A wavelength modulated laser gas absorption spectrum parameter extraction method and system

Abstract

The present invention provides a method and system for extracting parameters of wavelength modulated laser gas absorption spectrum. The system includes a signal generation module, a laser, an interferometer, a fiber splitting module, an optical fiber, a signal acquisition module, a modulation spectrum wavelength recording module and an online harmonic analysis module based on FPGA. According to the time-wavelength relationship and the gas absorption function relationship recorded by the wavelength recording module, a training structure is a neural network model whose input is a normalized high-order harmonic peak point and output is a gas absorption spectrum parameter, and the model parameters are generated. The model structure and parameter values ​​are then transplanted to the FPGA to achieve online solution of harmonic signals and absorption parameters. The feature of the present invention is that it utilizes the determinism of the wavelength modulation model and the online calculation advantage of the FPGA, which not only reduces the complexity of the modulation absorption spectrum signal acquisition, but also broadens the measurement range of the wavelength modulation method for gas parameters such as temperature and concentration, and has broad application prospects.

Description

Technical Field

[0001] The present invention relates to a method for solving gas concentration and temperature, and in particular to a parameter extraction method and system based on wavelength modulated laser gas absorption spectrum, belonging to the technical field of laser spectrum and gas parameter measurement. Background Art

[0002] Tunable diode laser absorption spectroscopy (TDLAS) uses narrow-band laser to scan the absorption spectrum of gas, obtains the absorption rate of gas molecules to laser, and then obtains the concentration, temperature and other information of the gas to be measured. Since the 1970s, this technology has been used to measure parameters in combustion fields, and has the characteristics of non-invasiveness, fast measurement speed and high sensitivity.

[0003] At present, the main methods of TDLAS are direct absorption spectroscopy (DAS) and wavelength modulation spectroscopy (WMS). The purpose of these two methods is to accurately obtain the gas absorption linear function. DAS directly fits the ratio of the transmitted light intensity to the incident light intensity to obtain the gas absorption linear function. The fitted absorption linear function includes information such as the temperature, concentration and pressure of the gas to be measured. However, the direct absorption method is easily affected by factors such as particle concentration, laser intensity fluctuations and spectral line overlap under high pressure, which leads to an increase in measurement errors. The direct absorption method is more suitable for strong absorption conditions. In 1981, Adam J Reid et al. published a paper titled "Second-harmonic detection with tunable diode lasers—comparison of experiment and theory" on page 203 of Volume 26 of the Journal of Applied Physics. They proposed a wavelength modulation method, which performed high-frequency modulation on the narrow-band laser scanning process and used the high-order harmonics of the high-frequency signal to calculate the gas absorption rate and extract the gas parameters. This method effectively reduced the background interference in the measurement system and improved the measurement accuracy of TDLAS. However, the calculation of gas parameters using high-order harmonic signals alone requires calibration of known gases, which increases the complexity of the system. Wei Wei et al. published the paper "Demonstration of non-absorbing interference rejection using wavelength modulation spectroscopy in high-pressure shock tubes" in the 9th paper "Demonstration of non-absorbing interference rejection using wavelength modulation spectroscopy in high-pressure shock tubes" in the 125th volume of Applied Physics B in 2019. The non-absorbing interference rejection ability of wavelength modulation spectroscopy was directly compared with the measurement values ​​of the direct absorption method in the shock tube experiment. The WMS detection system was evaluated under the reflected shock pressure of 3.5atm, 8.5atm and 18atm. The results showed that the signal-to-noise ratio of the WMS detection system was improved compared with the parallel DAS measurement. The WMS method generally controls the laser by superimposing a sinusoidal signal on the sawtooth wave, and then demodulates the collected signal to obtain some parameters of the laser absorption line shape.Due to the modulation and demodulation process, the signal-to-noise ratio of the signal measured by WMS is higher than that of the DAS method, but at the same time, the absorption spectrum cannot be fully restored due to the filtering and demodulation process.

[0004] In addition, during the modulation process, there is a strong nonlinear relationship between the wavelength and intensity of the laser, which also affects the realization of the measurement. In response to this problem, relevant research work has also been widely carried out. Abhishek Upadhyay et al. published a paper titled "A New RAM Normalized 1f-WMS Technique for the Measurement of Gas Parameters in Harsh Environments and a Comparison With 2f / 1f" on page 99 of Volume 1 of the IEEE Photonics Journal in 2018. It introduces a calibration-free first harmonic (1f) wavelength modulation spectroscopy technique for measuring gas component parameters. In this technique, the total amplitude of the 1f-WMS signal is normalized by a component of the 1f residual amplitude modulation signal. This method retains the advantages of traditional nf / 1f, such as immunity to non-absorption system losses and accurate recovery of gas parameters, and does not require normalization of the non-absorption region at high pressure or high modulation depth. In 2019, Jianxin Liu et al. published a paper titled "A novel methodology to directly predetermine the relative wavelength response of DFB laser in wavelength modulation spectroscopy" on page 1249 of Volume 27 of Optics Express. The paper introduced a novel method for directly measuring the relative wavelength response and gave a DFB laser modulation method based on current linear scanning wavelength response and current modulation wavelength response in wavelength modulation spectroscopy (WMS), which effectively improved the measurement accuracy. However, it brought difficulties to the modulation of the laser, limiting the versatility of the method.

[0005] In order to obtain absorption spectrum parameters based on WMS signals, a method of linear fitting directly for harmonic signals of WMS signals in combination with databases has been widely proposed. In 2009, Hong Duan et al. published a paper titled "Harmonic wavelet analysis of modulated tunable diode laser absorption spectroscopy signals" on page 401 of Volume 48 of Applied Optics. The paper analyzed the tunable semiconductor laser absorption spectroscopy signals with wavelet analysis, which can simultaneously extract 1f, 2f and higher harmonics from direct absorption data. There are also relevant papers reporting on the use of harmonic signals. In 1999, Ove Axner et al. published a paper titled "Theoretical description based on Fourier analysis of wavelength-modulation spectrometry in terms of analytical and background signals" in Applied Optics, Vol. 38, p. 5803. The paper gave a theoretical description of wavelength-modulation spectrometry in terms of analytical and background signals based on Fourier analysis, proposed a form that is not limited to low modulation amplitudes, and established a method for eliminating background signal interference using 2f harmonic signals. Ronald K Hanson et al. published a paper titled "Calibration-free wavelength-modulation spectroscopy for measurements of gas temperature and concentration in harsh environments" in Applied Optics, Vol. 48, p. 5546 in 2009. The paper introduced a calibration-free method for 2f / 1f signals. This method corrected the second harmonic signal with the first harmonic signal, eliminated the influence of laser intensity fluctuations, photoelectric gain and other factors, and derived an expression for absolute gas concentration.Afterwards, many methods for utilizing wavelength modulation and frequency doubling signals have emerged, such as the article "Analysis of calibration-free wavelength-scanned wavelength modulation spectroscopy for practical gas sensing using tunable diode lasers" published by Sun.Kai et al. in 2013 in Measurement Science & Technology, Vol. 24, p. 125203, which proposed a method for analyzing and extracting harmonic information based on the Fourier model. Christopher S. Goldenstein et al. published an article "Fitting of calibration-free scanned-wavelength modulation spectroscopy spectra for determination of gas properties and absorption line shapes" in Applied Optics, Vol. 53, p. 356 in 2014, which introduced a method for fitting the nf / 1f harmonics and obtaining relevant absorption spectrum information. In 2015, Zhechao Qu et al. published a paper titled "Calibration-free scanned wavelength modulation spectroscopy-application to H2O and temperature sensing in flames" on page 164929 of Optics Express, volume 23. They proposed a calibration-free scanned wavelength modulation spectroscopy scheme that is not affected by the modulation depth. In a single fit of a set of 2f / 1f-WMS line shapes obtained in a laser scan, the species concentration and temperature are obtained simultaneously. However, most of these analysis methods for wavelength modulation harmonics require simulation of processes including filtering, so as to iteratively fit the experimental data with the simulation data, which is a complicated process.

[0006] Based on the above problems, some researchers began to abandon the direct use of WMS harmonic signals and began to extract new characteristic information from WMS harmonic signals for solving absorption spectrum parameters. Zhou Peili et al. published a paper titled "Reproduction Method Based on Even Harmonic Absorption Rate Function" on page 293 of Volume 50 of Laser & Infrared in 2020. They proposed a method for recovering the absorption spectrum line type using the even harmonic components of the wavelength modulated signal. This method constructs the relationship between the absorption spectrum line type and the higher harmonics of the modulation signal based on the Taylor expansion form of the absorption spectrum line type, and is therefore limited by the conditions of Taylor expansion, that is, the approximation under small modulation conditions. Yun Pan et al. published an article titled "Improvement of Concentration Inversion Model Based on Second Harmonic Valley Spacing in Wavelength Modulation Spectroscopy" on page 164929 of Volume 23 of IEEE Access in 2020. The article introduces the change of the second harmonic peak due to the change of modulation depth, and proposes a method to calibrate the modulation depth and use it as a parameter of the concentration inversion model. The method verifies the effectiveness of using valley spacing for modulation depth calibration and the accuracy of the improved model. Although this method avoids many error problems in line fitting, it requires different calibrations for different lasers, which brings great difficulties to the application of the method. Although the above methods can establish a mapping relationship between harmonic information and gas parameters, the characteristic points of these harmonics are often related not only to the absorption spectrum, but also to the characteristics of the laser. These methods are derived after idealizing the laser model as much as possible. Therefore, in actual use, it is necessary to control the working range of the laser, which limits such methods in terms of modulation depth and test speed.

[0007] In recent years, relevant domestic patents have also described wavelength modulation measurement devices and methods. Chinese patent CN102680428A "A method for online measurement of gas temperature and concentration based on first harmonic signal" (patent number: 201210152689.0) introduces a measurement scheme for extracting the first harmonic component of the modulated laser signal detected by the laser detector after absorption by a phase-locked amplifier, and gives a method for recovering gas parameters from the first harmonic signal. The representative feature of this method is that the phase-locked amplifier is used to extract the harmonic component, which is a common means of signal extraction in general wavelength modulation methods. Chinese patent CN 103868885 A (patent number: 201410119964.8) "Online measurement method of gas concentration based on composite multiple harmonics" uses the fixed point technology of 4f / 2f signals to estimate the spectrum width on the basis of combining the calibration-free method of 2f / 1f, and makes good use of the information carried by the high-frequency signal to supplement the parameter solution. However, this method has strict requirements on the modulation signal and also needs to make corresponding adjustments according to the laser model, and the operation process is relatively complicated. Chinese patent CN 109696415 A "An online measurement method for gas absorption rate based on fast Fourier transform" (patent number: 201910036652.3) proposes an online measurement method for gas absorption rate based on fast Fourier transform, which uses a data acquisition card to digitize the light intensity signal detected by the photodetector and then perform fast Fourier transform to obtain the high-order harmonic signal of the related wavelength modulation signal. The introduction of these methods has greatly promoted the development of wavelength modulation technology, but these methods mainly focus on the solution of 2f / 1f signals when analyzing wavelength modulation signals, and in order to achieve accurate analysis of harmonic signals, the original photoelectric signal of the wavelength modulation method needs to be sampled at high speed, which puts high demands on the acquisition system and signal transmission system of the wavelength modulation method signal. Chinese patent CN 104535528 A "BP neural network real-time extraction of TDLAS gas absorption spectrum absorbance method" (patent number: 201410699109.9) proposes a method based on FPGA embedded BP (Back Propagation) neural network real-time extraction of TDLAS gas absorption spectrum absorbance method, using neural network technology to fit and reconstruct DAS signals, obtain the half width, absorption area and peak value information of the linear function, and then extract gas concentration, temperature and other parameter information, making good use of BP neural network to solve the nonlinear problem of DAS method to solve gas parameters. When solving the problem of nonlinear solution of gas parameters of wavelength modulation method signals, there is currently a lack of effective and fast extraction methods for neural network input parameters and nonlinear solution methods for gas parameters that integrate high-order harmonic signals.

[0008] The present invention proposes a parameter extraction method and system for wavelength modulated laser gas absorption spectrum, which is used to solve the above-mentioned signal online analysis and nonlinear solution of gas parameters in the wavelength modulation method. It is mainly divided into two steps. The first step is to construct a training set based on the wavelength-intensity modulation characteristics of the distributed feedback laser and the absorption spectrum information under the set temperature, pressure and concentration conditions provided by the HITRAN database, and train a network model based on the BP neural network from the nf / 1f signal to the temperature, pressure and concentration. The second step is to use an FPGA-based digital online harmonic analyzer to extract the frequency-doubled signal of the modulation frequency of the wavelength modulated laser signal after gas absorption, calculate the nf / 1f signal, and then input the extracted nf / 1f signal into the trained network model to obtain the gas parameters of temperature, pressure and concentration. The feature of the present invention is that the signal acquisition and processing process in the traditional wavelength modulation method is redistributed between the work of the host computer system and the slave computer system, and has the following advantages: the normalized high-order harmonic signal based on the wavelength modulation absorption spectrum signal is used to calculate the parameters such as gas temperature, concentration and pressure, which maintains the test characteristics of the traditional wavelength modulation method signal without calibration and improves its anti-noise ability; the parallel online computing capability of FPGA is fully utilized to extract the real-time high-order harmonic signal online, which greatly reduces the data upload amount and reduces the complexity of the system; and the fitting model of the neural network is used for training, so that the nonlinear relationship between the harmonic signal of the absorption spectrum and the gas parameter in a large range can be fully solved; the neural network model of appropriate scale can be customized according to the actual need of solving the error and the need of the test range, and has better adaptability and robustness. The present invention makes the wavelength modulation absorption spectrum method have a wider range of application, and the signal acquisition and extraction device is simple, and the network model can be flexibly adjusted according to the specific solution range and accuracy target, which has very practical significance and value for the promotion of the wavelength modulation method. Summary of the invention

[0009] 1. Technical issues to be resolved

[0010] The purpose of the present invention is to provide a parameter extraction method and system for wavelength modulated laser gas absorption spectrum, which are used to reduce the complexity of the real-time processing system of the wavelength modulation method and solve the problem of solving the gas parameters using high-order harmonic signals. The design based on FPGA realizes the online analysis of high-order harmonics of the wavelength modulation method signal and the extraction of normalized high-order harmonic peak points. At the same time, the neural network method uses the high-order harmonic components extracted by FPGA to solve the information such as gas temperature, concentration and pressure. While reducing the complexity of the system, it broadens the measurement range of the wavelength modulation method and has broad application prospects.

[0011] (II) Technical solution

[0012] The present invention, namely a method and system for extracting parameters of wavelength modulated laser gas absorption spectrum, comprises the following steps:

[0013] Step 1: Based on the measured laser wavelength intensity relationship, combined with the HITRAN database and the numerical simulation results of gas absorption, a training set of a neural network model is obtained with high-order harmonic peak points as input and gas concentration, temperature and pressure as output.

[0014] The wavelength modulated laser signal is generated by the distributed feedback laser controlled by the laser controller module. The modulation formula for the wavelength, i.e. the optical frequency, is as follows:

[0015]

[0016] in, is the center frequency of modulation, v is the modulated laser light wave frequency, ω is the fundamental frequency of the modulated harmonic, a k is the modulation amplitude of different multiple frequencies, where k can generally be between 1 and 5, and a k It is related to the selected absorption peak position, and is generally related to the absorption peak half-width Δv as follows:

[0017]

[0018] The modulated laser passes through the split fiber, and one beam is directly connected to the photodetector as a reference unabsorbed modulated laser signal V ref , a beam of wavelength change information v(t) is recorded by a wavelength recorder.

[0019] According to the HITRAN database, the gas absorption rate formula can be simulated and calculated as follows:

[0020] α(t)=exp(-P·l·C·S(T)·φ(v(t))) (3)

[0021] Where l[m] is the absorption path length, P[atm] is the pressure, T[K] is the temperature of the gas to be measured, C is the concentration of the gas to be measured, and φ(v(t)) is a linear function that satisfies the normalization condition: S(T)[cm -2 atm -1 ] is the line intensity of the absorption spectrum at temperature T. Line intensity function S(T)[cm -2 atm -1 ] is only related to temperature and can be expressed as:

[0022]

[0023] Where k is the Boltzmann constant, S(T0) is the line intensity at the reference temperature (usually room temperature, 296K), Q(T) and Q(T0) are the partition functions at temperature T and reference temperature T0, respectively, which can be obtained from the HITRAN database, h is Planck's constant, and c is the speed of light.

[0024] Take a set of temperature, concentration, and pressure combinations (T, P, C) within the range to be solved, and calculate its absorption rate α under the conditions of the recorded wavelength v according to the HITRAN database. From this, the measured laser modulation signal V ref According to formula (5), the absorption signal V of the gas to be tested at a certain temperature, concentration and pressure is simulated and calculated:

[0025] V=V ref α (5)

[0026] Design a matrix to extract the higher harmonic components of V0:

[0027]

[0028] Where N = 2πf s / ω,f s is the sampling frequency of the modulation signal, ω is the modulation base frequency of the laser modulation signal, M is the sampling length of the recorded signal, and n is the harmonic order of the signal.

[0029] The set of n-order harmonic components of the signal is calculated as follows:

[0030]

[0031] f (n) Represents the nth order harmonic component of the data set, and normalizes F to obtain

[0032]

[0033] Then, the maximum value of each column of F′ is extracted as the sample matrix F for inputting the neural network into this group of data. peak .

[0034] At the same time, within the temperature, pressure and concentration range to be solved, samples are taken according to the set density and then several sets of matrices F are obtained in turn according to the above steps. peak As the output of the neural network model, a training set of the neural network can be formed;

[0035] Initialize the neural network according to the complexity of the problem, set the network depth L, and select a four-layer neural network, that is, L = 3, m lis the number of nodes in the lth layer (l = 0, 1, 2, 3), m0 is the number of nodes in the input layer, m1 is the number of nodes in the first hidden layer, m2 is the number of nodes in the second hidden layer, and m3 is the number of nodes in the output layer; the weights W of each layer are randomly assigned l and bias β l ,in:

[0036]

[0037]

[0038] Select the activation function φ of the hidden layer of the neural network. For the commonly used Tan-Sigmoid function, it has the form of formula (8). Then, for the above input sample, each node H of the hidden layer of the neural network l The output is as follows:

[0039]

[0040]

[0041] Among them, when l = 1 is the relationship from the input layer to the first hidden layer, H0 = F peak ;

[0042] The output Y of each node in the output layer p (b) is:

[0043]

[0044] The error E(b) can be calculated according to the following formula:

[0045]

[0046] in, is the expected output at each output node, that is, the obtained gas parameter information such as temperature, concentration, pressure, etc. If the actual error does not meet the set error lower limit, the weight matrix W is updated according to the following formula l :

[0047] w l (a,b)=w l (a,b)+η·H′ l (b) H l-1 (b) (15)

[0048] Among them, H′ l (b) indicates H l (b), η represents the learning rate of the neural network. When the error between two consecutive calculations is less than the preset error parameter of the neural network, the training is stopped and the neural network structure and parameters are saved.

[0049] Step 3: Online harmonic peak point extraction and gas parameter calculation based on FPGA.

[0050] After the structure and parameters of the neural network model are determined, the neural network model structure and parameters are transplanted to the online harmonic analysis module based on FPGA. The online harmonic analysis module extracts the high-order harmonic peak points of the laser signal after being absorbed by the actual gas to be measured, inputs the neural network model, and calculates the gas parameters online and outputs them; the online harmonic analysis module designed based on FPGA includes a data acquisition module, a high-order harmonic reference signal generation module, an orthogonal demodulation module, a peak extraction module and a neural network calculation module. The data acquisition module is used to digitize the light intensity signal detected by the photodetector into V test , as described in step 2 above, equations (6)-(9) for V test Perform the same operation as V0, generate a synchronous reference signal through the high-order harmonic reference signal generation module, and then use the orthogonal demodulation module to solve the high-order harmonics and input them into the peak extraction module to obtain the peak point data F peak The neural network calculation module includes a parameter storage module, a fixed-point multiplier and an activation function module, wherein the parameter storage module is used to store the neuron weights w and bias β of the trained neural network. l ; The fixed-point multiplier performs the multiplication and addition calculation process required for the calculation process of a single neuron; and the activation function module uses a lookup table method to implement the activation function.

[0051] (III) Beneficial effects

[0052] The beneficial effects of the present invention are that the normalized high-order harmonic signal based on the wavelength modulation method signal is used to calculate parameters such as gas temperature, concentration and pressure. It has the anti-noise and calibration-free test characteristics of the traditional wavelength modulation method signal and fully utilizes the parallel online computing capability of FPGA to perform online real-time extraction of high-order harmonic signals, greatly reducing the amount of data uploaded and reducing the complexity of the system; training is performed using the fitting model of the neural network to realize online gas parameter solution based on the high-order harmonic signal; at the same time, a neural network model of appropriate size can be customized according to the actual solution error needs and test range needs, which has better robustness and a wider absorption rate adaptation range and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Attached Figure 1 :Structure diagram of gas parameter extraction method based on wavelength modulated laser gas absorption spectroscopy

[0054] Attached Figure 2 :Picture of experimental device for laser wavelength recording and online harmonic analysis

[0055] Attached Figure 3:Wavelength modulation method photoelectric signal and normalized harmonic signal

[0056] Attached Figure 4 :Online harmonic analyzer structure based on FPGA DETAILED DESCRIPTION

[0057] Reference Figure 1 This is a flow chart of a wavelength modulation signal solution method based on numerical simulation neural network training and FPGA online harmonic analysis. Figure 2 The following is a diagram of the experimental setup of the system. The operation steps combined with an example are as follows:

[0058] Step 1, the modulation signal is generated by the signal generator 101, and the laser controller 102 controls the laser 103 to emit a wavelength modulation signal near the gas absorption peak, one beam of which is divided into three beams by the optical fiber splitter 104, one beam passes through the gas area to be measured 105 and is detected by the photoelectric detector 106 and then input into the FPGA online harmonic analyzer 107 to extract the peak value of the high-order harmonics, one beam passes through the optical fiber 108 and enters the photoelectric detector 109 to obtain the relationship between the laser modulation amplitude and time, and the remaining beam passes through the interferometer 110 to connect to the photoelectric detector, and the relationship between the laser wavelength after wavelength modulation and time is recorded according to the interference principle.

[0059] Step 2: The oscilloscope 112 collects the data and uploads it to the computer 113. The computer 113 Figure 1 According to the steps introduced in , the gas absorption rate information under the set temperature, concentration and pressure conditions is retrieved from the HITRAN database based on the wavelength and time information, and the absorption result of the laser signal that is not absorbed by the gas to be measured is simulated and calculated, and then the normalized high-order harmonic component F is calculated. peak In this example, F is selected peak , respectively by the attached Figure 3 The high-order harmonic signal f normalized by the first harmonic described in (n) / f (1) The absorption peak near the absorption peak (n is an integer from 2 to 6). The actual sampling point here is the peak of the normalized higher harmonic between 3500 and 5000. The largest 50 points are selected. The laser modulation band selected is 7185.1-7186.6cm -1 The fundamental frequency of laser modulation is a sine wave of 100kHz, the sampling frequency is 20MHz, and a total of 7000 sampling points with uniform wave number are set. Among them, the pressure condition is set to 1atm, the mole fraction of water vapor concentration is set to 0 to 0.6, and a sampling point is set every 0.01. The temperature of water vapor is set to 298-397K, and a sampling point is set every 1K. A total of 6000 sampling points are set and constitute the training sample set of the neural network with the corresponding high-order harmonic components.

[0060] Step 3: In this example, 6000 sets of 50-dimensional data are input into the BP neural network. The input layer nodes of the neural network here are 50, the number of nodes in the first hidden layer is set to 100, the number of nodes in the second hidden layer is set to 50, and the number of nodes in the output layer is set to 2. The weight and threshold of each node are set to a random number uniformly distributed between -0.05 and 0.05 to complete the initialization of the neural network. Set the threshold of the total average error E to ε = 10 -6 , when E<ε, the neural network stops training.

[0061] Step 4: Figure 2 In the process, the photoelectric signal of the photoelectric detector 106 is transmitted to the FPGA online harmonic analyzer 107, and the online high-order harmonic calculation part is calculated by the attached Figure 3 The FPGA-based online harmonic analyzer shown in the figure is completed. The photoelectric signal is transmitted to the signal conditioning circuit 202 with high fidelity from the photoelectric signal interface 201. The signal is adjusted to match the signal of the analog-to-digital converter ADC 203. The digitized signal enters the orthogonal demodulation module 206, which is programmed and implemented on the FPGA chip together with the peak extraction module 204, the high-order harmonic reference signal generation module 205, and the peak upload module 207. The digitized photoelectric signal is calculated by the orthogonal demodulation module 206 according to the high-order harmonic orthogonal operator generated by the high-order harmonic reference signal generation module 205, and the peak harmonic signal after normalization is extracted by the peak extraction module 204. Then it is uploaded to the attached file through the peak upload module 207. Figure 2 In this example, the peak upload module is an Ethernet upload module based on the UDP protocol.

[0062] Step 5: Figure 1 In the process, when the estimated error of the neural network model is less than the set threshold, the training of the model is terminated. Figure 2 The FPGA-based online harmonic analyzer 107 shown in FIG. 1 inputs the peak point information of the normalized high-order harmonics into the FPGA-based neural network calculation module to solve the temperature and concentration information of the gas to be measured.

[0063] The above description of the present invention and its implementation methods is not limited thereto, and the accompanying drawings are only one of the implementation methods of the present invention. Without departing from the purpose of the present invention, any structure or embodiment similar to the technical solution without creative design shall fall within the protection scope of the present invention.

Claims

1. A wavelength modulated laser gas absorption spectrum parameter extraction system, characterized in that: The system consists of a wavelength recording module and an online harmonic analysis module, wherein the wavelength recording module consists of a signal generation module, a laser, an interferometer, a fiber splitting module, an optical fiber, and a signal acquisition module, and is used to modulate the laser wavelength emitted by the laser and determine the wavelength-intensity relationship of the luminescence; then, the laser signal of the modulated absorption is simulated by combining several groups of gas spectral absorption rate data, and each group of normalized high-order harmonic peak points and the corresponding gas temperature, concentration and pressure information are extracted to form a data set, and a neural network model with high-order harmonic peak points as input and gas concentration, temperature and pressure as output is trained, and the network model parameters are determined according to the training function error; after the parameters of the neural network model are determined, the neural network model structure and parameters are transplanted to the online harmonic analysis module based on FPGA, and the online harmonic analysis module extracts the high-order harmonic peak points of the laser signal after being absorbed by the actual gas to be measured, inputs the neural network model, and calculates the gas parameters online and outputs them; the specific implementation method is as follows: The wavelength modulated laser signal is generated by a sine signal generated by a signal generator 101, and a laser controller module 102 controls a distributed feedback laser 103 to generate a wavelength modulated laser signal. The modulated laser passes through a beam splitting optical fiber 104, and one beam passes through an optical fiber 108 and is directly connected to a photodetector 109, which is recorded as a reference unabsorbed modulated laser signal V by a high-speed data acquisition module 112. ref , a laser beam passes through the interferometer 110 and then enters the photoelectric detector 111, and the real-time wavelength information v is recorded by the signal acquisition module 112; Take a set of temperature, concentration, and pressure combinations (T, P, C) within the range to be solved, calculate its absorption rate α under the conditions of the recorded wavelength v according to the HITRAN database, and use the measured laser modulation signal V ref According to formula (1), the signal y after being absorbed by the gas to be measured at the set temperature, concentration and pressure is simulated and calculated: V=V ref ·α(1) Design a matrix to extract the higher harmonic components of V0: Where N = 2πf s / ω,f s is the sampling frequency of the modulation signal, ω is the modulation base frequency of the laser modulation signal, M is the recorded signal sampling length, and n is the harmonic order of the signal; the set of n-order harmonic components of the signal is calculated as: Among them, f (n) Represents the nth order harmonic component of the data set, and normalizes F to obtain Then extract the maximum value of each column of F′ as the matrix F for inputting this group of data into the neural network peak ; At the same time, within the temperature, pressure and concentration range to be solved, samples are taken according to the set density and then several sets of matrices F are obtained in turn according to the above steps. peak As the output of the neural network model, a training set of the neural network can be formed; Initialize the neural network according to the complexity of the problem, set the network depth L, and select a four-layer neural network, that is, L = 3, m l where m0 is the number of nodes in the first and second hidden layers (l=0, 1, 2, 3), m1 is the number of nodes in the first hidden layer, m2 is the number of nodes in the second hidden layer, and m3 is the number of nodes in the output layer. The weights of each layer are randomly assigned. W l and bias β l ,in: Select the activation function φ of the hidden layer of the neural network. For the commonly used Tan-Sigmoid function, it has the form of formula (8). Then, for the above input sample, each node H of the hidden layer of the neural network l The output is as follows: Among them, when l = 1 is the relationship from the input layer to the first hidden layer, H0 = F peak ; The output Y of each node in the output layer p (b) is: The error E(b) can be calculated according to the following formula: in, is the expected output at each output node, that is, the obtained temperature, concentration and pressure information; if the actual error does not meet the set error lower limit, the weight matrix W is updated according to the following formula l : w l (a,b)=w l (a,b)+η·H′ l (b)·H l-1 (b) (11) Among them, H' l (b represents H l (b is the derivative, η represents the learning rate of the neural network; until the error of two consecutive calculations is less than the preset error parameter of the neural network, the training is stopped and the neural network structure and parameters are saved.

2. A wavelength modulated laser gas absorption spectrum parameter extraction system according to claim 1, characterized in that: After the structure and parameters of the neural network model are determined, the structure and parameters of the neural network model are transplanted to the online harmonic analysis module based on FPGA. The online harmonic analysis module extracts the high-order harmonic peak points of the laser signal absorbed by the actual gas to be measured, inputs the neural network model, and calculates the gas parameters online and outputs them. The online harmonic analysis module based on FPGA includes a data acquisition module, a high-order harmonic reference signal generation module, an orthogonal demodulation module, a peak extraction module and a neural network calculation module. The data acquisition module is used to digitize the light intensity signal detected by the photodetector into V test The high-order harmonic reference signal generation module generates a synchronous reference signal, and then the orthogonal demodulation module solves the high-order harmonics and inputs the peak extraction module to obtain the peak point data F peak ; The neural network calculation module includes a parameter storage module, a fixed-point multiplier and an activation function module; the parameter storage module is used to store the neuron weights w and bias β of the trained neural network. l ; The fixed-point multiplier and adder performs the multiplication and addition calculation process required for the calculation process of a single neuron; The activation function module uses a lookup table method to implement the activation function.

Images