CN117747008B - A method and system for baseline fitting and noise reduction of gas laser absorption spectrum - Google Patents

Abstract

The present application relates to the field of signal processing technology, and provides a method and system for baseline fitting and noise reduction of gas laser absorption spectrum. The method comprises: sending a detection light signal to the gas to be tested; establishing a first data set and a baseline fitting neural network; using the first data set to train and test the baseline fitting neural network to obtain a baseline fitting signal; establishing a second data set and a noise suppression neural network; using the second data set to train and test the noise suppression neural network to obtain a noise reduction signal. The baseline fitting and noise reduction method does not require additional hardware equipment, nor does it introduce additional hardware noise. In addition, there is no frequency limit on the noise. Compared with the existing processing methods, the baseline signal drift and noise interference problems existing in gas concentration detection are improved. The method has the advantages of fast processing and good noise reduction effect. Moreover, the method is not limited by the structure, and is suitable for small systems and equipment with short optical paths, which effectively reduces the equipment cost and has a wide range of applications.

Description

A method and system for baseline fitting and noise reduction of gas laser absorption spectrum

Technical Field

The present application relates to the field of signal processing technology, and in particular to a baseline fitting and noise reduction method and system for gas laser absorption spectrum.

Background technique

The demand for gas concentration detection is widespread in many fields, and direct absorption spectroscopy technology in absorption spectroscopy is usually used to obtain gas concentration.

Direct Absorption Spectroscopy (DAS) technology has shown great potential in the miniaturization of sensing systems due to its simple system structure, convenient operation, and relatively low requirements for detectors. However, in actual use, DAS technology requires a set of non-absorption signals as baseline signals. Usually, the baseline signal is obtained by selecting the "non-absorption area" data in the transmission spectrum signal for low-order polynomial fitting. However, the selection of "non-absorption area" data and polynomials are both based on experience and there are no fixed rules. Therefore, the baseline signal obtained by the polynomial fitting method often has drift and has large uncertainty, which will cause measurement errors.

In order to obtain a more accurate baseline signal, two beams of detection light signals can be used, one for normal detection and the other as a reference light. The baseline signal obtained in this way is more accurate, but it is easy to generate interference noise. Due to the existence of interference noise, the detection accuracy of gas concentration will be reduced.

Summary of the invention

The present application provides a baseline fitting and noise reduction method and system for gas laser absorption spectrum to improve the technical problems of baseline signal drift and noise interference in gas concentration detection.

The first aspect of the present application provides a baseline fitting and noise reduction method for gas laser absorption spectrum, comprising: sending a detection light signal to the gas to be tested; establishing a first data set and a baseline fitting neural network; using the first data set to train and test the baseline fitting neural network to obtain a baseline fitting signal; wherein the input of the baseline fitting neural network is a transmitted light signal, and the output of the baseline fitting neural network is a non-absorption baseline signal, and both the transmitted light signal and the non-absorption baseline signal are obtained by processing the detection light signal; establishing a second data set and a noise suppression neural network; wherein the second data set includes an absorbance signal, and the absorbance signal is the ratio of the non-absorption baseline signal to the transmitted light signal; using the second data set to train and test the noise suppression neural network to obtain a noise reduction signal, wherein the input of the noise suppression neural network is an absorbance signal, and the output of the noise suppression neural network is a high signal-to-noise ratio absorbance signal or a noise-free simulated absorbance signal, and the high signal-to-noise ratio absorbance signal is obtained by processing the absorbance signal, and the noise-free simulated absorbance signal is obtained by processing the simulation program.

In some feasible implementations, establishing a first data set includes: acquiring multiple baseline training data and multiple baseline test data; wherein the baseline training data and the baseline test data are both obtained by detecting light signals, the baseline training data include a first transmitted light signal and a first non-absorption baseline signal, and the baseline test data include a second transmitted light signal and a second non-absorption baseline signal; the baseline training data and the baseline test data are selected according to a first feature selection principle to obtain the first data set.

In some feasible implementations, establishing the second data set includes: acquiring multiple noise reduction training data and multiple noise reduction test data, wherein the noise reduction training data includes a first absorbance signal and a first high signal-to-noise ratio absorbance signal or a noise-free first simulated absorbance signal, and the noise reduction test data includes a second absorbance signal and a second high signal-to-noise ratio signal or a noise-free second simulated absorbance signal; selecting the noise reduction training data and the noise reduction test data according to the second feature selection principle to obtain the second data set.

In some feasible implementations, the first feature selection principle is to set the values corresponding to the absorption peaks of the first transmitted light signal and the second transmitted light signal to zero; the second feature selection principle is to fit the absorption peaks of the first absorbance signal and the second absorbance signal using a Lorentzian linear function; wherein the Lorentzian linear function for:

;

is the center frequency of the molecular absorbance spectrum of the gas to be measured, /> Represents the full width at half maximum of the molecular absorbance spectrum of the gas to be measured,/> is the frequency of the laser's output light.

In some feasible implementations, the baseline fitting neural network and the noise suppression neural network both include an input layer, a first hidden layer, a second hidden layer, and an output layer; wherein the activation functions between the input layer and the first hidden layer, between the first hidden layer and the second hidden layer, and between the second hidden layer and the third hidden layer are all Sigmoid functions:

;

in, is a natural constant, /> is the input spectrum data of the input layer or the first hidden layer or the second hidden layer; the activation function between the third hidden layer and the output layer is/> ; Among them, /> is the output spectral data of the third hidden layer.

In some feasible implementations, the loss function of the baseline fitting neural network and the noise suppression neural network is the same; the training method of the baseline fitting neural network and the noise suppression neural network is the same.

In some feasible implementations, the baseline fitted neural network is tested using the mean square error, and the mean square error MSE is:

;

in, The output value of the baseline fitted neural network after training, /> is the second non-absorption baseline signal; i is the ith spectral data, and n is the number of input spectral data.

In some feasible implementations, the noise suppression neural network is tested using a signal-to-noise ratio function, where the signal-to-noise ratio function SNR is:

;

in, is the peak value of the absorption peak of the second absorbance signal, /> is the non-absorption peak part of the second absorbance signal.

In some feasible implementations, the baseline fitting and noise reduction method for gas laser absorption spectrum further includes: calculating the concentration of the gas to be measured based on the noise reduction signal.

The second aspect of the present application provides a baseline fitting and noise reduction system for gas laser absorption spectra, including a sending module; configured to send a detection light signal to the gas to be tested; a first establishment module, configured to establish a first data set and a baseline fitting neural network; a first training and testing module, configured to use the first data set to train and test the baseline fitting neural network to obtain a baseline fitting signal; wherein the input of the baseline fitting neural network is a transmitted light signal, and the output of the baseline fitting neural network is a non-absorption baseline signal, and both the transmitted light signal and the non-absorption baseline signal are obtained by processing the detection light signal; the second establishment module, configured to establish a second data set and a noise suppression neural network; wherein the second data set includes an absorbance signal, and the absorbance signal is the ratio of the non-absorption baseline signal to the transmitted light signal; the second training and testing module, configured to use the second data set to train and test the noise suppression neural network to obtain a noise reduction signal; wherein the input of the noise suppression neural network is an absorbance signal, and the output of the noise suppression neural network is a high signal-to-noise ratio absorbance signal obtained by processing the absorbance signal, a noise-free simulated absorbance signal, and the noise-free simulated absorbance signal is obtained by processing the simulation program.

The baseline fitting and noise reduction method of the gas laser absorption spectrum in the present application can adaptively fit the baseline signal according to the transmitted light signal, and can effectively suppress the noise in the original absorbance signal. Moreover, the method is not limited by the structure, is suitable for small systems and short optical path devices, and has a wide range of applications.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

FIG1 is a schematic flow chart of a baseline fitting and noise reduction method for gas laser absorption spectrum provided in an embodiment of the present application;

FIG2a is a schematic diagram of a transmission light signal provided in an embodiment of the present application;

FIG2b is a schematic diagram of feature selection using the first feature selection principle provided in an embodiment of the present application;

FIG3 is a schematic diagram of a process of training and testing a baseline fitting neural network provided in an embodiment of the present application;

FIG4 is a schematic diagram showing a comparison between a simulated and an original absorbance signal provided in an embodiment of the present application;

FIG5a is a schematic diagram of the original absorbance signal of the mixed gas provided in an embodiment of the present application;

FIG5 b is a schematic diagram of the absorbance of the scaled mixed gas provided in an embodiment of the present application;

FIG6 is a schematic diagram of feature selection using the second feature selection principle provided in an embodiment of the present application;

FIG7 is a schematic diagram of a process of training and testing a noise suppression neural network provided in an embodiment of the present application;

FIG8 is a schematic flow chart of a specific application of a method for baseline fitting and noise reduction of a gas laser absorption spectrum provided in an embodiment of the present application;

FIG9 is a schematic diagram of the effect of a baseline fitting provided in an embodiment of the present application;

FIG10a is a schematic diagram of an original absorbance signal provided in an embodiment of the present application;

FIG10b is a schematic diagram of an absorbance signal after noise reduction provided in an embodiment of the present application;

FIG11 is a structural block diagram of a baseline fitting and noise reduction system for gas laser absorption spectrum provided in an embodiment of the present application;

FIG. 12 is a schematic diagram of the structure of a baseline fitting and noise reduction system for gas laser absorption spectra provided in an embodiment of the present application.

Graphic marking:

100 - baseline fitting and noise reduction system; 101 - sending module; 102 - first establishment module; 103 - first training and testing module; 104 - second establishment module; 105 - second training and testing module.

Detailed ways

The technical solutions in the embodiments of the present application will be described clearly below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments of the present application, other embodiments obtained by ordinary technicians in this field without making creative work all belong to the protection scope of the present application.

In the following, the terms "first", "second", etc. are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first", "second", etc. may explicitly or implicitly include one or more of the feature. In the description of this application, unless otherwise specified, "plurality" means two or more.

In addition, in the present application, directional terms such as "upper", "lower", "inner" and "outer" are defined relative to the orientation of the components in the drawings. It should be understood that these directional terms are relative concepts. They are used for relative description and clarification, and they can change accordingly according to the changes in the orientation of the components in the drawings.

The demand for gas concentration detection is widely present in the fields of industrial production safety, environmental monitoring, deep space and deep sea exploration, energy production and utilization, etc. Optical detection methods have become one of the most promising technical routes in current gas detection technology due to their fast response speed, high safety, non-invasive measurement, and great potential in miniaturization and low power consumption.

Direct absorption spectroscopy (DAS) has shown great potential in the miniaturization of sensing systems due to its simple system structure, convenient operation, and relatively low requirements for detectors. However, in actual use, DAS technology requires a set of non-absorption signals as baseline signals. Usually, the baseline signal is obtained by selecting the "non-absorption area" data in the transmission spectrum signal for low-order polynomial fitting. However, the selection of "non-absorption area" data and the selection of polynomials are both based on experience and there are no fixed rules. Therefore, the baseline signal obtained by the polynomial fitting method often has drift and has large uncertainty, which will cause measurement errors. Under weak absorption conditions such as low concentration of the gas to be measured, weak absorption line intensity, and too short effective optical path, the measurement error caused by the polynomial fitting baseline is often unacceptable.

In order to calculate the gas concentration, it is necessary to calculate the original absorbance signal by comparing the baseline fitting signal with the transmission spectrum signal. The original absorbance signal is inevitably interfered by various noises, including various noises in the optical path and circuit. The existence of these noises will reduce the detection performance of the system, especially under the conditions of system miniaturization and short optical path. The impact of noise is relatively greater, so it is necessary to perform noise reduction on the detection signal. Commonly used noise reduction methods include multiple signal accumulation and averaging and S-G filtering. The multiple averaging method has a very good noise reduction effect on white noise, but in theory, the noise reduction effect often depends on the number of averages. In practical applications, the number of averages cannot be increased indefinitely due to the measurement response time requirements and hardware costs. S-G filtering is widely used in data smoothing processing. It can filter out high-frequency noise while ensuring that the shape of the signal remains unchanged, but the processing of local mutation signals may cause large errors. In addition, filtering out low-frequency noise often leads to signal distortion and amplitude reduction, resulting in poor noise reduction effect.

In order to solve the above technical problems, the embodiment of the present application provides a baseline fitting and noise reduction method for gas laser absorption spectrum. The method does not require additional hardware equipment, does not introduce additional hardware noise, and has no frequency limit on the noise. Compared with the existing processing methods, the method has the advantages of fast processing and good noise reduction effect.

FIG1 is a schematic flow chart of a baseline fitting and noise reduction method for a gas laser absorption spectrum provided in an embodiment of the present application.

Referring to FIG. 1 , the method may be implemented by the following steps S100 to S500 .

Step S100: sending a detection light signal to the gas to be tested.

Step S200: Establishing a first data set and a baseline fitting neural network.

In step S200, a first data set may be established first, and then a baseline fitting neural network may be established.

When establishing the first data set, the following steps S201 and S202 may be adopted.

Step S201: Acquire a plurality of baseline training data and a plurality of baseline test data.

The baseline training data and the baseline test data are both obtained by detecting light signals. The baseline training data is used for subsequent training operations on the baseline fitting neural network, and the baseline test data is used for subsequent testing operations on the baseline fitting neural network.

Specifically, the baseline training data includes a first transmitted light signal and a first non-absorption baseline signal, and the baseline test data includes a second transmitted light signal and a second non-absorption baseline signal. The transmitted light signal provided in the embodiment of the present application is a spectral signal.

That is, both the baseline training data and the baseline test data include the transmitted light signal and the non-absorption baseline signal.

Specifically, direct absorption spectroscopy technology usually requires a set of stable non-absorption signals as baseline signals, but due to the use of non-cooled photodetectors, their detection effect is susceptible to changes in the external ambient temperature and the heat accumulation caused by their own work, resulting in a baseline signal offset, causing detection errors. However, the temperature changes caused by the external ambient temperature and the detector work are negligible in a short period of time. Therefore, in the embodiment of the present application, the first transmitted light signal of different concentrations of gas and its corresponding first non-absorption baseline signal, the second transmitted light signal and its corresponding second non-absorption baseline signal can be continuously measured in a short period of time, and a mapping relationship between them can be established to obtain baseline training data and baseline test data.

In this way, the number of the first transmission light signal and the second transmission light signal are both multiple, and the first transmission light signal and the second transmission light signal can be transmission light signals of the gas to be measured under different concentration conditions. The number of the first non-absorption baseline signal is the same as the number of the first transmission light signal, and the number of the second non-absorption baseline signal is the same as the number of the second transmission light signal.

Step S202: Select the baseline training data and the baseline test data according to the first feature selection principle to obtain a first data set.

The principle of the first feature selection is: setting the values corresponding to the absorption peaks of the first transmitted light signal and the second transmitted light signal to zero.

Specifically, the first transmitted light signal and the second transmitted light signal are affected by the concentration effect, which will reduce the correlation between the transmitted light signal and the non-absorption baseline signal and increase the difficulty of network training. Therefore, in order to eliminate the influence of the concentration effect as much as possible, the part of the first transmitted light signal and the second transmitted light signal affected by the concentration, that is, the corresponding value within the absorption peak range is set to zero, and the rest is not processed.

FIG. 2 a is a schematic diagram of a transmitted light signal provided in an embodiment of the present application.

FIG2b is a schematic diagram of feature selection using the first feature selection principle provided in an embodiment of the present application.

In a specific implementation, see Figures 2a and 2b, which show the feature selection results of the baseline fitting neural network, where the horizontal axis is the wave number and the vertical axis is the transmitted light signal. The transmitted light signal in Figure 2a is obtained by detecting the detection light signal. The transmitted light signal will produce a depression due to the absorption of the gas to be measured, and the different concentrations of the gas to be measured will cause different degrees of depression, thereby introducing the concentration effect. In order to reduce the influence of the concentration effect, it is necessary to set the value of the above-mentioned depressed part to 0, and use the remaining part of the transmitted light signal as a feature. At this point, the feature selection of the input signal of the baseline fitting neural network is completed, and the final signal input to the baseline fitting neural network is shown as the dot-dash curve in Figure 2b.

In this way, after processing the first transmitted light signal and the second transmitted light signal, a first data set can be obtained. Specifically, the first data set may include a first training data set and a first test data set, the first training data set includes the processed first transmitted light signal and the first non-absorption baseline signal, and the first test data set includes the processed second transmitted light signal and the second non-absorption baseline signal.

After the first database is established, a baseline fitting neural network can be established.

The establishment of the baseline fitting neural network mainly takes the transmission spectrum signal of the gas to be tested with different concentrations as the input signal, and the corresponding non-absorption baseline signal as the output signal of the baseline fitting neural network.

Step S300: using the first data set to train and test the baseline fitting neural network to obtain a baseline fitting signal.

The input of the baseline fitting neural network is the transmitted light signal, and the output of the baseline fitting neural network is the non-absorption baseline signal. Both the transmitted light signal and the non-absorption baseline signal are obtained by processing the detected light signal.

Specifically, the baseline fitting neural network may be trained by using a first training data set in the first data set, and then the trained baseline fitting neural network may be tested by using a first test data set in the first data set.

FIG3 is a schematic diagram of a process of training and testing a baseline fitting neural network provided in an embodiment of the present application.

Referring to FIG. 3 , in a specific implementation, step S300 may be implemented by the following steps S301 to S306 .

Step S301: The detector detects the transmission spectrum signal (training data) obtained, and the target baseline signal (label) corresponding to the transmission spectrum signal.

Among them, the transmission spectrum signal is the transmission light signal, and the target baseline signal is the non-absorption baseline signal.

In step S301, the training data is used as input data of the baseline fitting neural network, and the label is used as output data of the baseline fitting neural network.

Step S302: Initialize the baseline fitting neural network.

Step S303: training.

Step S304: Fitting baseline signal.

The fitted baseline signal is the baseline fitting signal.

Step S305: Determine whether the mean square error between the fitted baseline and the target baseline meets the requirement. If the result is yes, the training and testing process is completed, and if the result is no, step S306 is executed.

Step S306: Adjust the baseline fitting neural network structure and parameters, and execute step S303.

In step S306, the mean square error (MSE) can be used to determine whether the trained network meets the requirements. The mean square error expression is as follows:

;

in, is the predicted value, that is, the output value of the baseline fitting neural network after training,/> is the label value, that is, the target baseline signal, that is, the second non-absorption baseline signal, i is the i-th spectrum data, and n is the number of input spectrum data.

Specifically, after the training is completed, the baseline fitting neural network may be tested using the first test data set, thereby optimizing the baseline fitting neural network.

Step S400: Establish a second data set and a noise suppression neural network.

The second data set includes an absorbance signal, which is a ratio of a non-absorption baseline signal to a transmitted light signal.

Specifically, in step S400, the second data set may be established first, and then the noise suppression neural network may be established.

The second data set can be established by following steps S401 to S402.

Step S401: Acquire a plurality of denoising training data and a plurality of denoising test data.

The noise reduction training data includes a first absorbance signal and a first high signal-to-noise ratio absorbance signal or a noise-free first simulated absorbance signal, and the noise reduction test data includes a second absorbance signal and a second high signal-to-noise ratio absorbance signal or a noise-free second simulated absorbance signal. The first absorbance signal can be calculated by comparing the first non-absorption baseline signal to the first transmitted light signal, and the second absorbance signal can be calculated by comparing the second non-absorption baseline signal to the second transmitted light signal.

Among them, the noise-free simulated absorbance signal is output first. When the noise-free simulated absorbance signal cannot be output or the noise-free simulated absorbance signal cannot be output better, the high signal-to-noise ratio absorbance signal is selected for output. Because there is sometimes a non-negligible difference between the original absorbance signal and the simulated absorbance signal under the same conditions, the simulated absorbance signal at this time cannot be used as the output signal of the noise suppression neural network. At this time, the high-concentration original absorbance signal that has a linear relationship with the input signal is often used as the output signal through a linear scaling method. Among them, the original absorbance signal represents the absorbance signal that has not been subjected to noise reduction processing, that is, the input signal of the noise suppression neural network.

FIG. 4 is a schematic diagram showing a comparison between a simulated and an original absorbance signal provided in an embodiment of the present application.

Referring to FIG. 4 , taking the absorbance signal of a mixed gas of 50 ppm ethane and 40 ppm methane at room temperature and pressure as an example, it is shown that there is an obvious difference between the simulated absorbance of the mixed gas and the original absorbance signal.

Direct absorption spectroscopy technology is based on the Beer-Lambert law. When the wavelength of the light emitted by the laser covers the characteristic absorption spectrum of the gas to be measured, the photon energy at the response wavelength will be absorbed by the gas molecules to be measured, causing the optical power of the outgoing light at the detection end to attenuate and conform to the following formula:

;

in Indicates the intensity of transmitted light, which is the detection signal after being absorbed by the gas to be tested; /> Indicates the intensity of emitted light, which is obtained by the detection signal when nitrogen is passed in actual detection; /> is the output light frequency of the laser, /> is the absorption coefficient, /> is the concentration of the gas to be measured, /> is the effective optical path.

In actual detection, the gas concentration is reflected by absorbance to achieve detection. defined as:

;

Where S(T) is the intensity of the absorption line, T is the temperature, and P is the pressure. is a Lorentzian linear function.

From the above formula, it can be found that when other conditions are the same, there is a linear relationship between the absorbance signal and the gas concentration. Therefore, in theory, the absorbance signal of low concentration can be obtained by linearly scaling the high concentration absorbance signal according to the concentration ratio. Since the system noise is fixed, and the absorbance signal is larger under high concentration conditions and the noise is relatively small, the absorbance signal with a higher signal-to-noise ratio can be obtained through the above method.

Step S402: selecting the denoised training data and the denoised test data based on the second feature selection principle to obtain a second data set.

The second feature selection principle is that the numerical values corresponding to the absorption peaks in the first absorbance signal and the second absorbance can be fitted by a Lorentz linear function, and the values of the remaining parts are set to zero.

Specifically, the Lorentz line function for:

;

is the center frequency of the molecular absorbance spectrum of the gas to be measured, /> is the half-maximum full width of the molecular absorbance spectrum of the gas to be measured, /> is the output light frequency of the laser.

FIG5 a is a schematic diagram of the original absorbance signal of the mixed gas provided in an embodiment of the present application.

FIG5 b is a schematic diagram of the absorbance of the scaled mixed gas provided in an embodiment of the present application.

For the problem in FIG. 4 above, see FIG. 5a and FIG. 5b , correction can be performed by reducing the original absorbance signal of the mixed gas of 500 ppm ethane and 400 ppm methane by 10 times, thereby approximately obtaining an absorbance signal of the mixed gas of 50 ppm ethane and 40 ppm methane with a high signal-to-noise ratio.

In this way, multiple detection and data processing steps are performed for the gas to be tested with different concentration gradients, and multiple noise reduction training data and multiple noise reduction test data can be initially obtained. The second data set can be obtained after the second feature selection.

FIG6 is a schematic diagram of feature selection using the second feature selection principle provided in an embodiment of the present application.

In a specific implementation, see Figure 6, the horizontal axis is the wave number, the vertical axis is the absorbance, and the solid line curve in Figure 6 is the original absorbance signal, but because the gas to be tested has absorption in the used band range, the original signals are all greater than 0, so it is corrected in the embodiment of the present application, and the absorption peak is pulled down to near the value of 0, and finally the absorption peak is fitted with the Lorentz linear function to obtain the dotted line curve in the figure, which is the input signal of the noise suppression neural network. This example is the absorbance signal of a mixed gas of 50ppm ethane and 40ppm methane, where the absorption peaks at 2986.6cm ^-1 and 2990cm ^-1 are ethane absorption peaks, so only one of the absorption peaks needs to be fitted to obtain all the information about the concentration of ethane gas. There are multiple absorption peaks for methane at the position of 2989cm ^-1 , and the superposition of multiple absorption peaks causes its shape to not conform to the Lorentz linear type, but a part of it still conforms to the Lorentz linear type, so only a part of the methane peak needs to be fitted with the Lorentz linear type. In this way, feature selection is completed.

In this way, after processing the first absorbance signal and the second absorbance signal, a second data set can be obtained. Specifically, the second data set may include a second training data set and a second test data set, the second training data set includes the processed first absorbance signal and a first high signal-to-noise ratio absorbance signal or a noise-free first simulated absorbance signal; the second test data set includes the processed second absorbance signal and a second high signal-to-noise ratio absorbance signal or a noise-free second simulated absorbance signal.

After the second database is established, a noise suppression neural network can be established.

The establishment of the noise suppression neural network is to establish a mapping relationship between the first absorbance signal and the second absorbance signal converted from the detected detection light signal, and the absorbance signal with a high signal-to-noise ratio or noise-free. The first absorbance signal and the second absorbance signal are used as input signals of the noise suppression neural network, and the output signal is usually a simulated absorbance signal under the same conditions as the input signal or a high signal-to-noise ratio absorbance signal after linear scaling. Among them, the high signal-to-noise ratio absorbance signal is obtained by absorbance signal processing, and the noise-free simulated absorbance signal is obtained by simulation program processing.

In a specific implementation, the training methods of the baseline fitting neural network and the noise suppression neural network are the same, and both the baseline fitting neural network and the noise suppression neural network include an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer.

Among them, the activation functions between the input layer and the first hidden layer, between the first hidden layer and the second hidden layer, and between the second hidden layer and the third hidden layer are all Sigmoid functions:

;

in, is a natural constant, /> is the input spectral data of the input layer or the first hidden layer or the second hidden layer;

The activation function between the third hidden layer and the output layer is ; Among them, /> is the output spectral data of the third hidden layer.

The loss functions of the baseline fitting neural network and the noise suppression neural network can be the same, and both can use the mean square error function as the loss function.

Step S500: using the second data set to train and test the noise suppression neural network to obtain a noise reduction signal.

The input of the noise suppression neural network is an absorbance signal, and the output of the noise suppression neural network is an absorbance signal with a high signal-to-noise ratio or a noise-free simulated absorbance signal.

Specifically, the noise suppression neural network is first trained using the second training data set in the second data set, and then the trained noise suppression neural network is tested using the second test data set in the second data set.

In a specific implementation, the training methods of the baseline fitting neural network and the noise suppression neural network can be the same, and both can use the Bayesian regularization method. In the Bayesian regularization training algorithm, it is necessary to first define the prior distribution of the model parameters. . Among them/> For the first/> The network parameters of each layer. Then we need to calculate the posterior distribution. According to the Bayesian formula, we can calculate the posterior distribution of each layer of the network by observing the data set D. As shown in the first The output of the layer network model is/> , the corresponding target value is/> .

According to the Bayesian formula:

;

in, Represents the probability of observing the target value given the model output and parameters.

Furthermore, we need to maximize the posterior probability, that is, select the parameter value with the largest posterior probability. Then, we update the parameters based on the result of maximizing the posterior probability.

Finally, the obtained model is applied to test its effect and adjustments and optimizations are made according to the experimental results.

FIG. 7 is a schematic diagram of a process of training and testing a noise suppression neural network provided in an embodiment of the present application.

Referring to FIG. 7 , in a specific implementation, step S500 may be implemented by the following steps S501 to S506 .

Step S501: absorbance signal with noise (training data), absorbance signal without noise (label). The absorbance signal with noise is the original absorbance signal, that is, the ratio of the non-absorption baseline signal to the transmitted light signal, and the absorbance signal without noise is the absorbance signal with high signal-to-noise ratio or the simulated absorbance signal without noise.

In step S501, the training data is used as input data of the noise suppression neural network, and the label is used as output data of the noise suppression neural network.

Step S502: Initialize the noise suppression neural network.

Step S503: training.

Step S504: absorbance signal after noise reduction.

Step S505: Determine whether the signal-to-noise ratio requirement is met. If the result is yes, the training and testing process is completed, and if the result is no, step S506 is executed.

Step S506: Adjust the noise suppression neural network structure and parameters, and execute step S503.

In step S506, the signal-to-noise ratio function SNR may be:

;

in, is the peak value of the absorption peak of the second absorbance signal, /> is the non-absorption peak part of the second absorbance signal.

Specifically, after the trained noise suppression neural network meets the condition of the signal-to-noise ratio function SNR, the training and testing process of the noise suppression neural network is completed.

In some feasible implementations, the baseline fitting and noise reduction method may further include step S600.

Step S600: Calculate the concentration of the gas to be measured according to the noise reduction signal.

In step S600, after the training and testing of the two neural networks are completed, the concentration of the gas to be measured can be calculated using the obtained noise reduction signal.

FIG8 is a flow chart of a specific application of a baseline fitting and noise reduction method for a gas laser absorption spectrum provided in an embodiment of the present application.

Referring to FIG. 8 , in a specific implementation, the concentration of the gas to be measured can be calculated by the baseline fitting and noise reduction method of the gas laser absorption spectrum provided in the embodiment of the present application, which can be specifically implemented by the following steps S001 to S600 - 1.

Step S001: The detector obtains a transmitted light signal.

Specifically, the transmitted light signal can be converted from the detected light signal.

Step S201 - 1 : performing feature selection on the transmitted light signal.

Among them, step S201-1 can be a sub-step in step S201.

Step S300-1: training the baseline fitting neural network.

Step S300 - 2 : Fitting baseline signal.

Among them, step S300-1 and step S300-2 can be two steps in step S300.

Step S401 - 1 : Compare the fitted baseline signal with the transmitted light signal to obtain the absorbance signal of the gas to be measured.

After obtaining the fitted baseline signal, it is necessary to compare the fitted baseline signal with the transmission spectrum signal to obtain the original absorbance signal.

Step S402-1: performing feature selection on the absorbance signal.

Among them, step S401-1 and step S402-1 may be two steps in step S400.

Step S501-1: Trained noise suppression neural network.

Step S502-1: De-noised absorbance signal of the gas to be measured.

Among them, step S501-1 and step S502-1 may be two steps in step S500.

Step S600 - 1 : inverting the concentration of the gas to be measured according to the peak value of the absorption peak of the gas to be measured in the absorbance signal.

Among them, step S600-1 can be a step in step S600.

FIG. 9 is a schematic diagram of the effect of a baseline fitting provided in an embodiment of the present application.

Referring to FIG9 , the signal after feature selection is input into the trained and tested baseline fitting neural network, and the baseline fitting neural network outputs a baseline fitting signal. The fitting effect is shown in FIG9 .

FIG. 10 a is a schematic diagram of an original absorbance signal provided in an embodiment of the present application.

FIG10 b is a schematic diagram of the absorbance signal after noise reduction provided in an embodiment of the present application.

10a and 10b, the absorbance signal after feature selection will be used as the input signal of the noise suppression neural network, and the noise suppression neural network after training and testing will output the absorbance signal after noise reduction.

The concentration of the gas to be measured can be inverted according to the peak value of the absorbance signal after noise reduction. Among them, according to the peak values of methane and ethane, it can be calculated that the SNR of methane has increased from 16.6dB before noise reduction to 20.5dB, and the SNR of ethane has increased from 22.69dB before noise reduction to 26.26dB.

Specifically, the baseline fitting and noise reduction method of the gas laser absorption spectrum provided in the embodiment of the present application can adaptively fit the baseline signal according to the transmitted light signal, improve the problems of baseline signal drift and noise interference in gas concentration detection, and can effectively suppress the noise in the absorbance signal. This method does not require additional hardware equipment, nor does it introduce additional hardware noise. In addition, there is no frequency limit on the noise. Compared with the existing processing methods, this method has the advantages of fast processing and good noise reduction effect. Moreover, this method is not limited by the structure and is suitable for small systems and equipment with short optical paths, effectively reducing equipment costs and having a wide range of applications.

FIG11 is a structural block diagram of a baseline fitting and noise reduction system for gas laser absorption spectrum provided in an embodiment of the present application.

Referring to FIG11 , corresponding to the above-mentioned embodiments, the present application also provides an embodiment of a baseline fitting and noise reduction system 100 for gas laser absorption spectrum, which is applied to the baseline fitting and noise reduction method provided in any of the above-mentioned embodiments. The baseline fitting and noise reduction system 100 includes a sending module 101, a first establishment module 102, a first training and testing module 103, a second establishment module 104, and a second training and testing module 105.

The sending module 101 is used to send a detection light signal to the gas to be tested.

The first establishing module 102 is used to establish a first data set and a baseline fitting neural network.

The first training and testing module 103 is used to train and test the baseline fitting neural network using the first data set to obtain a baseline fitting signal; wherein the input of the baseline fitting neural network is the transmitted light signal, and the output of the baseline fitting neural network is the non-absorption baseline signal, and both the transmitted light signal and the non-absorption baseline signal are obtained by processing the detection light signal.

The second establishing module 104 is used to establish a second data set and a noise suppression neural network, wherein the second data set includes an absorbance signal, which is a ratio of a non-absorption baseline signal to a transmitted light signal.

The second training and testing module 105 is used to train and test the noise suppression neural network using the second data set to obtain a noise reduction signal; wherein, the input of the noise suppression neural network is an absorbance signal, and the output of the noise suppression neural network is a high signal-to-noise ratio absorbance signal or a noise-free simulated absorbance signal, and the high signal-to-noise ratio absorbance signal or the noise-free simulated absorbance signal is obtained by processing the absorbance signal.

FIG12 is a schematic diagram of the structure of a baseline fitting and noise reduction system provided in an embodiment of the present application.

In a specific implementation, see Figure 12, the modulation signal generated by the signal generator is input into the laser controller, and then the laser is tuned so that the output wavelength of the laser covers the target detection band. Then the laser emits a detection light signal, which is collimated by a collimating lens and then shot into the gas chamber. The gas chamber is filled with the gas to be tested, and then the photoelectric detection module converts the detected light signal into an electrical signal and amplifies it. Finally, the data acquisition card collects the transmitted light signal and uploads it to the computer.

The laser may be the sending module 101 in the above embodiment, and the computer may include the first establishment module 102, the first training and testing module 103, the second establishment module 104 and the second training and testing module 105 provided in the above embodiment. In order to adjust the temperature and pressure, the gas chamber may be provided with a thermometer and a pressure gauge.

The computer may further include a calculation module, and the calculation module is used to calculate the concentration of the gas to be measured according to the noise reduction signal.

Specifically, the baseline fitting and noise reduction system for gas laser absorption spectrum provided in the embodiment of the present application can adaptively fit the baseline signal according to the transmitted light signal, improve the baseline signal drift and noise interference problems existing in gas concentration detection, and can effectively suppress the noise in the original absorbance signal. The device can be miniaturized and the production cost can be effectively reduced.

It should be noted that those skilled in the art will easily think of other embodiments of the present application after considering the specification and practicing the application disclosed herein. The present application is intended to cover any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary technical means in the art that are not disclosed in the present application. The specification and examples are only regarded as exemplary, and the true scope of the present application is indicated by the claims.

It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims (8)

1. A baseline fitting and noise reduction method for gas laser absorption spectrum, characterized by comprising: Sending a detection light signal to the gas to be tested; Establishing a first data set and a baseline fitting neural network; The first data set is used to train and test the baseline fitting neural network to obtain a baseline fitting signal; wherein the input of the baseline fitting neural network is a transmitted light signal, the output of the baseline fitting neural network is a non-absorption baseline signal, and both the transmitted light signal and the non-absorption baseline signal are obtained by processing the detection light signal; Establishing a second data set and a noise suppression neural network; wherein the second data set includes an absorbance signal, and the absorbance signal is a ratio of the non-absorption baseline signal to the transmitted light signal; The noise suppression neural network is trained and tested using the second data set to obtain a noise reduction signal; wherein the input of the noise suppression neural network is the absorbance signal, and the output of the noise suppression neural network is a high signal-to-noise ratio absorbance signal or a noise-free simulated absorbance signal, the high signal-to-noise ratio absorbance signal is obtained by processing the absorbance signal, and the noise-free simulated absorbance signal is obtained by processing a simulation program; The establishing of the first data set comprises: Acquire a plurality of baseline training data and a plurality of baseline test data; wherein the baseline training data and the baseline test data are both obtained by the detected light signal, the baseline training data includes a first transmitted light signal and a first non-absorption baseline signal, and the baseline test data includes a second transmitted light signal and a second non-absorption baseline signal; The baseline training data and the baseline test data are selected according to a first feature selection principle to obtain the first data set; The establishing of the second data set comprises: Acquire a plurality of noise reduction training data and a plurality of noise reduction test data; wherein the noise reduction training data comprises a first absorbance signal and a first high signal-to-noise ratio absorbance signal or a noise-free first simulated absorbance signal, and the noise reduction test data comprises a second absorbance signal and a second high signal-to-noise ratio absorbance signal or a noise-free second simulated absorbance signal; The denoising training data and the denoising test data are selected according to the second feature selection principle to obtain the second data set; The first feature selection principle is to set the values corresponding to the absorption peaks of the first transmitted light signal and the second transmitted light signal to zero; The second feature selection principle is to fit the absorption peaks of the first absorbance signal and the second absorbance signal using a Lorentzian linear function. 2. The baseline fitting and noise reduction method for gas laser absorption spectrum according to claim 1, characterized in that: The Lorentzian Linear Function for: is the center frequency of the molecular absorbance spectrum of the gas to be measured, /> is the half-maximum full width of the molecular absorbance spectrum of the gas to be measured, /> is the output light frequency of the laser. 3. The baseline fitting and noise reduction method for gas laser absorption spectrum according to claim 2, characterized in that: The baseline fitting neural network and the noise suppression neural network both include an input layer, a first hidden layer, a second hidden layer, a third hidden layer and an output layer; wherein the activation functions between the input layer and the first hidden layer, between the first hidden layer and the second hidden layer, and between the second hidden layer and the third hidden layer are all Sigmoid functions: Among them, e is a natural constant,/> is the input spectral data of the input layer or the first hidden layer or the second hidden layer; The activation function between the third hidden layer and the output layer is ; Among them, /> is the output spectral data of the third hidden layer. 4. The baseline fitting and noise reduction method for gas laser absorption spectrum according to claim 3, characterized in that: The loss functions of the baseline fitting neural network and the noise suppression neural network are the same; The training methods of the baseline fitting neural network and the noise suppression neural network are the same. 5. The baseline fitting and noise reduction method for gas laser absorption spectrum according to claim 4, characterized in that: The baseline fitting neural network is tested using mean square error, and the mean square error MSE is: Among them,/> is the output value of the baseline fitting neural network after training, is the second non-absorption baseline signal; i is the ith spectral data, and n is the number of input spectral data. 6. The method for baseline fitting and noise reduction of gas laser absorption spectrum according to claim 5, characterized in that: The noise suppression neural network is tested using a signal-to-noise ratio function, where the signal-to-noise ratio function SNR is: Among them,/> is the peak value of the absorption peak of the second absorbance signal, /> is the non-absorption peak part of the second absorbance signal. 7. The baseline fitting and noise reduction method for gas laser absorption spectrum according to claim 1, characterized in that: The baseline fitting and noise reduction method for gas laser absorption spectrum also includes: calculating the concentration of the gas to be measured according to the noise reduction signal. 8. A baseline fitting and noise reduction system for gas laser absorption spectrum, characterized in that it is applied to the baseline fitting and noise reduction method for gas laser absorption spectrum according to any one of claims 1 to 7, and the baseline fitting and noise reduction system for gas laser absorption spectrum comprises: A sending module; configured to send a detection light signal to the gas to be tested; A first establishment module is configured to establish a first data set and a baseline fitting neural network; A first training and testing module is configured to train and test the baseline fitting neural network using the first data set to obtain a baseline fitting signal; wherein the input of the baseline fitting neural network is a transmitted light signal, and the output of the baseline fitting neural network is a non-absorption baseline signal, and both the transmitted light signal and the non-absorption baseline signal are obtained by processing the detection light signal; A second establishment module is configured to establish a second data set and a noise suppression neural network; wherein the second data set includes an absorbance signal, and the absorbance signal is a ratio of the non-absorption baseline signal to the transmitted light signal; A second training and testing module is configured to train and test the noise suppression neural network using the second data set to obtain a noise reduction signal; wherein the input of the noise suppression neural network is the absorbance signal, and the output of the noise suppression neural network is a high signal-to-noise ratio absorbance signal or a noise-free simulated absorbance signal, wherein the high signal-to-noise ratio absorbance signal is obtained by processing the absorbance signal, and the noise-free simulated absorbance signal is obtained by processing a simulation program; The establishing of the first data set comprises: acquiring a plurality of baseline training data and a plurality of baseline test data; wherein the baseline training data and the baseline test data are both obtained through the detection light signal, the baseline training data comprises a first transmission light signal and a first non-absorption baseline signal, and the baseline test data comprises a second transmission light signal and a second non-absorption baseline signal; the baseline training data and the baseline test data are selected according to a first feature selection principle to obtain the first data set; The establishing of the second data set comprises: acquiring a plurality of noise reduction training data and a plurality of noise reduction test data; wherein the noise reduction training data comprises a first absorbance signal and a first high signal-to-noise ratio absorbance signal or a noise-free first simulated absorbance signal, and the noise reduction test data comprises a second absorbance signal and a second high signal-to-noise ratio absorbance signal or a noise-free second simulated absorbance signal; the noise reduction training data and the noise reduction test data are selected according to the second feature selection principle to obtain the second data set; The first feature selection principle is to set the values corresponding to the absorption peaks of the first transmitted light signal and the second transmitted light signal to zero; the second feature selection principle is to fit the absorption peaks of the first absorbance signal and the second absorbance signal using a Lorentzian linear function.