1. Optical frequency comb gas detection system based on neural network, its characterized in that includes: the device comprises an optical frequency comb, an air chamber, a photodiode, a data acquisition card and a detection unit; the gas chamber is connected with the photodiode through an optical fiber, and the data acquisition card acquires an electric signal transmitted by the photodiode and sends the electric signal to the detection unit; the detection unit calculates and processes the received electric signals to measure the types and the concentrations of the gases; the air chamber comprises a reference detection air chamber and a reference air chamber.

2. The optical frequency comb gas detection system based on the neural network as claimed in claim 1, wherein two ends of the gas chamber are respectively provided with a reflector with opposite mirror surfaces, and the reflector is a spherical mirror with 500mm curvature radius.

3. The neural network-based optical frequency comb gas detection system as claimed in claim 2, wherein the spherical mirror is a spherical mirror array composed of a plurality of small spherical mirrors.

4. The neural network-based optical frequency comb gas detection system as claimed in claim 3, wherein the small spherical mirror is fixed on an angle adjustment bracket.

5. The optical frequency comb gas detection system based on the neural network as claimed in claim 4, wherein the angle adjustment bracket comprises a base plate and mirror supports distributed on the base plate in an array manner, and the mirror supports are hinged with the base plate; the small spherical mirror is embedded on the mirror support.

6. The neural network-based optical frequency comb gas detection system as claimed in claim 1, wherein the detection unit comprises a modem unit, a calculation unit, a gas identification unit and a concentration determination unit; the modulation and demodulation unit is used for demodulating an input electric signal to obtain a second harmonic of the gas to be detected; the calculating unit processes the second harmonic and the frequency of the gas to be detected into an input characteristic vector; the gas identification unit obtains the category of the gas to be detected through an artificial neural network model; the concentration measuring unit is used for calculating the concentration of the gas to be measured.

7. An optical frequency comb gas detection method based on a neural network comprises the following steps:

constructing an artificial neural network model for identifying gas categories;

introducing gas to be detected into a detection gas chamber, and determining the category of the gas to be detected by using an artificial neural network model;

introducing gas to be detected with the determined category and the concentration of 1% into a reference gas chamber, and determining the concentration x of the gas to be detected according to the following formula:

in the formula: p is a radical of_t,l_t,p_r,l_r,x_rRespectively measuring the pressure intensity of the air chamber, measuring the optical path length of the air chamber, the pressure intensity of a reference air chamber, the optical path length of the reference air chamber and the gas concentration of the reference air chamber; r_SThe ratio of the normalized second harmonic signal intensity of the measurement gas chamber and the normalized second harmonic signal intensity of the reference gas chamber is obtained.

8. The optical frequency comb gas detection method based on the neural network as claimed in claim 7, wherein in the construction of the artificial neural network model, the method for acquiring the training sample comprises the following steps:

sequentially introducing M kinds of trace gases into a detection gas chamber for data acquisition; a category label Y is set for each gas_mM is 1, 2, … …, M; for each gas, the data acquisition card transmits n groups of data of the gas to the detection unit, and the detection unit transmits each group of data: frequency X₁Second harmonic X₂Using normalization formula to carry out feature scaling processing and adding class label Y_m(ii) a Then get the tagged feature vector:i∈(1，n)，m∈(1，M)；

for datai belongs to (1, n), M belongs to (1, M) to carry out high latitude mapping, and then a new feature vector is obtainedi belongs to (1, n), M belongs to (1, M), and the new characteristic vector data set is a training sample of the artificial neural network.

9. The method according to claim 8, wherein the high-order map is processed using a kernel function as follows:

10. the method of claim 8, wherein the activation function of the artificial neural network is a ReLu activation function.

Background

With the rapid development of national economy, the continuous progress of industrial technology leads to increasingly serious environmental pollution problems such as air pollution, soil pollution and the like. Especially, the environmental problems of global climate warming and the like caused by air pollution, industrial waste gas, domestic waste gas, leaked toxic gas, automobile exhaust and the like, and the diseases of human respiratory system infection, oxidation damage of heart, cerebral vessels, blood vessel skin cells and the like are caused. Under the large background, the novel laser spectrum gas detection technology becomes a new research hotspot for environmental monitoring application.

Optical Frequency Comb (Optical Frequency Comb) technology arose in the background described above. The technology is a high-precision high-resolution gas detection technology by virtue of the advantages of wide frequency spectrum, narrow pulse width, high frequency stability, traceability of time domain and frequency domain characteristics to microwave frequency reference and the like. Optical frequency combing has been implemented at home and abroad in various ways, such as optical frequency comb cavity ring-down spectroscopy, optical frequency comb cavity enhancement spectroscopy, and double optical frequency comb multi-heterodyne spectroscopy. The foreign subject group applies the cavity ring-down optical frequency comb to the gas absorption measurement, so that the sensitivity of a single spectral line is improved to 7 multiplied by 10^-13cm^-1And the absorption spectrum detection of industrial waste gas and atmospheric greenhouse gas is realized, and a solution is provided for monitoring the air purification grade. The measurement precision is improved by 10 times by utilizing the double-optical frequency comb multi-heterodyne spectrum technology, and the trace component of the exhaled gas of the patient is measured. Currently, developed countries in europe and america are at the international leading level in the research of optical frequency comb technology. And the whole countryThe research level is relatively laggard, the related research in foreign countries is mainly tracked and expanded, and the gap is very obvious particularly in the aspect of the application of optical frequency comb gas detection.

However, the current optical frequency comb gas detection technology is realized by: knowing the components of the target gas, selecting the characteristic spectral line of the target gas, and designing a laser light source corresponding to the frequency of the characteristic spectral line corresponding to the selected characteristic spectral line. Namely: the gas measurement is carried out on the premise of knowing the gas components, and the identification and effective measurement of unknown gas cannot be carried out. Gas identification research based on optical frequency combs is not available at home and abroad. The artificial neural network can provide a powerful solution algorithm, and can automatically generate a black box function by repeatedly training and learning the mapping relation among a large number of input and output, so that gas identification is realized without establishing an equation expression of gas response. Based on the method, a dual-optical frequency comb system based on artificial neural network gas identification and concentration detection optimization is provided.

Disclosure of Invention

The invention aims to provide an optical frequency comb spectrum gas detection system and method based on artificial neural network gas identification, and the system and method have the advantages of strong nonlinear mapping capability, high training speed, good self-learning capability, simple structure, avoidance of overfitting and the like, can optimize the neural network training process, simplifies various parameters involved in gas concentration calculation, and enables the optical frequency comb spectrum gas detection system to have higher gas identification capability and higher measurement accuracy.

In order to achieve the purpose, the invention adopts the technical scheme that: an optical frequency comb gas detection system based on a neural network, comprising: the device comprises an optical frequency comb, an air chamber, a photodiode, a data acquisition card and a detection unit; the gas chamber is connected with the photodiode through an optical fiber, and the data acquisition card acquires an electric signal transmitted by the photodiode and sends the electric signal to the detection unit; the detection unit calculates and processes the received electric signals to measure the types and the concentrations of the gases; the air chamber comprises a reference detection air chamber and a reference air chamber.

Furthermore, the two ends of the air chamber are respectively provided with a reflector with a mirror surface opposite to each other, and the reflector is a spherical mirror with a curvature radius of 500 mm.

Furthermore, the spherical mirror is a spherical mirror array formed by a plurality of small spherical mirrors.

Further, the small spherical mirror is fixed on the angle adjusting bracket.

Furthermore, the angle adjusting bracket comprises a bottom plate and mirror supports distributed on the bottom plate in an array manner, and the mirror supports are hinged with the bottom plate; the small spherical mirror is embedded on the mirror support.

In a preferred embodiment of the present invention, the detection means includes a modem means, a calculation means, a gas identification means, and a concentration measurement means; the modulation and demodulation unit is used for demodulating an input electric signal to obtain a second harmonic of the gas to be detected; the calculating unit processes the second harmonic and the frequency of the gas to be detected into an input characteristic vector; the gas identification unit obtains the category of the gas to be detected through an artificial neural network model; the concentration measuring unit is used for calculating the concentration of the gas to be measured.

The invention also provides an optical frequency comb gas detection method based on the neural network for realizing the purpose, which comprises the following steps:

1. constructing an artificial neural network model for identifying gas categories;

2. introducing gas to be detected into a detection gas chamber, and determining the category of the gas to be detected by using an artificial neural network model;

3. the method comprises the following steps of introducing 1% of gas to be detected with determined category into a reference gas chamber, and calculating the concentration x of the gas to be detected by a concentration measuring unit according to the following formula:

in the formula: p is a radical of_t，l_t，p_r，l_r，x_rRespectively measuring pressure intensity of the gas chamber, measuring optical path of the gas chamber, pressure intensity of the reference gas chamber, optical path length of the reference gas chamber and gas of the reference gas chamberConcentration; r_SIn order to measure the ratio of the normalized second harmonic signal intensity of the air chamber to the normalized second harmonic signal intensity of the reference air chamber, the normalized second harmonic is equal to the ratio of the second harmonic signal intensity to the first harmonic signal intensity, and the second harmonic signal and the first harmonic signal are obtained by system measurement.

Further, in the construction of the artificial neural network model, the method for obtaining the training sample comprises the following steps:

sequentially introducing M kinds of trace gases into a detection gas chamber for data acquisition; a category label Y is set for each gas_m1, 2.. said, M, denotes the mth gas of the M gases; for each gas, the data acquisition card transmits n groups of data of the gas to the detection unit, and the detection unit transmits each group of data: frequency X₁Second harmonic X₂Using normalization formula to carry out feature scaling processing and adding class label Y_m(ii) a Then get the tagged feature vector:

for dataHigh latitude mapping is carried out, and then a new feature vector is obtainedThe new feature vector data set is a training sample of the artificial neural network.

Further, the high-order mapping is processed by using the following kernel function:

in a preferred embodiment of the present invention, the activation function of the artificial neural network is a ReLu activation function.

Compared with the prior art, the invention has the beneficial effects that: the invention provides a double-optical frequency comb gas detection system based on a neural network, which is provided with double gas chambers, wherein spherical mirrors with the curvature radius of 500mm are arranged at two ends of each gas chamber, each spherical mirror comprises a plurality of small spherical mirrors to form an array, so that an incident light beam can be reflected for multiple times in the gas chambers, the effective optical path is increased, the absorption degree of the gas to the incident light is effectively increased due to low trace gas concentration and weak absorption, and an error correction mechanism of concentration calculation is provided due to the arrangement of a reference gas chamber, the measurement error of the system is effectively eliminated, and the accuracy of a concentration measurement result is improved. According to the method, after the trace gas components are determined by adopting the neural network model, the concentration of the specific gas is calculated, meanwhile, the reference concentration is introduced into the concentration calculation, the system error is eliminated, and finally the accurate gas components and concentration of the gas to be measured are obtained. The method has the advantages of strong nonlinear mapping capability, high training speed and the like.

Drawings

FIG. 1 is a structural connection diagram of an optical frequency comb gas detection system based on a neural network provided by the invention;

FIG. 2 is a side view of the plenum;

FIG. 3 is a schematic view of the mirror of FIG. 2;

FIG. 4 is a side view of the angle adjustment bracket;

FIG. 5 is a flow chart of a neural network based optical frequency comb gas detection method provided by the present invention;

fig. 6 is a schematic structural diagram of an artificial neural network.

Detailed Description

In order to facilitate an understanding of the invention, the invention is described in more detail below with reference to the accompanying drawings and specific examples. Preferred embodiments of the present invention are shown in the drawings. This invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

In order to solve the problem of gas detection in the prior art, the present embodiment provides an optical frequency comb gas detection system based on a neural network, as shown in fig. 1, the system mainly includes: the device comprises a laser controller 1, an optical frequency comb 2, a detection gas chamber 3, a reference gas chamber 4, an InGaAs avalanche photodiode 5, a data acquisition card 6 and a detection unit 7. The laser controller 1 is connected with the optical frequency comb 2 to generate a scanning sawtooth wave signal, and controls the laser pulse and the temperature of the optical frequency comb through current. The optical frequency comb 2 is connected with the detection gas chamber 3 and the reference gas chamber 4 through an 99/1 optical fiber beam splitter 8, the 99/1 optical fiber beam splitter 8 divides a laser beam generated by the optical frequency comb 2 into two beams, wherein one beam (99% energy) enters the detection gas chamber 3 filled with gas to be detected, and the other beam (1% energy) enters the reference gas chamber 4 filled with the gas to be detected with the calibrated concentration of 1%.

The absorption characteristic spectral lines of various gases are different, and because the optical frequency comb has the advantages of wide spectrum, high precision and the like, in the embodiment, the optical frequency comb of the erbium-doped optical fiber system is selected as a laser light source.

As shown in fig. 2 and 3, the detection gas cell 3 and the reference gas cell 4 have the same structure, and two opposite ends of the gas cells are respectively provided with a mirror 30. The mirrors 30 at each end are formed of dimensions 50 x 50mm²An array of spherical mirrors. The spherical mirror array is composed of 64 pieces of 6.25mm by 6.25mm²A square spherical mirror 34 of the size. The square spherical mirror 34 is embedded on the mirror support 33, the mirror support 33 is connected with the bottom plate 36 through the universal hinge 35, and the mirror support 33, the universal hinge 35 and the bottom plate 36 form an angle adjusting support, as shown in fig. 4.

The 64 square spherical mirrors 34 are fixed on an angle adjusting bracket, and the reflection angle of each square spherical mirror can be changed through manual adjustment. Each square spherical mirror can be independently adjusted in the horizontal and vertical directions for beam alignment.

At one end of the gas cell there is an entrance port 31 and at the other end there is an exit port 32, the entrance port 31 and the exit port 32 being located in the middle of the mirror.

The laser generated by the optical frequency comb 2 enters the detection air chamber 3 and the reference air chamber 4 through the optical fiber from the incident port 31, the reflection times of the laser in the air chamber can be changed by adjusting the angles of the square spherical mirrors at the two ends of the air chamber, and the laser is output through the exit port 32 after being reflected for multiple times. Generally, the trace gas has low concentration, so that the trace gas is absorbed weakly in a common gas chamber, and the subsequent concentration detection is not facilitated. By adopting the air chamber provided by the embodiment, the laser is reflected for multiple times in the air chamber, so that the absorption degree of the gas to the laser can be effectively increased, and the subsequent concentration detection and calculation are more facilitated. The mirror coating has a high reflection (> 99.98%) in the near infrared range [1550nm, 1825nm ].

The exit ports 32 of the detection gas chamber 3 and the reference gas chamber 4 are respectively connected with the InGaAs avalanche photodiode 5 through optical fibers. The InGaAs avalanche photodiode 5 converts the reflected light output from the gas chamber into an electrical signal and transmits the electrical signal to the data acquisition card 6, and the data acquisition card 6 transmits the acquired signal to the gas detection unit 7.

As shown in fig. 1, the detection unit 7 includes 4 subunits, which are a modem unit, a calculation unit, a gas identification unit, and a concentration measurement unit, respectively. The modulation and demodulation unit is used for modulating a laser emergent signal of the optical frequency comb to modulate a high frequency band; meanwhile, the device is also used for demodulating and low-pass filtering the input electric signal to obtain a second harmonic signal of the gas to be measured. The calculating unit is used for carrying out normalization processing on the second harmonic and the frequency of the gas to be detected to form an input characteristic vector; the gas identification unit is an artificial neural network model and identifies the type of the gas to be detected by inputting the characteristic vector. The concentration measuring unit is stored with a formula for calculating the gas concentration, and the formula is called to calculate the concentration of the gas to be measured through the detection data transmitted by the acquisition card.

The present invention provides another embodiment, which is a method for detecting an optical frequency comb gas based on a neural network, and the flow of the method is shown in fig. 5, wherein nitrous oxide (N) is used₂The detection of four trace gases of O), Nitric Oxide (NO), methane (CH4) and carbon dioxide (CO2) is taken as an example, and the specific steps are as follows:

1. setting an artificial neural network structure

According to the four gases to be trained: nitrous oxide (N)₂O), Nitric Oxide (NO), methane (CH4), and carbon dioxide (CO2), the number of neurons in the output layer is set to 4, the number of input features is set to 2, and the input layer bias means is added to set the number of neurons in the input layer to 2Is 1. Setting the number of hidden layers to 1 layer, setting the number of hidden layer neurons to 4, adding a hidden layer bias unit to set to 1, and completing the setting of the neural network structure, as shown in fig. 6.

2. Training of neural networks

The laser controller and the optical frequency comb are opened, the current generated by the laser controller is used for generating scanning sawtooth waves and controlling the temperature of the optical frequency comb, the modulation and demodulation unit generates modulation signals to be added to laser output signals, the output signals are modulated to be in a high frequency range and are divided into two beams through an 99/1 optical fiber beam splitter, and the first beam (99% of energy) passes through a detection air chamber filled with gas to be detected.

For the four gases to be trained: nitrous oxide (N)₂O), Nitric Oxide (NO), methane (CH4) and carbon dioxide (CO2), wherein the four gases with the concentration of 10% are sequentially introduced into the detection gas chamber, the temperature of the gas chamber is controlled, and the pressure of the gas chamber is controlled by controlling the flow rate of the gases. The laser is reflected in the air chamber for multiple times and then is emitted from the exit, and is transmitted into the InGaAs avalanche photodiode through the optical fiber, and the signal is subjected to photoelectric conversion and then is transmitted to the data acquisition card.

3. For each gas to be trained, the data acquisition card transmits the data of the gas to the detection unit, and the modulation and demodulation unit demodulates and low-pass filters the data to obtain 500 groups of data points of second harmonic signal intensity and frequency of each gas. The calculation unit calculates the data: frequency X₁Intensity of second harmonic X₂Feature scaling is performed using a normalization formula:

wherein, x and x_max、x_minInput values, maximum values and minimum values of the features, respectively.

Add class label Y to each set of data for each gas_mAnd m is equal to 1 and 4. Then obtaining the feature vector of the normalized labeled training sampleWhereinRespectively, the ith data of the mth gas.

Y_mDenotes the m-th class gas label:

N₂O：Y₁＝(1，0，0，0)；NO：Y₂＝(0，1，0，0)；

CH₄：Y₃＝(0，0，1，0)；CO₂：Y₄＝(0，0，0，1)。

then 2000 feature vector datasets with gas category labels are obtained.

4. And performing high latitude mapping on the data, and performing data processing by using the following kernel function, namely:

then the following results are obtained:

K¹＝K((X₁,X₂)¹,(X₁,X₂)^j)

K²＝K((X₁,X₂)²,(X₁,X₂)^j)

…

Kⁿ＝K((X₁,X₂)ⁿ,(X₁，X₂)^j)，j∈(1,n)。

a new feature vector is obtainedA new data set is obtained:

the purpose of using the Gaussian kernel function is to map the features to high latitude, so that the features can be linearly separable at the high latitude under the condition that the current dimension hyperplane is linearly separable, and the embarrassment that low latitude cannot be effectively decoupled when gas absorption spectral lines are aliased is solved.

5. Initializing the weighting coefficients, i.e. the weighting values of each layer, using He initialization Where N represents a random floating point value from 0 to the number of neurons of layer (l +1) of the neural network. Unified initialization to 0 for bias facilitates computation.

6. Inputting the new feature vector data obtained in the step 4 into a neural network, and performing forward and backward propagation on the neural network to obtain a backward propagation error iterative formula:

δ⁽³⁾＝(a⁽³⁾-Y).*g′(z⁽³⁾)

in the formula, delta^(l)Denotes the propagation error, z, of the l, l ∈ (2,3) layer^(l)＝θ^(l)a^(l-1)+b^(l)Matrix of state values, theta, representing the l-th layer^(l)Representing the weight matrix from layer (l-1) to layer (l), a^(l)A matrix of activation values representing the l-th layer, a^(l)＝g(z^(l)). Denotes dot multiplication, i.e. matrix element multiplication.

Wherein g' (z)^(l)) Is g (z)^(l))＝max(0,z^(l)) The derivative of (2) has no saturation region by using the ReLu activation function, has no problem of gradient disappearance, has higher actual convergence rate, and is more in line with a biological nerve activation mechanism.

7. The partial derivatives of the cost function versus weight and bias terms for each training data were found using the FP and BP algorithms:

the parameters are updated according to the following formula:

θ^(l)＝θ^(l)-αD^(l)

b^(l)＝b^(l)-αbD^(l)

and (4) when the cost function meets the precision error, performing the next step, otherwise, checking whether the iteration times is greater than the maximum iteration times, if not, returning to the step 6, and if so, indicating that the network training is finished, and performing the next step.

A random gradient descent method is used, namely, each time one training data is used for gradient descent to update the weight, so that the iteration time is effectively accelerated, and the network training is faster.

8. The method comprises the steps of firstly opening a detection gas chamber, introducing trace gas to be detected into the detection gas chamber, normalizing the second harmonic signal intensity and frequency of the gas to be detected by a calculation unit to form an input characteristic vector, inputting the input characteristic vector into an artificial neural network model of a gas identification unit, and identifying unknown gas to obtain a classification result of the gas to be detected.

9. Then, opening the reference gas chamber, introducing the gas with the determined category and the concentration of 1% into the reference gas chamber, and calculating the concentration x of the gas to be measured by the concentration measuring unit according to the following formula:

in the formula: p is a radical of_t，l_t，p_r,l_r,x_rRespectively measuring the pressure intensity of the air chamber, the optical path of the air chamber, the pressure intensity of the reference air chamber, the optical path length of the reference air chamber and the gas concentration of the reference air chamber; r_STo measure the ratio of the normalized second harmonic signal strength of the air cell to the normalized second harmonic signal strength of the reference air cell, the normalized second harmonic is equal to the ratio of the second harmonic signal strength to the first harmonic signal strength.

The derivation process of the concentration calculation formula is as follows:

the first harmonic signal intensity s of the measurement air chamber is obtained by a modulation and demodulation unit_1fSecond harmonic signal intensity s_2fAnd the first harmonic signal intensity s of the reference gas cell_1frSecond harmonic signal intensity s_2fr。

The ratio of the normalized second harmonic signal intensity of the measurement air chamber to the normalized second harmonic signal intensity of the reference air chamber is:

according to the first harmonic and second harmonic expressions:

wherein: x is the concentration of the gas to be measured in the measuring gas chamber, x_rTo be referenced to the concentration of the respective gas in the gas chamber, I_0t、I_0rThe emergent laser light intensity of the measurement air chamber and the reference air chamber is respectively.

Then the 2f signal after 1f normalization is:

the ratio of the second harmonic signal intensity of the measurement air chamber to the reference air chamber is:

therefore, the calculation formula for measuring the concentration of the gas to be measured in the gas chamber is as follows:

10. so far, the species identification and concentration determination of the trace gas to be detected are completed.