Method for identifying VOCs in mixed environment based on machine learning algorithm

Abstract

The invention discloses a method for identifying VOCs (Volatile Organic Compounds) in a mixed environment based on a machine learning algorithm, which comprises the following steps of: firstly, explaining the principles and mechanisms of data preprocessing and deep learning (mainly CNN (Convolutional Neural Network)); then, a VOCs identification and prediction method based on a mixed background library is provided, and a series of experiments are designed to verify the effectiveness of the method. Finally, through analysis and discussion of experimental results, the specific single-phase VOCs gas is successfully identified from the mixed environment. The research result of the invention is helpful for improving the accurate identification and prediction capability of VOCs in the fields of environment monitoring, indoor air quality detection, personal health protection and the like, and provides theoretical and technical support for the detection of mixed gases in practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of VOCs identification, and more specifically to a method for identifying VOCs in a mixed environment based on a machine learning algorithm. Background Art

[0002] Volatile organic compounds (VOCs) are a class of volatile, low-boiling organic compounds that are widely present in industrial production, transportation, building decoration, household goods and other fields, and pose potential risks to human health and environmental quality. Therefore, accurately identifying and predicting the composition and concentration of VOCs in a mixed gas environment is of great significance for environmental monitoring, air treatment and human health.

[0003] However, due to the wide variety and complex structure of VOCs, traditional gas analysis methods often require expensive and cumbersome instruments and complicated sample preparation processes, which limits their popularity and promotion in practical applications. With the rapid development of artificial intelligence and machine learning, it is possible to use these technologies to develop intelligent VOCs identification and prediction methods.

[0004] Accurate identification of VOCs in mixed environments faces many challenges. These challenges include, but are not limited to: the complexity of VOCs composition, the impact of changing environmental conditions on VOCs behavior, and the interaction between VOCs. In addition, monitoring low-concentration VOCs and distinguishing them from other gas components are also technical obstacles that must be overcome during the identification process.

[0005] To accurately predict the presence and concentration levels of VOCs, researchers have developed a variety of models, including statistical models, artificial intelligence (AI) models, and machine learning and deep learning techniques. These models can help predict the likely behavior of VOCs under specific environmental conditions, thereby optimizing monitoring strategies and air quality management. Machine learning has a wide range of applications in identification and prediction tasks. It trains and learns models using existing data, and then uses these models to identify and predict new data. Machine learning and pattern recognition play an important role in data analysis and information extraction. By applying different methods and techniques, it can help understand and utilize various types of data and achieve accurate prediction and classification results. However, continued research and exploration of new algorithms and methods is still necessary to overcome the challenges and limitations that may arise in practical applications. Summary of the invention

[0006] The present invention aims to study how to qualitatively and quantitatively analyze different single-phase VOCs from a mixed environment. First, the principles and mechanisms of data preprocessing and deep learning (mainly CNN) are explained. Then, a VOCs identification and prediction method based on a mixed background library is proposed, and a series of experiments are designed to verify the effectiveness of the method. Finally, through the analysis and discussion of the experimental results, specific single-phase VOCs gases are successfully identified from a mixed environment. The results of the present invention's research will help improve the ability to accurately identify and predict VOCs in the fields of environmental monitoring, indoor air quality testing, personal health protection, etc., and provide theoretical and technical support for the detection of mixed gases in practical applications.

[0007] The technical solution of the present invention is as follows:

[0008] A method for identifying VOCs in a mixed environment based on a machine learning algorithm, comprising:

[0009] (1) Fabrication of Fe2O3@SnO2 gas sensor

[0010] Preferably, the gas sensor is a MOS gas sensor.

[0011] Preferably, in the present invention, a ceramic sheet is selected as the main structure of the sensor, the sensing element is mainly supported by a six-pin base, the gold electrode on the front of the heating ceramic plate is used as a test electrode, and the RuO2 resistor on the back is used as a heating electrode to provide the required operating temperature for the target gas test.

[0012] Preferably, the manufactured sensor needs to be subjected to a stable aging treatment, generally by being properly stored under dry conditions for 12-24 hours.

[0013] (2) Parameter control of temperature modulation technology

[0014] Including: adjustment of frequency, duty cycle and voltage parameters.

[0015] Preferably, a Keithley 6487 picoammeter is used as voltage source and current reader, applying a membrane voltage of 5 V to the transducer and recording two values ​​per second.

[0016] Preferably, a RIGOL DP832A is used to apply a square wave pulse heating voltage to the resistance wire, and the parameters are as follows: low voltage 0V, high voltage 6.5V and 50% duty cycle; frequency adjustment range: 0.01Hz-0.05Hz.

[0017] (3) Dynamic gas sensing performance test

[0018] Preferably, the gas sensor is a Fe2O3@SnO2 gas sensor.

[0019] Preferably, the sensing measurements are performed in a closed glass container (500 mL) equipped with appropriate gas flow inlet and outlet. Temperature modulated sensing of the detected gas is performed in dry air, and each measurement is repeated three or more times.

[0020] Preferably, the response is defined as S = R at a characteristic temperature point a / R g , where R a and R g They are the resistance in dry air and in the presence of the detected gas at the corresponding characteristic temperature points. Finally, the characteristic response curve of the target gas under temperature modulation is obtained by computer data acquisition. It should be emphasized that all gas sensing measurements are performed in dry air.

[0021] (4) Identification of single-phase gas in a mixed environment:

[0022] Preferably, LDA (Linear Discriminant Analysis) is used; and K-means algorithm is combined at the same time.

[0023] Preferably, a convolutional neural network (CNN) is employed.

[0024] Preferably, support vector machine, neural network, decision tree, linear discriminant, naive Bayes, K nearest neighbor method are used. Preferably, deep learning model (CNN, SVM, PCA-LR) is used.

[0025] Technical effects and advantages of the present invention:

[0026] The present invention mainly takes the 7 VOCs specifically identified by the Fe2O3@SnO2 gas sensor as the identification objects, and on this basis establishes the corresponding mixed gas background library. Through the preprocessing of test data; training; model establishment and pattern recognition and prediction, the characteristics between different mixtures are analyzed in detail, and the possible reasons for their existence are explained. At the same time, a variety of deep learning models are used to perform pattern recognition and concentration prediction on seven single-phase gases. The results show that the recognition accuracy of n-butanol using the CNN model reaches 99.8%, and the accuracy of the remaining gases is above 98%. The concentration prediction results show that the predicted concentration is highly consistent with the actual concentration, which provides strong theoretical support for the subsequent detection of unknown concentrations of actual samples. . BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the technical solution of the invention.

[0028] Figure 2 Schematic diagram of machine learning (CNN) data preprocessing training.

[0029] Figure 3 Schematic diagram of different model selections for machine learning.

[0030] Figure 4 Schematic diagram of mixed gas pattern recognition and concentration prediction mechanism.

[0031] Figure 5 Schematic diagram of establishing a mixed gas test library using Fe2O3@SnO2 sensor.

[0032] Figure 6 Schematic diagram of establishing mixed gas test library 1 using Fe2O3@SnO2 sensor.

[0033] Figure 7 Schematic diagram of establishing mixed gas test library 2 using Fe2O3@SnO2 sensor.

[0034] Figure 8 Schematic diagram of establishing mixed gas test library 3 using Fe2O3@SnO2 sensor.

[0035] Fig. 9 Schematic diagram of establishing mixed gas test library 4 using Fe2O3@SnO2 sensor.

[0036] Fig.10 Schematic diagram of establishing mixed gas test library 5 using Fe2O3@SnO2 sensor.

[0037] Fig.11 Target gas recognition results based on different machine learning models: (a) SVM; (b) Decision Tree; (c) Linear Discrimination; (d) Naive Bayes; (e) K-nearest Neighbor; (f) CNN.

[0038] Fig.12 Target gas recognition results based on different machine learning models Target gas concentration prediction results based on different deep learning models, (a) n-butanol; (b) n-propanol; (c) ethanol; (d) triethylamine; (e) formaldehyde; (f) acetone; (g) ether. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Embodiment 1

[0041] like Figure 1 The figure shows the technical scheme of the invention. First, the Fe2O3@SnO2 gas sensor can be prepared, and the temperature modulation technology is used to optimize the gas sensing performance. The LDA technology is used to identify the single-phase gases of 7 VOCs (n-propanol, n-butanol, ethanol, triethylamine, formaldehyde, acetone, and ether). At the same time, multiple deep learning models are used to perform pattern recognition and concentration prediction on the seven single-phase gases. Finally, the CNN model is used to achieve an accuracy rate of 99.8% for n-butanol, and the accuracy rates of the remaining gases are all above 98%.

[0042] Embodiment 2

[0043] The technical solution of the present invention mainly includes the following parts:

[0044] (1) Fabrication of Fe2O3@SnO2 gas sensor

[0045] In the present invention, porous SnO2 nanofibers embedded with Fe2O3 hollow spheres (Fe2O3@SnO2) are prepared as gas sensitive sensing materials.

[0046] (2) Parameter control of temperature modulation technology

[0047] The present invention uses a Keithley 6487 picoammeter as a voltage source and current reader, applies a membrane voltage of 5V to the transducer, and records two values ​​per second. A square wave pulse heating voltage is applied to the resistance wire using a RIGOL DP832A, and the parameters are as follows: low voltage 0V, high voltage 6.5V and 50% duty cycle; frequency adjustment range: 0.01Hz-0.05Hz. Different frequencies show different temperature changes at this voltage. The sensing measurement is carried out in a closed glass container (500mL) equipped with appropriate airflow inlet and outlet. The detected gas is subjected to temperature modulation sensing behavior in dry air, and each measurement is repeated more than three times.

[0048] (3) Dynamic gas sensing performance test

[0049] This paper will illustrate the differences in temperature changes under different frequency regulation: the real-time temperature change curve of the Fe2O3@SnO2 sensor tested by the gas sensor assembly manufactured in three cycles under square wave pulse heating voltage with different frequencies and 50% duty cycle shows periodic changes. More specifically, there are differences in the temperature change range of each frequency, and the temperature change ranges of different frequencies are 29.8-445.1℃, 38.7-445.1℃, 60.5-436.1℃, 95.3-428.6℃ and 112.1-412.2℃ respectively.

[0050] (4) Single-phase gas identification in mixed environment

[0051] This paper proposes a VOCs identification and prediction method based on a mixed background library, and designs a series of experiments to verify the effectiveness of this method. The specific steps are as follows:

[0052] 4) Data preprocessing training

[0053] The present invention preprocesses and classifies the data in all mixed gas libraries tested by the Fe2O3@SnO2 gas sensor, 30% of the data is used as a label set, and the rest of the data is used as a training set. In addition, interference data is introduced to provide error analysis for the establishment of the model. Secondly, the convolutional neural network (CNN) is used to train and optimize the data. The input data is normalized and preprocessed, and different features are extracted using a 6-core convolution layer. Each convolution layer consists of several convolution units, and the parameters of each convolution unit are optimized by the back propagation algorithm. Then, the activation function is applied to the rectified linear unit (ReLU), and the recognition architecture is 128, 64, 32, 16 and 8 layers respectively. As a fully connected layer, it merges all local features into global features to calculate the final score for each category. Finally, the two important nodes as the output layer are species identification and concentration prediction. Here, the classification nodes are marked with 01-07 as the symbols of the final output for easy final identification. As a linear regression analysis, concentration prediction reflects the quality of its prediction by comparing the unknown concentration with the known concentration. The detailed process is shown in Figure 2 .

[0054] 5) Model selection of data

[0055] It is necessary to further explain that in the calculation process of deep learning, various models are used for data classification analysis and optimization (support vector machine, neural network, decision tree, linear discriminant, naive Bayes, K nearest neighbor), and the specific process is reflected in Figure 3 The choice of deep learning model plays an extremely important role in solving practical problems. Different models are suitable for different data characteristics and task types. Therefore, choosing a suitable model can significantly improve the performance of the model and reduce resource consumption. At the same time, the choice of model will also affect the training process of the model and the interpretability of the results.

[0056] 6) Data pattern recognition and prediction

[0057] The present invention uses all the mixed gas data tested by the Fe2O3@SnO2 gas sensor under temperature control (frequency is 0.02Hz, duty cycle is 50%) as the input end, and the single-phase gas is represented by (a, b, c, d, e, f, g) respectively. All the data are pre-processed and trained for deep learning. As object recognition, data features are extracted from the mixed environment gas library and compared with the features of the single phase. Finally, the features of the single phase are found from the mixed gas phase, and the recognition accuracy is compared by training. Next, three different deep learning models (CNN, SVM, PCA-LR) are used to predict the unknown concentrations that may exist in the single phase in different mixed phases. The specific process is shown in Figure 4 .

[0058] 7) Establishment of mixed gas library

[0059] The present invention successfully achieved the purpose of identifying seven VOCs using Fe2O3@SnO2 gas sensors, including n-propanol, n-butanol, ethanol, triethylamine, formaldehyde, acetone, and ether. On this basis, a mixed gas library was established, and the characteristic curves of different mixture phases were tested and used as the data of the gas library. Here, a total of 21 different mixture phase response characteristic curves were cross-tested, which can be specifically obtained from Figure 5 In the test, in order to better obtain the characteristics of the single-phase gas in the mixed phase, the concentration of the single-phase gas is fixed before the test, and new phase gases of different concentrations are continuously added to observe and compare the changes between the two and analyze the reasons for the changes.

[0060] 8) Real-time characteristic response curves of different mixtures in mixed environment testing

[0061] Example 1: Real-time characteristic response curves of different mixtures tested by the prepared Fe2O3@SnO2 gas sensor in a mixed environment Figure 6 As shown. Test conditions: square wave pulse heating voltage with a frequency of 0.02Hz and a duty cycle of 50%. (Take n-propanol as the research object, and fix its single-phase concentration to 10ppm) It should be emphasized here that there is a certain degree of similarity between different mixed gases, and it can be found that the inherent characteristics of the single phase are an important factor in determining the degree of similarity after mixing.

[0062] Example 2: Real-time characteristic response curves of different mixtures tested by the prepared Fe2O3@SnO2 gas sensor in a mixed environment Figure 7 Test conditions: square wave pulse heating voltage with a frequency of 0.02Hz and a duty cycle of 50%. (Take n-butanol as the research object, and fix its single-phase concentration to 10ppm)

[0063] Example 3: Real-time characteristic response curves of different mixtures tested by the prepared Fe2O3@SnO2 gas sensor in a mixed environment Figure 8 Test conditions: square wave pulse heating voltage with a frequency of 0.02Hz and a duty cycle of 50%. (Ethanol was used as the research object, and its single-phase concentration was fixed at 10ppm)

[0064] Example 4: Real-time characteristic response curves of different mixtures tested by the prepared Fe2O3@SnO2 gas sensor in a mixed environment Fig. 9 Test conditions: square wave pulse heating voltage with a frequency of 0.02Hz and a duty cycle of 50%. (Triethylamine alcohol was used as the research object, and its single-phase concentration was fixed at 10ppm)

[0065] Example 5: Real-time characteristic response curves of different mixtures tested by the prepared Fe2O3@SnO2 gas sensor in a mixed environment Fig.10 Test conditions: square wave pulse heating voltage with a frequency of 0.02Hz and a duty cycle of 50%. (Formaldehyde, acetone, and ether were used as research objects, and their single-phase concentration was fixed at 10ppm)

[0066] 9) Mixture identification results based on machine learning

[0067] The output results of the final confusion matrix model of the present invention show that among the six models, the recognition and classification effect of the convolutional neural network (CNN) on n-butanol is 99.8%, such as Fig.11As shown. Secondly, the accurate recognition rates of n-propanol, ethanol, triethylamine, formaldehyde, acetone, and ether are 99.4%, 99.5%, 99.10%, 98.90%, 98.50%, and 98.70%. At the same time, due to its powerful learning and data processing capabilities, CNN has good classification performance when used as the output classification of seven target objects. Further research on the results of support vector machine (SVM) for different gases from the mixed gas library shows that its accurate recognition rate for n-butanol is 97.5%, which indirectly reflects that the modeling of n-butanol in the seven gases has better recognition and generalization capabilities. Similarly, the accuracy rates for other target gases are 96.10%, 95.90%, 95.70%, 95.30%, 94.50%, and 93.50%, respectively. However, for the output results of the decision tree, naive Bayes, and K nearest neighbor models, it can be clearly found that the accurate recognition rate for formaldehyde is not very high, and the results of the confusion matrix output are consistent. This may be related to the number of iterations of learning and training in the process of model data processing and optimization, but it depends on the target gas itself to a greater extent. The successful and accurate identification of different target single-phase gases (n-butanol, n-propanol, ethanol, triethylamine, formaldehyde, acetone, ether) from mixed environments not only provides powerful assistance for the subsequent detection and identification of environmental pollutants, but also provides a certain model basis for the qualitative identification and prediction of VOCs.

[0068] 10) VOCs gas concentration prediction results based on machine learning

[0069] The present invention tests different concentrations of different single-phase target gases, and also tests the concentrations after different mixed phases, and on this basis predicts and compares the actual concentration and the unknown concentration. Subsequently, three specific models (CNN, SVM, PCA-LR) are used to perform linear regression analysis of the concentration. Fig.12 The output results of the concentrations of different target gases (n-butanol, n-propanol, ethanol, triethylamine, formaldehyde, acetone, and ether) in the corresponding mixed gases are explained in detail, indicating that their actual concentrations are consistent with the predicted concentrations, and proving the reliability of the three models. In addition, it can be observed that the closer the point is to the straight line, the higher its accuracy. This is not only a predicted display of the concentration, but also a trend of the concentration change of a single-phase gas in different mixed phases. In fact, the true concentration and the actual concentration are different to a certain extent, such as Fig.12In (b), the predicted concentrations of n-propanol using support vector machine (SVM) and linear regression analysis (PCR-LR) are 66ppm and 67ppm respectively, while the actual concentration is 60ppm. This difference may be caused by the influence of parameters during model parameterization training. What needs to be emphasized here is that the parameter values ​​of the model will directly affect its concentration prediction ability. The selection and generalization of different models provide strong support for the quantitative analysis of different gases in the actual mixed environment.

[0070] The present invention uses the CNN model to achieve an identification accuracy of 99.8% for n-butanol, and the accuracy of the remaining gases is above 98%. The concentration prediction results show that the predicted concentration is highly consistent with the actual concentration, providing strong theoretical support for the subsequent detection of unknown concentrations of actual samples.

[0071] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying VOCs in a mixed environment based on a machine learning algorithm, characterized in that: include: (1) Fabrication of Fe2O3@SnO2 gas sensor Fe2O3 hollow spheres embedded in porous SnO2 nanofibers (Fe2O3@SnO2) were prepared as gas sensing materials; The ceramic sheet is selected as the main structure of the sensor. The sensing element is mainly supported by a six-pin base. The gold electrode on the front of the heating ceramic plate is used as a test electrode, and the RuO2 resistor on the back is used as a heating electrode to provide the required working temperature for the target gas test. (2) Parameter control of temperature modulation technology Including: frequency, duty cycle, voltage parameter adjustment; A membrane voltage of 5 V was applied to the transducer using a Keithley 6487 picoammeter as a voltage source and current reader, and two values ​​were recorded per second; Use RIGOL DP832A to apply square wave pulse heating voltage to the resistance wire; (3) Dynamic gas sensing performance test The sensing measurements were performed in a closed glass container (500 mL) equipped with appropriate gas flow inlet and outlet; the temperature-modulated sensing behavior of the detected gas was performed in dry air, and each measurement was repeated more than three times; The response is defined as S = R at a characteristic temperature point a / R g , where R a and R g They are the resistance in dry air and in the presence of the detected gas at the corresponding characteristic temperature points; finally, the characteristic response curve of the target gas under temperature modulation is obtained by collecting data through a computer; it should be emphasized that all gas sensing measurements are carried out in dry air; (4) Identification and prediction of single-phase gas in mixed environment, the method is as follows: 1) Data preprocessing training The data in all the mixed gas libraries tested by the Fe2O3@SnO2 gas sensor were preprocessed and classified, 30% of the data were used as the label set, and the rest of the data were used as the training set; in addition, interference data were introduced to provide error analysis for the establishment of the model; secondly, the convolutional neural network (CNN) was used to train and optimize the data, and the input data was normalized and preprocessed, and different features were extracted using a 6-core convolution layer; each convolution layer consists of several convolution units, and the parameters of each convolution unit are optimized by the back propagation algorithm; then, the activation function is applied to the rectified linear unit (ReLU), and the recognition architecture is 128, 64, 32, 16 and 8 layers respectively; as a fully connected layer, it merges all local features into global features to calculate the final score of each category; finally, the two important nodes as the output layer are species identification and concentration prediction; here, the classification nodes are marked with 01-07 as the symbols of the final output for easy final identification; as a linear regression analysis, the concentration prediction reflects the quality of its prediction by comparing the unknown concentration with the known concentration; 2) Model selection of data In the calculation process of deep learning, various models are used for data classification analysis and optimization, including support vector machines, neural networks, decision trees, linear discriminants, naive Bayes, and K nearest neighbors; 3) Data pattern recognition and prediction All mixed gas data tested by the Fe2O3@SnO2 gas sensor under temperature control, frequency of 0.02Hz and duty cycle of 50% are used as input. Single-phase gases are represented by (a, b, c, d, e, f, g) respectively. All data are preprocessed and trained for deep learning. As object recognition, data features are extracted from the mixed environmental gas library and compared with the features of the single phase. Finally, the features of the single phase are found from the mixed gas phase, and the recognition accuracy is trained and compared. Next, three different deep learning models (CNN, SVM, PCA-LR) are used to predict the unknown concentrations of the single phase in different mixed phases.

2. The method for identifying VOCs in a mixed environment based on a machine learning algorithm according to claim 1, characterized in that: The gas sensor is a MOS gas sensor.

3. The method for identifying VOCs in a mixed environment based on a machine learning algorithm according to claim 1, characterized in that: A square wave pulse heating voltage was applied to the resistance wire using RIGOL DP832A, and the parameters were as follows: low voltage 0 V, high voltage 6.5 V, and 50% duty cycle; frequency adjustment range: 0.01 Hz-0.05 Hz.

4. The method for identifying VOCs in a mixed environment based on a machine learning algorithm according to claim 1, characterized in that: For the sensors that have been manufactured, they need to be subjected to stable aging treatment, generally by being properly stored in dry conditions for 12-24 hours.

5. The method for identifying VOCs in a mixed environment based on a machine learning algorithm according to claim 1, characterized in that: It also includes the establishment of a mixed gas library. The Fe2O3@SnO2 gas sensor was used to successfully identify seven VOCs, including n-propanol, n-butanol, ethanol, triethylamine, formaldehyde, acetone, and ether. On this basis, a mixed gas library was established, and the characteristic curves of different mixture phases were tested and used as the data of the gas library. Here, a total of 21 different mixture phase response characteristic curves were cross-tested. In the test, in order to better obtain the characteristics of the single-phase gas in the mixed phase, the concentration of the single-phase gas was fixed before the test, and new phase gases of different concentrations were continuously added to observe and compare the changes between the two and analyze the causes of the changes.

6. The method for identifying VOCs in a mixed environment based on a machine learning algorithm according to claim 1, characterized in that: The gas sensor is a Fe2O3@SnO2 gas sensor.

Images