A gas detection method based on data dimension expansion and data drift correction

Abstract

The application discloses a gas detection method based on data dimension expansion and data drift correction, and belongs to the technical field of gas detection.The method comprises the following steps: integrating multiple test electrodes with different key parameters on a single MEMS gas sensor to prepare a composite MEMS gas sensor; connecting the test electrodes of the composite MEMS gas sensor to a multi-channel data acquisition device so that the composite MEMS gas sensor enters a working state; collecting response signals output by the test electrodes in different gas environments; integrating the response signals of the multiple-point test electrodes to construct a virtual sensor array data set; detecting and correcting data drift of the virtual sensor array data set to obtain a corrected virtual sensor array data set and input the corrected virtual sensor array data set into a classification algorithm, so that gas recognition is realized.The method realizes data dimension expansion based on array test electrodes, and combines a data drift detection and correction method based on a sliding window, so that the selectivity and recognition performance of gas detection are improved.

Description

Technical Field

[0001] This invention belongs to the field of gas detection technology and relates to a gas detection method based on data dimension expansion and data drift correction. Background Technology

[0002] With the continuous development of metal-oxide-semiconductor (MOS) gas sensor technology, its advantages such as miniaturization, low power consumption, and low cost have attracted widespread attention. In recent years, related research has significantly improved its key performance characteristics such as sensitivity and response / recovery time, enabling MOS sensors to be widely used in scenarios such as gas leak monitoring and industrial production process monitoring for the detection of volatile organic compounds. However, gas sensors still suffer from insufficient selectivity, and researchers have proposed temperature modulation testing methods. Although this method improves gas selectivity to some extent, MOS sensors still face significant data drift problems. Specifically, most gas recognition methods typically rely on training data to build a recognition model, which is then used to distinguish unknown samples in subsequent measurements. This process requires the sensor output signal to maintain good stability over time; otherwise, it will undermine the effectiveness of the existing model, leading to a decline in pattern recognition performance. Therefore, research on sensor data drift problems is of great significance. For example, researchers have proposed a compensation method based on a lightweight artificial neural network to address the drift problem caused by unstable factors such as temperature, humidity, and heating power in static sensor testing. However, existing research focuses on using machine learning methods to correct drift data in order to compensate for sensor data drift, while systematic research on the direction and intensity of data drift is still insufficient. Summary of the Invention

[0003] This invention provides a gas detection method based on data dimension expansion and data drift correction. It expands the data dimension based on array test electrodes and combines it with a sliding window-based data drift detection and correction method to improve the selectivity and recognition performance of gas detection.

[0004] This invention provides a gas detection method based on data dimensionality expansion and data drift correction, comprising:

[0005] Step 1: Integrate multiple test electrodes with different key parameters on a single MEMS gas sensor to fabricate a composite MEMS gas sensor;

[0006] Step 2: Connect each test electrode of the composite MEMS gas sensor to the multi-channel data acquisition device, and apply a specific operating voltage to the virtual sensor array to put it into working state;

[0007] Step 3: Collect the response signals output by each test electrode under different gas environments, and integrate the response signals of multiple test electrodes to construct a virtual sensor array dataset;

[0008] Step 4: Perform data drift detection and correction on the virtual sensor array dataset based on a sliding window to obtain the corrected virtual sensor array dataset;

[0009] Step 5: Input the corrected virtual sensor array dataset into the classification algorithm to achieve gas identification.

[0010] This invention discloses a gas detection method based on data dimensionality expansion and data drift correction. It utilizes a sliding window mechanism to accurately capture dynamic changes in data distribution, thereby estimating the direction and intensity of data drift, and combines this with linear discriminant analysis to effectively correct the data drift. Furthermore, the drift in sensor data originates from the signal acquisition end (measurement end). To further improve gas selectivity and enhance the sensor's anti-drift capability from the signal acquisition end, a data dimensionality expansion of the array test electrodes is proposed. Traditional gas sensors typically use only a single test electrode to acquire the response signal; when fluctuations occur at the measurement end, the output data is prone to drift. This invention integrates multiple test electrodes in different areas on the sensor's sensitive material surface, enabling a single sensor to simultaneously acquire multi-point response information. This is functionally equivalent to constructing multiple sensing units, thus mitigating the data drift problem caused by the signal acquisition end. Moreover, the multidimensional data acquired by this method can be used to construct a virtual sensor array, further improving the selectivity and recognition performance of gas detection. Attached Figure Description

[0011] Figure 1 This is a flowchart of a gas detection method based on data dimension expansion and data drift correction according to the present invention;

[0012] Figure 2 The diagram shows the response signals of two gas sensors with different sensing performances. (a) and (b) show the response signals of the two sensors with different sensing performances in dynamic working mode to a mixture of 200 ppm toluene and 10–25 ppm butanone. (c) is a comparison diagram of the response signals of the two gas sensors with different sensing performances under the condition of a mixture of 200 ppm toluene and 10 ppm butanone.

[0013] Figure 3 A schematic diagram of the sample structure;

[0014] Figure 4 A schematic diagram illustrating the construction of a virtual sensor array dataset;

[0015] Figure 5The dynamic response curves of different target gases are shown for the composite MEMS gas sensor when the operating voltage is a sinusoidal modulation waveform; (a) sinusoidal modulation waveform, (b) dynamic response curve of air, (c) dynamic response curve of 30 ppm xylene, (d) dynamic response curve of 60 ppm xylene, (e) dynamic response curve of 90 ppm xylene, and (f) dynamic response curve of 120 ppm xylene.

[0016] Figure 6 The diagram illustrates the drift correction process; (a) shows the drift direction results for air and 30–120 ppm xylene data estimated based on the sliding window theory; (b) shows the drift intensity results for air and 30–120 ppm xylene estimated based on the sliding window theory; and (c) shows the drift correction results for air and 30–120 ppm xylene based on linear discriminant analysis. Detailed Implementation

[0017] The specific embodiments of the method described in this invention will be described in detail below with reference to the accompanying drawings. This part is used to further explain the invention, but does not constitute a limitation on the scope of protection.

[0018] like Figure 1 As shown, a gas detection method based on data dimensionality expansion and data drift correction according to the present invention includes:

[0019] Step 1: Integrate multiple test electrodes with different key parameters onto a single MEMS gas sensor to fabricate a composite MEMS gas sensor, specifically as follows:

[0020] Step 1.1: Determine the number of test electrodes based on the complexity of the gas to be tested and the required data dimensions, and determine the electrode material, electrode geometry, and spatial layout between the electrodes.

[0021] In practice, the number of electrodes can be determined based on the complexity of the gas to be measured and the required data dimensionality; generally, the more electrodes there are, the higher the data dimensionality. To further obtain differentiated sensor responses, different electrode materials (such as gold, platinum, and graphene), electrode geometries (interdigitated electrode structures, parallel plate electrode structures, and grid electrode structures), and other relevant parameters of the test electrodes can be selected.

[0022] Step 1.2: Integrate the test electrodes with different parameters mentioned above onto the same chip using standard MEMS fabrication technology.

[0023] Step 1.3: Using drop-coating or inkjet printing, uniformly coat the surface of the integrated test electrode with a metal oxide semiconductor as a gas-sensitive material. Each test electrode and the gas-sensitive material constitute a composite sensing unit, and multiple composite sensing units constitute a virtual sensor array.

[0024] Step 1 transforms the inherent sensor performance inconsistencies in MEMS processes into an information gain advantage, thereby improving the selectivity and stability of gas detection. Specifically, a "virtual sensor array" is constructed by integrating multiple test electrodes with different preset parameters (such as materials and geometries) onto a single MEMS sensor. When this array is exposed to the target gas, each electrode will generate differentiated response signals, significantly increasing the data dimensionality. This step effectively utilizes the sensor performance inconsistencies that are difficult to completely eliminate in MEMS processes, turning them from a disadvantage into a favorable condition for providing diverse response information.

[0025] To illustrate the effectiveness of array test electrodes in improving gas detection performance, the response signals of two gas sensors with different sensing performances under dynamic operating modes were selected as research examples. In this example, a mixture of 200 ppm toluene and 10–25 ppm butanone was used as the target gas. Under temperature modulation conditions, a triangular wave modulation voltage with an amplitude of 6–2 V and a period of 60 seconds was used to drive the gas sensor. The corresponding dynamic response signal (data sampling frequency of 2 Hz, i.e., modulation period of 60 seconds, each sample containing 120 data points) is shown below. Figure 2 As shown. From Figure 2 As can be seen from (a) and 2(b), in dynamic operating mode, the output signals of both gas sensors simultaneously contain information on toluene and methyl ethyl ketone (MEK). For example, as the MEK concentration gradually increases, the amplitude of the response signals from both gas sensors shows a significant increasing trend. Further comparison... Figure 2 (c) It is evident that for two sensors with different sensing performance, the response signals generated under the same dynamic operating mode exhibit significant differences, but both can reflect the mixed information of toluene and methyl ethyl ketone. Furthermore, the information acquired by different sensors also differs in feature representation and information content, such as differences in the shape of the response signal. Therefore, sensors with different sensing performance can generate differentiated information; integrating these sensors can further increase the data dimensionality, thereby improving gas detection performance. Therefore, utilizing the differentiated sensing characteristics exhibited by sensitive materials in different regions can effectively improve gas detection performance.

[0026] Step 2: Connect each test electrode of the composite MEMS gas sensor to the multi-channel data acquisition device, and apply a specific operating voltage to the virtual sensor array to put it into working state;

[0027] In practical implementation, the operating states include static and dynamic modes. During actual detection, the sensor's static or dynamic operating mode can be flexibly selected based on different task requirements to achieve the best match between data acquisition and task complexity. Specifically, the dynamic operating mode is suitable for complex task scenarios, as its response data possesses rich time-varying characteristics and multi-dimensional information, typically capturing more comprehensive information about the gas being tested, thereby improving the accuracy and reliability of complex analyses such as gas type identification and concentration change tracking. The static operating mode, on the other hand, is suitable for simple task scenarios, as its response data is singular, helping to simplify subsequent processing and facilitating rapid extraction of key features. It is suitable for basic detection tasks such as gas presence determination or steady-state concentration measurement. Therefore, appropriately selecting the operating mode can not only improve the system's adaptability to different tasks but also help optimize data processing efficiency and analytical performance.

[0028] Step 3: Collect the response signals output by each test electrode under different gas environments, and integrate the response signals from multiple test electrodes to construct a virtual sensor array dataset, specifically:

[0029] Step 3.1: Place the composite MEMS gas sensor in the gas environment to be tested and expose it to different types of gas environments in sequence. At the same time, continuously collect the current-time response signals generated by each test electrode under voltage excitation to form the IT response curve.

[0030] Step 3.2: For each gas exposure experiment, extract the response data corresponding to a complete voltage scan cycle from each IT response curve.

[0031] Specifically, if the composite MEMS gas sensor is used in dynamic operating mode, and the modulation voltage is set to a triangular wave with an amplitude of 6V to 2V and a period of 60s, and the data sampling frequency is 2Hz, then each test electrode will collect a total of 120 response data points within one modulation cycle. This interception process ensures that the data used for subsequent analysis can fully reflect the sensor's dynamic response characteristics to the gas within one modulation cycle.

[0032] Step 3.3: The response data of multiple test electrodes in the same gas exposure experiment and the same voltage scan cycle are fused using a time-domain-based data stitching strategy to form a sample.

[0033] The time-domain-based data splicing strategy specifically involves connecting the response signals of each test electrode within the same time interval according to the electrode sequence number to form a longer combined response sequence, which constitutes a sample.

[0034] Step 3.4: The samples obtained from multiple gas exposure experiments constitute a virtual sensor array dataset.

[0035] In practice, each of the two test electrodes collected 60 seconds of dynamic response data. Through time-domain concatenation, these two data points were combined into a single 120-second comprehensive response dataset. Extending this to a general case, if the number of test electrodes is n, the sensor's dynamic operating mode scan cycle is m seconds, and the data sampling frequency is L Hz, then by concatenating the response data from the n test electrodes in the time domain, each sample can be represented as an n×m×L dimensional vector, such as... Figure 3 As shown. Furthermore, if the experiment contains o independent gas exposure samples (i.e., o data samples), then the constructed "virtual sensor array" dataset can be formally expressed as a two-dimensional matrix of size o rows and n×m×L columns, as shown. Figure 4 As shown in the diagram, the rows of this matrix correspond to different samples, while the columns integrate the joint response information of multiple electrodes at multiple time points, providing rich feature inputs for subsequent gas identification and concentration analysis.

[0036] Step 4: Perform data drift detection and correction on the virtual sensor array dataset based on a sliding window to obtain the corrected virtual sensor array dataset, specifically as follows:

[0037] Step 4.1: Calculate the local center of the data using a sliding window:

[0038] Assume the sample whose drift direction is to be estimated is ,in , The number of samples and , The feature dimension of the sample is ; the total length of the sliding window is . Take the length of two adjacent child windows as .

[0039] With the first Sample Starting from this point, define the first and second child windows as shown in equations (1) and (2):

[0040] (1)

[0041] (2)

[0042] The local center of the first child window and the local center of the second child window are shown in equations (3) and (4) respectively:

[0043] (3)

[0044] (4)

[0045] in, and These are the local centers of the first and second child windows, respectively.

[0046] Step 4.2: Calculate the center difference vector between adjacent sub-windows. This difference vector represents the data drift within that time period; where, the first... The drift vector corresponding to each window is the difference between the center of the second sub-window and the center of the first sub-window:

[0047] (5)

[0048] in, This represents the center difference vector between adjacent sub-windows.

[0049] Place it on the timeline , that is, a point in the window, therefore, if represented by the time index l, then:

[0050] (6)

[0051] in, Defined as the drift vector of the l-th sample, and further used to estimate the drift intensity and its angle, as shown in equations (7) and (8):

[0052] (7)

[0053] in, This represents the drift intensity of the l-th sample. The first dimension represents the drift vector of the l-th sample; This represents the second dimension of the drift vector of the l-th sample.

[0054] (8)

[0055] in, This represents the drift angle of the l-th sample, which is the angle relative to the x-axis.

[0056] Step 4.3: After performing drift capture in steps 4.1 and 4.2, determine whether the drift intensity is greater than the threshold. If it is greater than the threshold, the sample has drifted. Then, proceed to step 4.4 to perform the drift correction process based on linear discriminant analysis.

[0057] Step 4.4: Calculate the inter-class scatter matrix for each sample before the drift intensity reaches the threshold. and intra-class scatter matrix :

[0058] (9)

[0059] (10)

[0060] in, The dataset consists of samples of the k-th class preceding the l-th sample in the virtual sensor array dataset; the sample types are classified according to the types of gases in the gas exposure experiment; M is the total number of sample types; This is the mean vector of the k-th class samples preceding the l-th sample in the virtual sensor array dataset; This represents the number of samples of class k preceding the l-th sample in the virtual sensor array dataset. This is the mean vector of the r-th class samples preceding the l-th sample in the virtual sensor array dataset; This is the mean vector of the s-th class samples preceding the l-th sample in the virtual sensor array dataset.

[0061] Step 4.5: Based on the calculated between-class scatter matrix and within-class scatter matrix, solve for the matrix. eigenvalues ​​and eigenvectors.

[0062] Step 4.6: Select the eigenvectors corresponding to the first p eigenvalues ​​in descending order of their magnitude. This allows for the construction of a projection matrix for drift data correction. .

[0063] Step 4.7: Correct the drifted sample according to the following formula:

[0064] (11)

[0065] in, To generate the l-th sample with drift, and This is the corrected l-th sample.

[0066] In specific implementation, under sinusoidal modulation voltage conditions with an amplitude of 6.5–2V and a period of 60s, the dynamic response data of air and 30–120 ppm xylene are used as examples to illustrate the drift capture and correction process. Figure 5 As shown. Figure 5 (a) Sine wave modulated waveform Figure 5 (b) Dynamic response curve of air. Figure 5 (c) Dynamic response curve of 30 ppm xylene. Figure 5 (d) Dynamic response curve of 60 ppm xylene, Figure 5 (e) Dynamic response curve of 90 ppm xylene, Figure 5 (f) Dynamic response curve of 120 ppm xylene. As shown in the figure, the response curves for different gases exhibit varying degrees of drift, for example, in… Figure 5 In (b), the dynamic response curve of air shows a clear trend of increasing amplitude over time; while... Figure 5 In (d), the dynamic response curve of 60 ppm xylene shows a relatively gentle amplitude decay characteristic.

[0067] Subsequently, to facilitate visualization analysis, Figure 5 The data in the image is reduced to a two-dimensional plane using principal component analysis, such as... Figure 6 As shown in (a). By Figure 6 As shown in (a), the data distribution exhibits a significant drift phenomenon; for example, the distribution of the two-dimensional scatter plot of air data gradually shifts to the lower right. Next, regarding... Figure 6 The data shown in (a) is further processed using a drift capture procedure. Specifically, following the algorithm flow of steps 4.1 and 4.2, the total length parameter of the sliding window is... Setting it to 4, the drift direction of the data is estimated, and the result is... Figure 6 (a) shows the arrow representation. The results indicate that this method can accurately estimate the data drift direction; that is, the direction indicated by the drift arrow accurately estimates the direction of data distribution drift. Based on this, the data drift intensity was estimated according to the algorithm flow in steps 4.1 and 4.2, and the results are as follows. Figure 6 As shown in (b), the figure demonstrates that this method can effectively characterize the intensity of overall data drift, for example... Figure 6 (a) shows a significant variation in data distribution between data points 35 and 36. Figure 6 (b) The corresponding intensity value is also relatively large. In practical applications, a reasonable threshold for the drift intensity can be set. That is, when the drift intensity is higher than the threshold, the correction process is initiated, and if it is lower than the threshold, the correction process is skipped and gas identification is performed directly. Finally, due to the large data drift, the data correction process shown in steps 4.3-4.7 is initiated, and the correction result is as follows: Figure 6 As shown in (c). With Figure 6 Compared with the uncorrected results in (a), the drift phenomenon of the corrected data is significantly reduced, and the clustering between different gases is clearer, thus effectively improving the gas identification performance.

[0068] Step 5: Input the corrected virtual sensor array dataset into the classification algorithm to achieve gas identification, specifically:

[0069] The corrected virtual sensor array dataset is used as input to the backpropagation neural network (BPNN) to perform joint identification of gas types and concentrations.

[0070] During the model training and validation phase, all sample data were randomly divided into training and test sets at a ratio of 70% and 30%, respectively: 70% of the data was used to train the BPNN model to learn the complex mapping relationship between features and gas categories and concentrations; the remaining 30% of the data was used to evaluate the model's generalization ability and test its accuracy in identifying unknown samples.

[0071] This invention employs a data dimensionality expansion method based on arrayed test electrodes, which, compared to traditional physical sensor arrays, offers advantages such as lower cost and easier integration, thereby increasing data dimensionality and enhancing the selectivity of gas identification. This method, by integrating multiple test electrodes onto a single sensor, improves the sensor's anti-drift capability to a certain extent from the signal detection end. Furthermore, regarding drift capture and correction, compared to existing gas identification technologies, current research largely focuses on using algorithms such as machine learning to compensate for drift data, while systematic research on drift direction and intensity remains insufficient. This invention not only achieves effective correction of data drift but also quantitatively estimates drift direction and intensity, thus enabling real-time and accurate determination of sensor drift status in practical applications.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the ideas of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A gas detection method based on data dimensionality expansion and data drift correction, characterized in that, include: Step 1: Integrate multiple test electrodes with different key parameters on a single MEMS gas sensor to fabricate a composite MEMS gas sensor; Step 2: Connect each test electrode of the composite MEMS gas sensor to the multi-channel data acquisition device, and apply a specific operating voltage to the virtual sensor array to put it into working state; Step 3: Collect the response signals output by each test electrode under different gas environments, and integrate the response signals of multiple test electrodes to construct a virtual sensor array dataset; Step 4: Perform data drift detection and correction on the virtual sensor array dataset based on a sliding window to obtain the corrected virtual sensor array dataset; Step 5: Input the corrected virtual sensor array dataset into the classification algorithm to achieve gas identification.

2. The gas detection method based on data dimensionality expansion and data drift correction according to claim 1, characterized in that, Step 1 specifically involves: Step 1.1: Determine the number of test electrodes based on the complexity of the gas to be tested and the required data dimensions, and determine the electrode material, electrode geometry, and spatial layout between the electrodes. Step 1.2: Integrate the test electrodes with different parameters mentioned above onto the same chip using standard MEMS fabrication technology; Step 1.3: Using drop-coating or inkjet printing, uniformly coat the surface of the integrated test electrode with a metal oxide semiconductor as a gas-sensitive material. Each test electrode and the gas-sensitive material constitute a composite sensing unit, and multiple composite sensing units constitute a virtual sensor array.

3. The gas detection method based on data dimensionality expansion and data drift correction according to claim 1, characterized in that, The working states include: static working mode and dynamic working mode.

4. The gas detection method based on data dimensionality expansion and data drift correction according to claim 1, characterized in that, Step 3 specifically involves: Step 3.1: Place the composite MEMS gas sensor in the gas environment to be tested and expose it to different types of gas environments in sequence. At the same time, continuously collect the current-time response signals generated by each test electrode under voltage excitation to form the IT response curve. Step 3.2: For each gas exposure experiment, extract the response data corresponding to a complete voltage scan cycle from each IT response curve; Step 3.3: The response data of multiple test electrodes in the same gas exposure experiment and the same voltage scan cycle are fused using a time-domain-based data stitching strategy to form a sample; Step 3.4: The samples obtained from multiple gas exposure experiments constitute a virtual sensor array dataset.

5. The gas detection method based on data dimensionality expansion and data drift correction according to claim 4, characterized in that, The time-domain based data stitching strategy is as follows: The response signals of each test electrode within the same time interval are connected end to end according to the electrode number to form a longer combined response sequence, which constitutes a sample.

6. The gas detection method based on data dimensionality expansion and data drift correction according to claim 4, characterized in that, Step 4 specifically involves: Step 4.1: Calculate the local center of the data using a sliding window; Assume the sample whose drift direction is to be estimated is ,in , The number of samples and , The feature dimension of the sample is ; the total length of the sliding window is . Take the length of two adjacent child windows as ; With the first Sample Starting from this point, define the first and second child windows as shown in equations (1) and (2): （1） （2） The local center of the first child window and the local center of the second child window are shown in equations (3) and (4) respectively: （3） （4） in, and These are the local centers of the first and second child windows, respectively. Step 4.2: Calculate the center difference vector between adjacent sub-windows. This difference vector represents the data drift within that time period; where, the first... The drift vector corresponding to each window is the difference between the center of the second sub-window and the center of the first sub-window: （5） in, Represents the center difference vector between adjacent sub-windows; Place it on the timeline , that is, a point in the window, therefore, if represented by the time index l, then: （6） in, Defined as the drift vector of the l-th sample, and further used to estimate the drift intensity and its angle, as shown in equations (7) and (8): （7） in, This represents the drift intensity of the l-th sample. The first dimension represents the drift vector of the l-th sample; The second dimension represents the drift vector of the l-th sample; （8） in, This represents the drift angle of the l-th sample, i.e., the angle relative to the x-axis; Step 4.3: After performing drift capture in steps 4.1 and 4.2, determine whether the drift intensity is greater than the threshold. If it is greater than the threshold, the sample has drifted. Then, proceed to step 4.4 to perform the drift correction process based on linear discriminant analysis. Step 4.4: Calculate the inter-class scatter matrix for each sample before the drift intensity reaches the threshold. and intra-class scatter matrix ; （9） （10） in, The dataset consists of samples of the k-th class preceding the l-th sample in the virtual sensor array dataset; the sample types are classified according to the types of gases in the gas exposure experiment; M is the total number of sample types; This is the mean vector of the k-th class samples preceding the l-th sample in the virtual sensor array dataset; This represents the number of samples of class k preceding the l-th sample in the virtual sensor array dataset. This is the mean vector of the r-th class samples preceding the l-th sample in the virtual sensor array dataset; This is the mean vector of the s-th class samples preceding the l-th sample in the virtual sensor array dataset; Step 4.5: Based on the calculated between-class scatter matrix and within-class scatter matrix, solve for the matrix. eigenvalues ​​and eigenvectors; Step 4.6: Select the eigenvectors corresponding to the first p eigenvalues ​​in descending order of their magnitude. This allows for the construction of a projection matrix for drift data correction. ; Step 4.7: Correct the drifted sample according to the following formula: （11） in, To generate the l-th sample with drift, and This is the corrected l-th sample.

7. The gas detection method based on data dimensionality expansion and data drift correction according to claim 1, characterized in that, Step 5 specifically involves: The corrected virtual sensor array dataset is used as input to the backpropagation neural network for joint identification of gas types and concentrations.

Images