Infrared telemetry spectrum simulation method for chemical toxic agent

Abstract

The invention discloses a chemical toxic agent infrared telemetry spectrum simulation method, belongs to the technical field of chemical toxic agent detection, and solves the problems of unbalanced training data distribution, insufficient data volume, high data similarity and the like in the process of identifying a chemical toxic agent by using a deep neural network. The method comprises the following steps: firstly, generating a pure spectrum, performing interpolation, normalization and data cleaning on an actually measured spectrum to obtain parameters such as the number, the peak position, the peak height and the peak width of characteristic peaks of the toxic agent infrared spectrum, and generating a pure spectrogram through a Gaussian function and by adjusting a parameter range; then, on the basis of the pure spectrum, polynomial baselines with different shapes are superposed; and finally, random noise is added, Gaussian random white noise is added to the spectral data of the superposed base line, and the generated simulation spectrum is consistent with the actually measured spectrum. According to the method, the economic and time cost in the actual toxic agent spectrum collection process is effectively saved, the requirement of the neural network training process for the data volume can be met by collecting less data, and the poisoning injury risk of experimenters in the toxic agent spectrum collection process can be reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of chemical agent detection, and in particular relates to a chemical agent infrared remote sensing spectrum simulation method. Background Art

[0002] At present, almost all chemical detection methods are being used in the field of chemical agent detection. Passive infrared remote sensing spectroscopy technology can monitor chemical agent clouds and volatile gases of atmospheric organic pollution within a range of thousands of meters in real time, online, and non-contact within a few seconds, and realize qualitative identification and quantitative analysis of substances, which is of great significance for locating pollution sources and avoiding military operations. However, due to the long detection distance and serious background interference in the atmosphere, the infrared measured spectrum often contains high-frequency noise and the spectrum signal-to-noise ratio is low, and infrared spectrum recognition has high requirements for algorithm models.

[0003] Deep neural network models have high recognition accuracy, but the training process often requires massive data support to avoid overfitting. Therefore, increasing the data scale is a key means to improve network performance. However, data acquisition and manual data labeling not only require expensive economic costs, but are also time-consuming and labor-intensive. Affected by objective factors, it is difficult to obtain sufficient training data in many research fields. Spectral simulation technology, as a data expansion method for artificially generating new data, can use existing small sample data to obtain more diverse data when the data volume is insufficient, thereby improving data quality and the generalization ability of network models. Summary of the invention

[0004] 1. Technical issues to be solved

[0005] The present invention proposes an infrared remote sensing spectrum simulation method for chemical agents and toxic and hazardous substances to solve the problems of uneven distribution of training data, insufficient data volume and high data similarity in the process of identifying chemical agents using deep neural networks.

[0006] (II) Technical solution

[0007] In order to solve the above technical problems, the present invention proposes a simulation method based on infrared telemetry spectrum of chemical agents and toxic and hazardous substances. The simulation method comprises the following steps:

[0008] S1. Select polar chemical agents and collect their infrared spectra

[0009] Before collecting infrared spectrum information, adjust the angle and position of the blackbody radiator, gas absorption cell, and infrared spectrometer to ensure that the light flux of the optical path where the reagent is located is the largest. Before sampling, exhaust the residual gas in the gas absorption cell through the vacuum system and collect the background spectrum. Select common nerve agents (sarin, soman) and corrosive agents (mustard gas, Lewisite), fix the pipeline temperature at 40°C, and obtain different temperature differences between the gas and the pipeline by adjusting the surface source blackbody radiation temperature. Use a micro-injector to inject samples into the gas absorption cell. The injection dose is determined according to the volatility of the agent and the toxic gas. For gases with good volatility, a semi-emptying method is used when emptying to obtain low and medium signal-to-noise ratio data. For gases with poor volatility, try different temperature differences and doses to obtain low, medium, and high signal-to-noise ratio data.

[0010] The infrared remote sensing spectrometer developed by the company collects the difference spectrum information of each agent before and after mixing with air, after deducting the gas pool background. The sampling interval of the infrared remote sensing test system is 1 second, and the difference spectrum of each agent before and after mixing with air is collected for 15 minutes, totaling 1,000 to 2,000.

[0011] S2. Generate simulated spectrum

[0012] S2-1. Clean spectrum generation

[0013] The difference spectrum information collected in S1 was preprocessed by spline interpolation, normalization and data cleaning. 40 measured spectrum data were randomly selected as the simulation basis. The measured parameters such as peak height, peak position, peak width and peak number of characteristic peaks were obtained using Python built-in functions. A single main peak was generated using the Gaussian function, and multiple main peaks were combined according to the peak height ratio to obtain a pure spectrum containing multiple characteristic peaks. The Gaussian function formula is as follows:

[0014]

[0015] x is the band range after spline interpolation. The three parameters a (peak height), b (peak position), and c (peak width) of the Gaussian function are set to random numbers fluctuating around the measured parameters. 2000 different pure spectra are generated and the function values ​​are normalized.

[0016] S2-2. Adding polynomial baseline

[0017] Based on the pure spectrum generated in step S2-1, a polynomial baseline is added through the polynomial baseline formula to obtain a simulated spectrum with superimposed baselines of different shapes. A certain range is set for the polynomial coefficients to generate polynomial curves of different shapes. The polynomial baseline formula is as follows:

[0018]

[0019] Wherein, m, n, q are polynomial coefficients, and the horizontal coordinate x has the same meaning as in step S2-1. m, n, q are all set to random numbers between -1 and 1 to generate a random baseline. The horizontal coordinate x is divided by 1000 to control the consistency of the horizontal and vertical coordinate measurement scales.

[0020] S2-3. Adding random noise

[0021] Based on step S2-2, Gaussian white noise is added through the awgn function to obtain 2000 simulated spectra of each agent. The function expression is as follows:

[0022] ps=awgn(xin,snr,′measured′)

[0023] Among them, awgn is a function that can add Gaussian white noise with a specific signal-to-noise ratio to the signal, xin and ps are the spectral signals before and after adding Gaussian white noise, snr is the signal-to-noise ratio, which is set to a random number in the range of 30 to 40, and 'measured' is the power of the added noise calculated based on the actual power of the input signal and the signal-to-noise ratio.

[0024] S3. Construct infrared simulation spectrum recognition model

[0025] The simulated spectra of different toxic agents generated in step S2 were integrated and annotated with real category labels to form a sample set, and the sample set was randomly divided into a training set and a validation set at a ratio of 7:3. The infrared simulated spectrum recognition model was constructed on the Python software platform using a convolutional neural network, the input two-dimensional spectral data was converted into a four-dimensional vector, the convolutional neural network structure and training parameters were adjusted, and the validation set was internally verified using a classification function.

[0026] S4. External verification of infrared simulation spectrum recognition model

[0027] 300 spectra were randomly selected from the 1000 to 2000 infrared spectra of each agent in step S2-1 and processed with polynomial fitting baseline correction as the test set for external verification of the infrared simulation spectrum recognition model in S3. The true labels were annotated, and the spectrum dimensions were converted into four dimensions as model input. The predicted labels were output through the classification function in Python software, and the probability score threshold was set to 0.5. If the probability score was below 0.5, it would not be recognized. The true labels were compared with the predicted labels.

[0028] Furthermore, the nerve agents in step S1 include sarin and soman, the corrosive agents include mustard gas and Lewisite, and the types of simulated agents can also be expanded to general toxic and harmful gases, such as ammonia, ethanol, styrene and other common chemical reagents. The self-developed infrared remote sensing spectrometer measurement and spline interpolation band range is the mid-infrared band: 1347cm -1 ~681cm -1 .

[0029] Furthermore, in steps S3 and S4, the internal and external verifications are performed by using the classification function of Python software to calculate and output the predicted labels. When the predicted labels are consistent with the true labels, it indicates that the simulation spectrum recognition model prediction is correct.

[0030] Furthermore, in step S4, when the number of correctly predicted samples is close to the total number of tested samples, that is, the prediction accuracy is close to 1, it indicates that the classification effect of the simulated spectrum recognition model is good.

[0031] (III) Beneficial effects

[0032] The present invention proposes a simulation method based on infrared telemetry spectrum of chemical agents, which includes three processes. The first is to generate a pure spectrum. After the measured spectrum is interpolated, normalized, and data cleaned, the parameters such as the number of characteristic peaks, peak position, peak height and peak width of the infrared spectrum of the agent are obtained. Through the Gaussian function and the adjustment of the parameter range, 2000 different pure spectra are generated; then a polynomial baseline is added. On the basis of the pure spectrum, polynomial baselines of various shapes are superimposed; finally, random noise is added. Gaussian random white noise is added to the spectrum data of the superimposed baseline, and the generated simulation spectrum is consistent with the measured spectrum.

[0033] By generating simulated spectra of chemical agents, the economic and time costs of the actual agent spectrum acquisition process can be effectively saved. Collecting less data can meet the data volume requirements of the neural network training process, and the risk of poisoning and injury to experimenters during the agent spectrum acquisition process can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The present invention is a flow chart of the infrared spectrum simulation method of chemical agents;

[0035] Figure 2 Schematic diagram of a portion of the polynomial baseline generated for the present invention;

[0036] Figure 3 This is a comparison diagram of the simulated spectrum and the measured spectrum of the mustard gas agent in the present invention.

[0037] Figure 2 ,3, the horizontal axis is the wave number, the unit is cm -1 , the vertical axis is the infrared absorption intensity, the unit is 1. DETAILED DESCRIPTION

[0038] In order to make the purpose, content and advantages of the present invention clearer, the specific implementation methods of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments.

[0039] This embodiment provides a chemical agent infrared remote sensing spectrum simulation method, such as Figure 1 As shown, the spectrum simulation method mainly includes the following steps:

[0040] S1. Select polar chemical agents and collect their infrared spectra

[0041] Before collecting infrared spectrum information, adjust the angle and position of the blackbody radiator, gas absorption cell, and infrared spectrometer to ensure that the light flux of the optical path where the reagent is located is the largest. Before sampling, exhaust the residual gas in the gas absorption cell through the vacuum system and collect the background spectrum. Select common nerve agents (sarin, soman) and corrosive agents (mustard gas, Lewisite), fix the pipeline temperature at 40°C, and obtain different temperature differences between the gas and the pipeline by adjusting the surface source blackbody radiation temperature. Use a micro-injector to inject samples into the gas absorption cell. The injection dose is determined according to the volatility of the agent and the toxic gas. For gases with good volatility, a semi-emptying method is used when emptying to obtain low and medium signal-to-noise ratio data. For gases with poor volatility, try different temperature differences and doses to obtain low, medium, and high signal-to-noise ratio data.

[0042] The infrared remote sensing spectrometer developed by the company collects the difference spectrum information of each agent before and after mixing with air, after deducting the gas pool background. The sampling interval of the infrared remote sensing test system is 1 second, and the difference spectrum of each agent before and after mixing with air is collected for 15 minutes, totaling 1,000 to 2,000.

[0043] S2. Generate simulated spectrum

[0044] S2-1. Clean spectrum generation

[0045] The difference spectrum information collected in S1 was preprocessed by spline interpolation, normalization and data cleaning. 40 measured spectrum data were randomly selected as the simulation basis. The measured parameters such as peak height, peak position, peak width and peak number of characteristic peaks were obtained using Python built-in functions. A single main peak was generated using the Gaussian function, and multiple main peaks were combined according to the peak height ratio to obtain a pure spectrum containing multiple characteristic peaks. The Gaussian function formula is as follows:

[0046]

[0047] x is the band range after spline interpolation. The three parameters a (peak height), b (peak position), and c (peak width) of the Gaussian function are set to random numbers fluctuating around the measured parameters. 2000 different pure spectra are generated and the function values ​​are normalized.

[0048] S2-2. Adding polynomial baseline

[0049] Based on the pure spectrum generated in step S2-1, a polynomial baseline is added through the polynomial baseline formula to obtain a simulated spectrum with superimposed baselines of different shapes. A certain range is set for the polynomial coefficients to generate polynomial curves of different shapes. The polynomial baseline formula is as follows:

[0050]

[0051] Wherein, m, n, q are polynomial coefficients, and the horizontal coordinate x has the same meaning as in step S2-1. m, n, q are all set to random numbers between -1 and 1 to generate a random baseline. The horizontal coordinate x is divided by 1000 to control the consistency of the horizontal and vertical coordinate measurement scales.

[0052] S2-3. Adding random noise

[0053] Based on step S2-2, Gaussian white noise is added through the awgn function to obtain 2000 simulated spectra of each agent. The function expression is as follows:

[0054] ps=awgn(xin,snr,′measured′)

[0055] Among them, awgn is a function that can add Gaussian white noise with a specific signal-to-noise ratio to the signal, xin and ps are the spectral signals before and after adding Gaussian white noise, snr is the signal-to-noise ratio, which is set to a random number in the range of 30 to 40, and 'measured' is the power of the added noise calculated based on the actual power of the input signal and the signal-to-noise ratio.

[0056] S3. Construct infrared simulation spectrum recognition model

[0057] The simulated spectra of different toxic agents generated in step S2 were integrated and annotated with real category labels to form a sample set, and the sample set was randomly divided into a training set and a validation set at a ratio of 7:3. The infrared telemetry spectrum simulation model was constructed on the Python software platform using a convolutional neural network, the input two-dimensional spectral data was converted into a four-dimensional vector, the convolutional neural network structure and training parameters were adjusted, and the validation set was internally verified using a classification function.

[0058] S4. External verification of infrared simulation spectrum recognition model

[0059] 300 spectra were randomly selected from the 1000 to 2000 infrared spectra of each agent in step S2-1 and processed with polynomial fitting baseline correction as the test set for external verification of the infrared simulation spectrum recognition model in S3. The true labels were annotated, and the spectrum dimensions were converted into four dimensions as model input. The predicted labels were output through the classification function in Python software, and the probability score threshold was set to 0.5. If the probability score was below 0.5, it would not be recognized. The true labels were compared with the predicted labels.

[0060] Specifically, the nerve agents in step S1 include sarin and soman, the corrosive agents include mustard gas and Lewisite, and the types of simulated agents can also be expanded to general toxic and harmful gases, such as ammonia, ethanol, styrene and other common chemical reagents; the self-developed infrared remote sensing spectrometer measurement and spline interpolation band range is the mid-infrared band: 1347cm -1~681cm -1 .

[0061] The collection conditions of different toxic agents and toxic and harmful gases are shown in the following table:

[0062] Table 1 Collection conditions of different toxic agents and toxic and harmful gases

[0063]

[0064]

[0065]

[0066] In steps S3 and S4, the internal and external verification is to use the Python software classification function to calculate and output the predicted label. When the predicted label is consistent with the true label, it indicates that the simulation spectrum recognition model predicts correctly. When the number of correctly predicted samples is close to the total number of test samples, that is, the prediction accuracy is close to 1, it indicates that the classification effect of the simulation spectrum recognition model is good. The model performance evaluation index is judged by the accuracy of the test set, and the accuracy A (Accuracy, A) is expressed as follows:

[0067]

[0068] Among them, N A is the number of samples whose predicted labels are consistent with the true labels, and N is the total number of test samples. The external verification results of the toxic agent remote sensing infrared spectrum simulation recognition model are shown in the following table (scores is the average probability score of the predicted correct samples, and Average is the average recognition accuracy of multiple categories with equal weights):

[0069] Table 2 External validation results of the toxic agent simulation spectrum recognition model

[0070]

[0071]

[0072]

[0073] According to Table 2, the infrared simulation spectrum recognition model of toxic agents has an average recognition accuracy of 95.65% for toxic and harmful gases, indicating that the infrared telemetry spectrum simulation method of toxic agents can be further used in other types of spectrum recognition fields.

[0074] The above are all preferred implementation methods of the present invention. It should be pointed out that for ordinary technicians in this technical field, several modifications and improvements can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for simulating infrared telemetry spectrum of chemical agents, characterized in that: The spectrum simulation method comprises the following steps: S1. Select polar chemical agents and collect their infrared spectra S2. Generate simulated spectrum S2-1. Clean spectrum generation The difference spectrum information collected in S1 was preprocessed by spline interpolation, normalization and data cleaning. 40 measured spectrum data were randomly selected as the simulation basis. The peak height, peak position, peak width and peak number of the characteristic peaks were obtained using Python built-in functions. A single main peak was generated using a Gaussian function. Multiple main peaks were combined according to the peak height ratio to obtain a pure spectrum containing multiple characteristic peaks. The Gaussian function formula is as follows: x is the band range after spline interpolation. The three parameters of the Gaussian function: peak height a, peak position b, and peak width c are set to random numbers fluctuating around the measured parameters. 2000 different pure spectra are generated and the function values ​​are normalized. S2-2. Adding polynomial baseline On the basis of the pure spectrum generated in step S2-1, a polynomial baseline is added through a polynomial baseline formula to obtain a simulated spectrum with superimposed baselines of different shapes; a certain range is set for the polynomial coefficients to generate polynomial curves of different shapes; The polynomial baseline formula is as follows: Where m, n, q are polynomial coefficients, and the horizontal coordinate x has the same meaning as in step S2-1. Set m, n, and q to random numbers between -1 and 1 to generate a random baseline. Divide the horizontal coordinate x by 1000 to ensure that the horizontal and vertical coordinates have the same measurement scale. S2-3. Adding random noise Based on step S2-2, Gaussian white noise is added through the awgn function to obtain 2000 simulated spectra of each agent. The function expression is as follows: ps=awgn(xin,snr,′measured′) Among them, awgn function is a function that can add Gaussian white noise with a specific signal-to-noise ratio to the signal, xin and ps are the spectral signals before and after adding Gaussian white noise, snr is the signal-to-noise ratio, which is set to a random number in the range of 30 to 40, and 'measured' is the power of the added noise calculated according to the actual power of the input signal and the signal-to-noise ratio; S3. Construct infrared simulation spectrum recognition model The simulated spectra of different toxic agents generated in step S2 are integrated and annotated with true category labels to form a sample set, and the sample set is randomly divided into a training set and a validation set in a ratio of 7:3; The infrared telemetry spectrum simulation model was constructed on the Python software platform using convolutional neural networks, the input two-dimensional spectrum data was converted into a four-dimensional vector, the convolutional neural network structure and training parameters were adjusted, and the classification function was used to perform internal verification on the verification set. S4. External verification of infrared simulation spectrum recognition model 300 spectra were randomly selected from the 1000 to 2000 infrared spectra of each agent in step S2-1 and processed with polynomial fitting baseline correction as the test set for external verification of the infrared simulation spectrum recognition model in S3. The true labels were annotated, and the spectrum dimensions were converted into four dimensions as model input. The predicted labels were output through the classification function in Python software, and the probability score threshold was set to 0.

5. If the probability score was below 0.5, it would not be recognized. The true labels were compared with the predicted labels.

2. The spectrum simulation method according to claim 1, characterized in that: The specific implementation process of step S1 is as follows: Adjust the angle and position of the blackbody radiator, gas absorption cell, and infrared spectrometer to ensure that the light flux of the reagent light path is the maximum; Before sampling, the residual gas in the gas absorption cell was exhausted through the vacuum system, and the background spectrum was collected; Common nerve agents and corrosive agents were selected, the pipeline temperature was fixed at 40°C, and different temperature differences between the gas and the pipeline were obtained by adjusting the surface source blackbody radiation temperature; Use a micro-injector to inject samples into the gas absorption cell. The injection dose is determined according to the volatility of the toxic agent and the toxic gas. For gases with good volatility, a semi-emptying method is adopted during emptying to obtain low and medium signal-to-noise ratio data. For gases with poor volatility, different temperature differences and doses are tried to obtain low, medium and high signal-to-noise ratio data. The difference spectrum information of each toxic agent before and after mixing with air and after deducting the gas pool background is collected by infrared remote sensing spectrometer.

3. The spectrum simulation method according to claim 2, characterized in that: The infrared remote sensing test system has a sampling time interval of 1 second, and the difference spectra before and after each agent is mixed with air are collected for 15 minutes each, for a total of 1000 to 2000 lines.

4. The spectrum simulation method according to claim 3, characterized in that: The self-developed infrared remote sensing spectrometer measures and interpolates the mid-infrared band: 1347cm -1 ~681cm -1 .

5. The spectrum simulation method according to claim 2, characterized in that: The nerve agents include sarin and soman, the corrosive agents include mustard gas and Lewisite, and the types of simulated agents can also be expanded to include general toxic and harmful gases.

6. The spectrum simulation method according to claim 1, characterized in that: In steps S3 and S4, the internal and external verifications are performed by using the classification function of Python software to calculate and output the predicted labels. When the predicted labels are consistent with the true labels, it indicates that the simulation spectrum recognition model prediction is correct.

7. The spectrum simulation method according to claim 1, characterized in that: In step S4, when the number of correctly predicted samples is close to the total number of tested samples, that is, the prediction accuracy is close to 1, it indicates that the classification effect of the simulated spectrum recognition model is good.

Images