows.

[0061] A first SMO thin film is deposited on each of the eight SMO unit sensor regions corresponding to the first and second rows on the substrate through first sputtering using a first shadow mask. In one embodiment, IGZO may be deposited as the first SMO thin film on each of the eight SMO unit sensor regions corresponding to the first and second rows (see FIG. 2(a)).

[0062] Next, the second sputtering is performed with the first shadow mask rotated 180 degrees horizontally. As the first shadow mask is rotated 180 degrees horizontally, the eight openings of the first shadow mask correspond to the openings in the third and fourth rows.

[0063] A second SMO thin film is deposited on each of the eight SMO unit sensor regions corresponding to the third and fourth rows on the substrate through a second sputtering process. In one embodiment, ZnO may be deposited as the second SMO thin film on each of the eight SMO unit sensor regions corresponding to the third and fourth rows (see FIG. 2 (b)).

[0064] Through the first sputtering and the second sputtering, all 16 SMO thin films of the SMO unit sensors are completely deposited. The first and second rows on the substrate consist of 8 first SMO thin films, and the third and fourth rows consist of 2nd SMO thin films.

[0065] Next, each of the eight types of SMO unit sensors is configured differently by depositing different catalyst layers on each SMO thin film (the first SMO thin film or the second SMO thin film) through the third sputtering and the fourth sputtering.

[0066] A third sputtering step is performed using a first shadow mask while the substrate is rotated 90 degrees clockwise. As the substrate is rotated 90 degrees clockwise, four first SMO thin films and four second SMO thin films in rows 3 and 4 of the substrate are exposed through the openings of the first shadow mask.

[0067] A first catalyst layer is formed on the first SMO thin film and the second SMO thin film constituting rows 3 and 4 through a third sputtering process. In one embodiment, the first catalyst layer can be formed of Ag (see (c) of FIG. 2).

[0068] Next, while maintaining the substrate in a state where it is rotated 90 degrees clockwise, a fourth sputtering is performed using a second shadow mask. As described above, the second shadow mask has eight openings, and the eight openings are a combination of two rows and three rows. Accordingly, four first SMO thin films and four second SMO thin films constituting the second and third rows before the substrate is rotated are exposed by the openings of the second shadow mask. At this time, a first catalyst layer is deposited on each of the two first SMO thin films and two second SMO thin films of the exposed third row by the third sputtering described above.

[0069] A second catalyst layer is formed on the first SMO thin film and the second SMO thin film constituting rows 2 and 3 by the fourth sputtering. In one embodiment, the second catalyst layer can be formed of Pt (see FIG. 2 (d)).

[0070] Through the above first to fourth sputtering, a total of 16 SMO unit sensors of 8 different materials are manufactured on a substrate, with 2 of each. In the above-described embodiment, the materials of the first SMO thin film, the second SMO thin film, the first catalyst layer, and the second catalyst layer were exemplified, but are not limited thereto.

[0071] As an example of the specific materials of the above-described embodiment, as shown in FIG. 3, it can be summarized that two SMO unit sensors composed of IGZO (row 1-column 1, column 2), two SMO unit sensors composed of ZnO (row 1-column 3, column 4), two SMO unit sensors with Pt deposited on IGZO (row 2-column 1, column 2), two SMO unit sensors with Pt deposited on ZnO (row 2-column 3, column 4), two SMO unit sensors with Ag and Pt deposited on IGZO (row 3-column 1, column 2), two SMO unit sensors with Ag and Pt deposited on ZnO (row 3-column 3, column 4), two SMO unit sensors with Ag deposited on IGZO (row 4-column 1, column 2), and two SMO unit sensors with Ag and Pt deposited on ZnO (row 4-column 3, column 4) are formed.

[0072] As explained above, when performing multiple sputtering processes using a shadow mask, the deposition material (SMO material and catalyst layer material) for each sputtering process is different, and a shadow mask with a different aperture arrangement for each sputtering process is used, or the same shadow mask is used but the substrate or shadow mask is rotated at a certain angle (e.g., 90 degrees or 180 degrees), thereby allowing multiple SMO unit sensors formed on the substrate to be composed of different materials.

[0073] Above, a gas detection method using transient current fluctuation data according to one embodiment of the present invention has been described. Below, the present invention will be explained in more detail through experimental examples.

[0074] Experimental Example 1: Fabrication of a Gas Sensor

[0075] A gas sensor having eight types of SMO unit sensors was fabricated using the gas sensor manufacturing method described in the above-described embodiment. A total of 16 SMO unit sensors were formed, with two of each. IGZO was used as the first SMO thin film and ZnO as the second SMO thin film, while Ag was used as the first catalyst layer and Pt as the second catalyst layer.

[0076] IGZO and ZnO were selected as SMO thin films due to their ability to adsorb various gases even at relatively low temperatures and their stable n-type semiconductor characteristics during AC sputtering. Pt and Ag nanoparticles act as catalysts to enhance gas adsorption and reaction on the surface of the SMO unit sensor, thereby improving sensitivity and reducing response time. IGZO and ZnO were deposited via AC sputtering, while Pt and Ag nanoparticles were deposited via DC sputtering.

[0077] In addition, interdigitated electrodes (channel length: 30 μm) made of Au were formed on each of the 16 individual SMO unit sensors. Fig. 4(a) is a schematic diagram of the gas sensor fabricated according to Experimental Example 1, and Fig. 4(b) is a photograph of the actual SMO unit sensor fabricated according to Experimental Example 1.

[0078] Experimental Example 2: Extraction of Transient Current Fluctuation (ΔI)

[0079] A gas detection experiment was conducted using the gas sensor fabricated according to Experimental Example 1.

[0080] As shown in Fig. 5, a gas chamber containing a Peltier device was prepared, and an array sensor fabricated according to Experimental Example 1 was mounted on the Peltier device of the gas chamber. A UV LED (λ: 365 nm, power: 80 W) for irradiating ultraviolet light onto the gas sensor was mounted on one side of the top of the gas chamber facing the gas sensor. While supplying the gas to be detected into the gas chamber, the Peltier device was maintained at 150°C, and ultraviolet light was irradiated onto the gas sensor through the UV LED. NO2, HCHO, and a mixture of NO2 and HCHO were each injected as the gas to be detected, and the injection of the gas to be detected for 2 minutes and the injection of the purging gas (N2) for 2 minutes were repeated. For reference, conventional SMO gas sensors require a temperature environment of 200 to 400°C for the adsorption and desorption of the reaction gas, but in the case of the present invention, it was confirmed that the adsorption and desorption of the target gas proceed smoothly even at a temperature of 150°C by combining a UV LED.

[0081] Before explaining the data measured by the gas sensor, we will explain the principle of adsorption and desorption of the gas to be detected in the gas sensor fabricated according to Experimental Example 1, and the principle of current generation according to the adsorption of the gas to be detected.

[0082] The IGZO and ZnO thin films of the gas sensor fabricated according to Experimental Example 1 exhibit n-type semiconductor characteristics due to oxygen vacancies generated during the sputtering process. Oxygen molecules in the air of the gas chamber are O 2- It is easily adsorbed onto the surface of the SMO thin film in this form, and as a result, the electron density on the surface of the SMO thin film decreases, along with the conduction band (E) of the SMO thin film. c ) and valence band (E v The energy level of ) increases (see Fig. 6). Oxygen molecules (O) on the surface of the SMO thin film 2- ) are holes (h) generated by ultraviolet irradiation +It can be detached from the surface of the SMO thin film in the form of free oxygen molecules (O2) by combining with ), and some of the free oxygen molecules (O2) combine with electrons generated by ultraviolet irradiation to form oxygen molecules (O 2- They can be weakly bound again to the surface of the SMO thin film in the form of ). Ag nanoparticles, Pt nanoparticles, or a mixture (or compound) of Ag and Pt nanoparticles immobilized on the surface of the SMO thin film act as electron sources. If the size of the catalyst nanoparticles is smaller than the wavelength of the incident ultraviolet light, a localized surface plasmon resonance (LSPR) phenomenon occurs on the surface of the catalyst nanoparticles, and the resonated electrons move to the SMO thin film, thereby improving the detection and response speed of the target gas. Since the LSPR effects of Ag nanoparticles and Pt nanoparticles are different, four types of catalytic effects can occur depending on the combination of catalyst nanoparticles.

[0083] Figure 7(a) shows the 'current values ​​over time' measured by the gas sensor according to the experiment described above. In Figure 7(a), the top graph is for NO2 gas, the middle graph is for HCHO gas, and the bottom graph is for a mixture of NO2 and HCHO. As previously described, in the experiment of Figure 7(a), the target gas (NO2, HCHO, NO2+HCHO) was injected at 5 ppm at 2 to 4 minutes and 6 to 8 minutes, and an inert gas (N2) was injected during the remaining time.

[0084] Figure 7(b) is a conversion of the ‘current value over time’ of Figure 7(a) into ‘change in current value per unit time’, that is, transient current fluctuation (ΔI) data (see Equation 1).

[0085] In the process of repeated adsorption and desorption of the gas to be detected, the result of Fig. 7(a) is expressed in a continuous form with various types of current values, such as a section where the current value decreases rapidly, a section where the current value increases gradually, and a section where the current value is maintained at a constant.

[0086] In contrast, looking at the results in FIG. 7(b), the transient current fluctuation (ΔI) decreases sharply only at the beginning of the time period when the gas to be detected is supplied, while the remaining states show a transient current fluctuation (ΔI) of almost the same value. In addition, there is a large quantitative difference between the transient current fluctuation (ΔI) when the gas to be detected is supplied and the transient current fluctuation (ΔI) in the remaining states. Therefore, it is easy to determine whether the gas to be detected is adsorbed through the transient current fluctuation (ΔI) data in FIG. 7(b).

[0087] Meanwhile, depending on the type of gas to be detected, the current value tends to increase or decrease, which is attributed to the oxidizing or reducing properties of the gas to be detected. In the case of NO2, when oxidizing NO2 is adsorbed onto the SMO unit sensor while the SMO unit sensor has n-type semiconductor characteristics as described above, electrons on the surface of the SMO thin film move to NO2, causing the current to decrease rapidly; consequently, the transient current fluctuation (ΔI) of NO2 also becomes less than 0 (ΔI < 0). When the adsorbed NO2 is desorbed, electrons move back to the SMO thin film, causing the current to gradually increase, and the transient current fluctuation (ΔI) also becomes greater than 0 (ΔI > 0) (see Fig. 8a). On the other hand, HCHO exhibits reducing properties and thus shows behavior opposite to that of NO2. That is, during adsorption, electrons from HCHO move to the SMO thin film, and as the current value increases, the transient current fluctuation (ΔI) also becomes greater than 0 (ΔI > 0); during desorption, electrons from the SMO thin film move to HCHO, and as the current value decreases, the transient current fluctuation (ΔI) also becomes less than 0 (ΔI < 0) (see Fig. 8b).

[0088] As seen above, it can be seen that the adsorption of the target gas can be clearly identified through the transient current fluctuation (ΔI) data in Fig. 7(b) rather than the 'current value over time' data in Fig. 7(a).

[0089] The gas detection utility of transient current fluctuation (ΔI) data compared to 'current value over time' data is also proven through standard deviation (σ) data. Figure 9 shows the standard deviation (red line) for the 'current value over time' data in Figure 7(a) and the standard deviation (blue line) for the transient current fluctuation (ΔI) data in Figure 7(b).

[0090] Referring to Figure 9, it can be seen that the standard deviation of the transient current fluctuation (ΔI) is significantly larger than the standard deviation of the 'current value over time'. In the case of NO2, for all 8 types of SMO unit sensors, the standard deviation of the transient current fluctuation (ΔI) is approximately 1, whereas the standard deviation of the 'current value over time' is approximately 0.1. In the case of HCHO, the standard deviation of the transient current fluctuation (ΔI) is approximately 0.4, while the standard deviation of the 'current value over time' is approximately 0.001. Furthermore, in the case of NO2+HCHO, the standard deviation of the transient current fluctuation (ΔI) is approximately 0.1, while the standard deviation of the 'current value over time' is approximately 0.01.

[0091] Experimental Example 3: Combination of Transient Current Fluctuation (ΔI) and Artificial Intelligence Model

[0092] An SRNN model was trained and evaluated using a dataset containing 3D input data and 2D label data. The dataset consists of 10 samples of NO2, 10 samples of HCHO, and 10 samples of a mixture of NO2 and HCHO. Each sample contains time-series data collected over 720 seconds from 8 types of SMO unit sensors and three class labels indicating the gas type. Each gas type was measured 7 times through 8 types of SMO unit sensors. For each SMO unit sensor vector, one measurement was randomly selected from the 7 measurements. That is, 10 samples were selected for each gas type in the dataset from 7 possible combinations. Additionally, two sets of 8 features were created by calculating both the transient current fluctuation (ΔI) value and the absolute ΔI value for each SMO unit sensor. The original data format was (30, 720, 8), but it was expanded to (30, 720, 24) with the addition of ΔI variations. Meanwhile, the output data format was maintained as (30, 3).

[0093] To ensure efficient training of the SRNN model, data normalization was performed to adjust all input features to a range between 0 and 1 based on their minimum and maximum values. For output labels, 3 columns are generated for each category, and binary values indicate the presence or absence of an attribute. This process is called one-hot encoding. The data set was split into a training set and a test set, with 80% (24 samples) allocated to training and 20% (6 samples) allocated to testing. Within the training set, 20% of the data was separately reserved for validation to monitor loss and implement early stopping with a patience parameter of 30. Model performance was evaluated on the test set using an accuracy metric, which measures the proportion of correctly classified instances. Monte Carlo cross validation (MCCV) was also used for comprehensive evaluation, where the data set was randomly split into training and test groups multiple times. This MCCV evaluation was repeated 4 times to evaluate the overall performance of the model.

[0094] The model architecture includes two RNN layers each with 64 neurons and three dense layers each with 64 neurons, and the parameters are optimized through hyperparameter tuning. The model can process 8 feature sets of sensor responses, namely 'current value over time' (response), ΔI values, and 24 feature sets including absolute ΔI. The RNN layers extract context information and deliver it to the dense layers. The final output layer, which has three neurons corresponding to each gas type, uses the softmax function to determine class probabilities. The model parameters ■, 盼 and the learning rate were set to 0.00005, 0.5 and 0.001, respectively.

[0095] Table 1 below briefly summarizes the accuracy of models trained with various sets of variables. In Table 1 below, RNN and SRNN are cases trained using only sensor response data, i.e., 'current values ​​over time' data, while RNN(ΔI) and SRNN(ΔI) are cases trained using transient current fluctuation (ΔI) data.

[0096] Referring to Table 1, it can be seen that when the SRNN model is trained using only 'current value over time' data, the gas detection accuracy is 0.87, whereas when trained using transient current fluctuation (ΔI) data (SRNN(ΔI)), the gas detection accuracy increases to 0.92 and the training time is reduced to 0.15 minutes.

[0097] [Table 1]

[0098]

[0099] Figure 10 shows the t-SNE plot of the SRNN model, where (a) is the case where the model was trained using transient current fluctuation (ΔI) data, and (b) is the case where the model was trained using only 'current value over time' data.

[0100] The t-SNE plot of the SRNN model visualizes the output of the last dense layer neuron prior to the output layer in two dimensions. When trained using only 'current values ​​over time' data (see Fig. 10(b)), the three gases in the MCCV subset (CV#4) formed distinct clusters with minimal overlap, whereas NO2 and the mixed gas (NO2+HCHO) in another MCCV subset (CV#1) showed feature similarities, resulting in no clear classification. On the other hand, when trained using transient current fluctuation (ΔI) data (see Fig. 10(a)), it was confirmed that sophisticated classification was possible by generating dense feature points.

[0101] Meanwhile, in neural network models, the magnitude of feature weights indicates the importance of individual features. Generally, a higher weight implies that the more influential the feature is. Figure 11 shows the feature weights for the 'current value over time' (response), ΔI, and absolute ΔI ([ΔI]) of RNN and SRNN, which were averaged through 4 iterations of MCCV using the test dataset.

[0102] Referring to Figure 11, in the case of the RNN, all input features had similar weights, which means they represent uniform importance across all features. On the other hand, the SRNN utilizing group normalization assigned lower weights to less important features. Additionally, in the SRNN, the ΔI variable had a higher weight than the response variable, but the absolute ΔI variable was assigned the smallest weight. This suggests that both the speed and direction of the sensor response change are important for gas identification. Through these results, it can be confirmed that including the ΔI variable in the SRNN model improves gas detection accuracy.

[0103] Table 2 below shows the gas detection characteristics using an SMO sensor and an artificial intelligence model according to the present invention and prior research.

[0104] Referring to Table 2, there are many existing research results showing results superior to the gas detection accuracy (92%) of the present invention, but there is a clear difference in that the existing research results are for gas detection of a single gas, whereas the present invention is the first research result for mixed gas detection.

[0105] [Table 2]

[0106]

[0107] [Reference]

[0108]

[23] M. Kang, I. Cho, J. Park, J. Jeong, K. Lee, B. Lee, D. Del Orbe Henriquez, K. Yoon, I. Park, High Accuracy Real-Time Multi-Gas Identification by a Batch-Uniform Gas Sensor Array and Deep Learning Algorithm, ACS Sens. 7 (2022) 430??440. .

[0109]

[24] K. Lee, I. Cho, M. Kang, J. Jeong, M. Choi, K.Y. Woo, K.-J. Yoon, Y.-H. Cho, I. Park, Ultra-Low-Power E-Nose System Based on Multi-Micro-LED-Integrated, Nanostructured Gas Sensors and Deep Learning, ACS Nano 17 (2023) 539??551. .

[0110]

[47] S. Acharyya, B. Jana, S. Nag, G. Saha, P.K. Guha, Single resistive sensor for selective detection of multiple VOCs employing SnO2 hollowspheres and machine learning algorithm: A proof of concept, Sensors and Actuators B: Chemical 321 (2020) 128484. .

[0111]

[48] S. Kanaparthi, S.G. Singh, Discrimination of gases with a single chemiresistive multi-gas sensor using temperature sweeping and machine learning, Sensors and Actuators B: Chemical 348 (2021) 130725. .

[0112]

[49] T. Wang, H. Ma, W. Jiang, H. Zhang, M. Zeng, J. Yang, X. Wang, K. Liu, R. Huang, Z. Yang, Type discrimination and concentration prediction towards ethanol using a machine learning??enhanced gas sensor array with different morphology-tuning characteristics, Phys. Chem. Chem. Phys. 23 (2021) 23933??23944. .

[0113]

[50] M.A.H. Khan, B. Thomson, R. Debnath, A. Motayed, M.V. Rao, Nanowire-Based Sensor Array for Detection of Cross-Sensitive Gases Using PCA and Machine Learning Algorithms, IEEE Sensors J. 20 (2020) 6020??6028. .

[0114]

[51] P. Peng, X. Zhao, X. Pan, W. Ye, Gas Classification Using Deep Convolutional Neural Networks, Sensors 18 (2018) 157. .

[0115]

[52] M. Tonezzer, Selective gas sensor based on one single SnO2 nanowire, Sensors and Actuators B: Chemical 288 (2019) 53??59. https: / / doi.org / 10.1016 / j.snb.2019.02.096.

Claims

Claim 1 A gas detection method using transient current fluctuation data, characterized by extracting transient current fluctuation (ΔI) data from 'current value over time' data measured by a gas sensor, detecting a specific gas adsorbed to the gas sensor from the transient current fluctuation (ΔI) data, and wherein the transient current fluctuation (ΔI) data is the 'amount of change in current value per unit time' converted from the 'current value over time' (see Equation 1 below). (Equation 1) ΔI n = I n+1 - I n ( n (refers to a specific time in a time series flow) Claim 2 A gas detection method using transient current fluctuation data according to claim 1, characterized by inputting transient current fluctuation (ΔI) data into an artificial intelligence model to detect a specific gas adsorbed to a gas sensor. Claim 3 A gas detection method using transient current fluctuation data according to claim 1, characterized in that the artificial intelligence model is a Sparse Recurrent Neural Networks (SRNN) model. Claim 4 A gas detection method using transient current fluctuation data according to claim 1, characterized in that the artificial intelligence model is an RNN (Recurrent Neural Networks) model. Claim 5 A gas detection method using transient current fluctuation data according to claim 1, characterized in that the gas sensor has a structure in which a plurality of SMO unit sensors are provided on a single substrate. Claim 6 A gas detection method using transient current fluctuation data according to claim 5, wherein each of the plurality of SMO unit sensors is composed of different materials, and the plurality of SMO unit sensors are manufactured through a plurality of shadow masks with different aperture arrangements and a plurality of sputtering processes. Claim 7 A gas detection method using transient current fluctuation data according to claim 6, characterized in that the deposition material of each sputtering process is different from one another, the aperture region corresponds to the SMO unit sensor region, and the deposition material is a metal oxide (SMO) having semiconductor properties or a catalyst layer material. Claim 8 A gas detection method using transient current fluctuation data according to claim 6, characterized in that, in addition to varying the deposition material during each sputtering process, a shadow mask with a different aperture arrangement during each sputtering process is used, or the same shadow mask is used but the substrate or shadow mask is rotated at a certain angle, thereby configuring a plurality of SMO unit sensors formed on a substrate with different materials.

Images