A convolutional neural network-based optical fiber inner cavity gas detection system

Abstract

This invention discloses a fiber optic cavity gas detection system based on a convolutional neural network, comprising a pump source, a wavelength division multiplexer, a thulium-doped fiber, an isolator, a tunable filter, a test gas chamber, a fiber optic circulator, a photodetector, a data acquisition card, and an industrial control computer. The thulium-doped fiber can generate amplified spontaneous emission light in the range of 1750 nm to 2100 nm. The test gas chamber is used to hold at least two mixed gases to be tested, and its length is less than 20 cm. The industrial control computer receives data from the data acquisition card, performs photoelectric demodulation on the interference signal, and uses a pre-trained gas classification and concentration model to classify and detect the concentration of the mixed gas in the test gas chamber. This invention can classify and detect the concentration of five gases: ammonia, hydrogen sulfide, methane, carbon dioxide, and hydrogen bromide.

Description

Technical Field

[0001] This invention belongs to the field of fiber optic sensing, specifically relating to a fiber optic cavity gas detection system based on a convolutional neural network. Background Technology

[0002] Industrial fumes and harmful gases in daily life can endanger the environment and human safety. Intracavity absorption spectroscopy based on fiber lasers offers advantages such as high stability, high resolution, and real-time monitoring, making it a crucial technology in fiber optic gas sensing. Fiber optic gas sensing systems can identify the type of gas being measured and monitor its concentration, meeting the needs of protecting industrial production and human safety.

[0003] Fiber optic gas sensing technology is based on spectral absorption theory. Traditional gas sensing systems employ a linear structure, using a laser as the light source and detecting the gas absorption spectrum after single-pass transmission or multiple reflections within a gas chamber. In their 2009 paper, "Gas Detection Method Based on Ring Cavity Fiber Laser," Jia Dagong et al. proposed a gas detection system based on the internal cavity design of a fiber laser. This system uses wavelength scanning technology to ensure the fiber laser output wavelength covers multiple gas absorption peaks, and the corresponding gas concentration is obtained by averaging the concentrations from these multiple absorption peaks. However, spectral absorption based on wavelength scanning technology suffers from low sensitivity. Multiple reflections require large-volume or more complex gas chamber structures. In contrast, gas sensing systems using ring cavity laser structures increase the effective absorption optical path by repeatedly reflecting the laser within a finite-length resonant cavity, thereby improving gas sensing sensitivity. This allows for both a small, simple gas chamber structure and a large effective absorption optical path. The cavity absorption sensing system uses thulium-doped fiber as the gain medium to form a fiber amplifier, which can generate broadband laser light that can cover the absorption peaks of various gases. Therefore, the cavity gas sensing system has the characteristics of high-precision analysis and rapid detection of various gases.

[0004] Convolutional Neural Networks (CNNs) are a type of deep learning model that can be used to compensate for the effects of experimental temperature and system fluctuations on gas concentration. Therefore, multi-gas detection in fiber optic cavities based on CNNs has profound significance for practical engineering applications. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and propose a fiber optic cavity gas detection system based on convolutional neural networks. By exciting thulium-doped fiber to generate a broadband light source covering the absorption peaks of various gases, the convolutional neural network compensates for the effects caused by light source fluctuations, changes in sensor end-face coupling efficiency, and experimental test temperature, thereby realizing a gas classification and concentration detection system with more types of gases and higher accuracy.

[0006] This invention relates to a fiber optic cavity gas detection system based on a convolutional neural network, comprising a pump source, a wavelength division multiplexer, a thulium-doped fiber, an isolator, a tunable filter, a test gas chamber, a fiber optic circulator, a photodetector, a data acquisition card, and an industrial control computer.

[0007] The pump source is a pump source with a center wavelength around 1570nm;

[0008] The thulium-doped fiber can generate amplified spontaneous emission light from 1750 nm to 2100 nm.

[0009] The test chamber is used to hold at least two mixed gases to be tested, and the length of the test chamber is less than 20 cm.

[0010] The output end of the fiber optic circulator with a splitting ratio of 10% is connected to the other input end of the wavelength division multiplexer, and the output end with a splitting ratio of 90% is connected to the input end of the photodetector.

[0011] The output terminal of the photodetector is connected to the input terminal of the data acquisition card, and the output terminal of the data acquisition card is connected to the industrial control computer.

[0012] The industrial control computer receives data from the data acquisition card, performs photoelectric demodulation on the interference signal, and uses a pre-trained gas classification and concentration model to classify and detect the concentration of the mixed gas in the test chamber.

[0013] Furthermore, the mixed gas in the test chamber is selected from at least two of the following gases: ammonia, hydrogen sulfide, methane, carbon dioxide, and hydrogen bromide.

[0014] Furthermore, the photoelectric demodulation step includes:

[0015] Step 1: The industrial control computer performs preliminary processing on the received electrical signal, extracts the absorption spectrum signal based on the absorption wavelength of the test gas, and after extracting the spectral lines, uses a low-pass filter to remove high-frequency noise, and uses Haar wavelet transform and empirical mode decomposition to denoise the original absorption spectral lines, thereby obtaining the absorption spectrum signal.

[0016] Step 2: Further process the noise-reduced absorption spectrum signal obtained in Step 1, obtain the spectral distribution of each component gas in the mixed gas through multi-peak fitting, separate the absorption lines, obtain the separated spectral signals, further reduce the influence of overlapping gas absorption lines, and improve the measurement accuracy of the mixed gas.

[0017] Step 3: Input the spectral signal processed by multi-peak fitting and spectral line separation in Step 2 into the gas classification and concentration model based on convolutional neural network for feature extraction and recognition, thereby outputting the type of mixed gas and the concentration of each gas; wherein the gas classification and concentration model is pre-trained on gas samples using convolutional neural network.

[0018] Furthermore, in step one, methane gas is used to extract the absorption line in the 1965 nm band, ammonia gas is used to extract the absorption line in the 1959 nm band, hydrogen sulfide gas is used to extract the absorption line in the 1945 nm band, and hydrogen bromide gas is used to extract the absorption line in the 1970 nm band. In order to reduce the influence of overlapping absorption lines, carbon dioxide gas is used to extract the absorption line in the 2004 nm band.

[0019] Compared with the prior art, the beneficial effects of the technical solution of the present invention are as follows:

[0020] 1. This invention utilizes a constructed cavity-type gas sensing system to extract laser light absorbed by the test gas, and then inputs the processed signal into a gas classification and concentration model based on a convolutional neural network for calculation, which can classify and detect the concentration of five gases, namely ammonia, hydrogen sulfide, methane, carbon dioxide, and hydrogen bromide.

[0021] 2. Since the absorption lines of the detected gases may overlap, affecting the detection accuracy, a wavelength with less impact is selected for absorption line detection. The 2004nm band, where CO2 absorption is more pronounced, is chosen to collect the absorption lines. At the same time, this invention performs noise reduction and multi-peak fitting processing on the spectrum before detection to improve the measurement accuracy of mixed gases.

[0022] 3. This invention requires only one light source to detect five gases, including ammonia, hydrogen sulfide, methane, carbon dioxide, and hydrogen bromide. It can achieve high-precision detection with a test chamber length not exceeding 20cm. The system has a simple structure, which simplifies the optical path design, saves hardware design costs, and enables real-time gas detection and sensing. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the fiber optic cavity gas detection system described in this invention;

[0024] Figure 2 This is a schematic diagram of the signal absorption processing flow.

[0025] Figure 3 This is a flowchart of the fiber optic cavity gas detection system described in this invention.

[0026] In the picture:

[0027] 1: Pump light source; 2: Wavelength division multiplexer; 3: Thulium-doped fiber

[0028] 4: Isolator; 5: Tunable filter; 6: Test gas chamber

[0029] 7: Fiber optic circulator; 8: Photodetector; 9: Data acquisition card

[0030] 10: Industrial PC Detailed Implementation

[0031] To make the objectives, technical solutions, beneficial effects, and significant advancements of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings provided in the embodiments of the present invention. Obviously, all the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "fixing," "connection," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, a movable connection, or even an integral part; they can refer to a direct connection, an indirect connection through an intermediate medium, or an invisible signal connection; they can refer to the internal communication of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0033] It should be noted that the terms "first," "second," "third," etc., in the specification and claims of this invention are only used to distinguish different objects, and not to describe a specific order.

[0034] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described are only for explanation and illustration of the present invention and are not intended to limit the present invention.

[0035] like Figure 1As shown, an optical fiber cavity gas detection system based on a convolutional neural network includes a pump light source 1, a wavelength division multiplexer 2, a thulium-doped fiber 3, an isolator 4, a tunable filter 5, a test gas chamber 6, an optical fiber circulator 7, a photodetector 8, a data acquisition card DAQ9, and an industrial control computer IPC10. The output of the pump light source 1 is connected to the input of the wavelength division multiplexer 2. The recombining end of the wavelength division multiplexer 2 is connected to one end of the thulium-doped fiber 3. The other end of the thulium-doped fiber 3 is connected to the input of the isolator 4. The output of the isolator 4 is connected to the input of the tunable filter 5. The output of the tunable filter 5 is connected to the input of the test gas chamber 6. The output of the test gas chamber 6 is connected to the input of the 10 / 90 fiber optic circulator 7. The output of the fiber optic circulator 7 with a splitting ratio of 10% is connected to the other input of the wavelength division multiplexer 2. The output of the fiber optic circulator 7 with a splitting ratio of 90% is connected to the input of the photodetector 8. The output of the photodetector 8 is connected to the input of the data acquisition card 9. The output of the data acquisition card 9 is connected to the industrial control computer 10.

[0036] The specific working principle is as follows: The pump light source 1 is a pump light source with a center wavelength around 1570nm. The wavelength division multiplexer 2 adds the pump light into the resonant cavity. The pump light enters the thulium-doped fiber 3 to generate amplified spontaneous emission light from 1750nm to 2100nm, which can cover the absorption peaks of various gases such as ammonia (1959nm), hydrogen sulfide (1945nm), methane (1965nm), carbon dioxide (2004nm), and hydrogen bromide (1970nm), allowing for the detection of a wide variety of gases. After passing through the isolator 4 to ensure the unidirectionality of the transmitted light, and after passing through the tunable filter 5 to select the frequency and narrow the linewidth, the gas absorption spectrum can be obtained by tuning and scanning through the test gas chamber 6. The test gas chamber contains a mixed gas to be tested, which includes ammonia, methane, carbon dioxide, and hydrogen sulfide with concentrations of 0.01%-3% and 0.01%-3%, respectively. The gas chamber contains 0.01%-3% hydrogen bromide gas. The measuring chamber is a hollow cylinder with an outer diameter of 5 cm and a height of 20 cm. The fiber optic circulator 7 splits the input light in a 1:9 ratio. The 10% output port outputs signal light, which is connected to the photodetector 8. The 90% output port is connected to the input port of the wavelength division multiplexer 1, so that most of the light is used as input light to re-enter the ring cavity and continue to participate in the optical amplification process in the resonant cavity. The light is continuously excited and amplified in the ring cavity. Finally, when the system stabilizes, the laser is output. Since the light repeatedly passes through the gas to be measured, the effective absorption optical path is greatly increased, thereby improving the gas sensing sensitivity. The last part of the light enters the photodetector 8 through the coupler, thereby obtaining the laser absorption spectrum data after the gas action. The photodetector 8 converts the signal light into an electrical signal and outputs it to the data acquisition card 9. The data acquisition card 9 collects the sensing electrical signal transmitted back from the photodetector 8 and outputs it to the industrial control computer 10 for signal processing.

[0037] The industrial control computer receives data from the data acquisition card 9, performs photoelectric demodulation on the interference signal, and thereby classifies and detects the concentration of the mixed gas in the test gas chamber 6.

[0038] The fiber optic cavity gas detection system based on convolutional neural networks described in this invention utilizes fiber optic laser cavity absorption spectroscopy technology, which features high selectivity, spectral resolution, and detection sensitivity. The gas sensing technology based on fiber optic ring cavity lasers is suitable for high-precision gas detection. When the gas to be tested is placed in the test gas chamber 6, it will pass through the mixed gas inside multiple times during the laser formation process, thereby greatly increasing the effective absorption optical path of the gas and improving detection accuracy and sensitivity. It also has the advantage of small gas chamber volume.

[0039] like Figure 2-3 As shown, the detection method of the fiber optic cavity gas detection system based on convolutional neural networks specifically includes:

[0040] Step 1: The mixed gas to be tested is placed in the test gas chamber 6, and an annular gas detection system as described in the fiber optic cavity gas detection system is built to obtain the absorption spectrum of the mixed gas to be tested in the test gas chamber; the photodetector 8 collects the analog signal of the signal light, converts the signal light into an electrical signal and outputs it to the data acquisition card 9; the data acquisition card 9 collects the sensing electrical signal transmitted back from the photodetector 8 and outputs it to the industrial control computer 10 for signal processing.

[0041] Step Two: The industrial computer 10 performs preliminary processing on the electrical signal, scanning in the vicinity of 1959nm (ammonia), 1965nm (methane), 1945nm (hydrogen sulfide), 1970nm (hydrogen bromide), and 2004nm (carbon dioxide), while simultaneously acquiring absorption spectrum data. Since system noise and environmental noise can reduce the signal-to-noise ratio of the spectral signal, affecting the acquisition of gas absorption lines, Haar wavelet transform is used to denoise the absorption spectrum data.

[0042] Step 3: Due to the output laser fluctuation problem during wavelength scanning, a baseline removal operation is performed after obtaining the data to subtract the baseline signal of the light that was not absorbed by the test gas, and then Lorentz line fitting is performed; the spectral distribution of each component gas in the mixture is obtained through multi-peak fitting, the absorption spectrum is separated, and then the transformed and fitted absorption spectrum signal is sent to a pre-trained convolutional neural network for recognition.

[0043] The basic principle of trace gas absorption sensing follows the Lambert-Beer law: when laser light of different wavelengths passes through a gas chamber containing the gas to be measured, the intensity will decrease. The transmitted light intensity can be expressed as...

[0044]

[0045] In the formula: I0(v) is the incident light intensity, I(v) is the transmitted light intensity, and v is the laser frequency (cm). -1 ), α g (v) represents the absorption cross section (cm) of the gas spectral line at different frequencies. 2 ·molecule -1 ·atm), C is the number density of gas molecules (molecule·cm). -3 ·atm -1 ), L eff The effective absorption optical path (cm) is the distance at which the gas interacts with the light.

[0046] When using the direct absorption method to detect different types and concentrations of gases, to characterize the strength of gas absorption, absorbance K(v) is defined as the attenuation of the system light intensity caused by the absorption of the gas to be measured. By transforming equation (1), the expression for absorbance K(v) can be obtained:

[0047]

[0048] As shown in equation (2), under the conditions of constant temperature, pressure, and effective absorption optical path, the gas concentration and absorbance are directly proportional in an ideal state. By calibrating the concentration and absorbance of the gas to be measured, quantitative analysis of the gas concentration can be achieved. Due to the absorption effect of the gas, the laser output from pump source 1 is amplified multiple times in the fiber optic loop and passes through the gas to be measured multiple times, thus increasing the effective absorption optical path and improving the gas sensing sensitivity.

[0049] Step 4: Input the pre-processed signal into the gas classification and concentration model based on convolutional neural networks for calculation.

[0050] In actual measurement, there may be problems such as aliasing and broadening of absorption lines in mixed gases, which can affect wavelength positioning and concentration demodulation. Therefore, this invention uses deep learning algorithms to realize gas identification and concentration prediction, which reduces the errors caused by inherent factors to a certain extent and avoids complicated data processing.

[0051] This embodiment detects five types of gases. The gas absorption lines overlap, affecting detection accuracy. To address this issue, we first selected wavelengths with minimal impact for absorption line detection; for example, we chose the 2004nm band, where CO2 absorption is particularly pronounced. Secondly, we used multi-peak fitting to obtain the spectral distribution of each component gas in the mixture, improving the measurement accuracy of the mixed gas.

[0052] S401: Training a Convolutional Neural Network

[0053] To avoid losing important information in the original data after 2D conversion and to improve the processing speed, a one-dimensional convolutional neural network is used to extract spatial features. Considering the temporal correlation of sensor response sequences, a long short-term memory network (LSTM) with strong learning ability in long-term temporal dependence is introduced to build a multi-task learning model, namely the MTL-1DCNN-LSTM model.

[0054] 2000 optical signal samples were pre-selected, each sample having a dimension of (2000, Y). This 2000 dimension was constructed by concatenating 500 spectral data points from the absorption wavelength ranges of carbon dioxide, hydrogen sulfide, ammonia, hydrogen bromide, and methane; Y = (y cla ,y reg_NH3 ,y reg_CO2 ,y reg_CH4 ,y reg_H2S ,y reg_HBr ), where y cla For gas classification, labels are used to classify the following five gas mixtures: pure nitrogen, ammonia mixed with nitrogen, methane mixed with nitrogen, carbon dioxide mixed with nitrogen, hydrogen sulfide mixed with nitrogen, hydrogen bromide mixed with nitrogen, and the mixture of NH3, CH4, H2S, HBr, and CO2, respectively labeled 0, 1, 2, 3, 4, 5, and 6. reg_NH3 y reg_CH4 y reg_CO2 y reg_H2S and y reg_HBr The values ​​represent the concentrations of NH3, CH4, CO2, H2S, and HBr, respectively. The convolutional neural network feature extraction part is a simplified version of the Vgg16 model. The kernel size of each convolutional layer is 1×3. First, shallow shared feature extraction is performed, with the number of kernels in the shared parameter layers being 16, 16, 64, and 64 respectively. Then, deep feature extraction specific to each branch task is performed, with 64 kernels in the classification task and 128, 128, 64, and 64 in the regression task. The activation function is set to ReLU. The features obtained after deep convolution are input into an LSTM layer for further extraction of temporal features. The num_layers parameter of the LSTM layer is set to 32. In the classification task, since it is necessary to classify five gas events, the batch_size of the fully connected (Dense) layer is set to 5, and the Softmax activation function is selected. In the regression task, only gas concentration values ​​are output, so the batch_size of the Dense layer is set to 1, and the linear activation function is selected.

[0055] The convolutional neural network (CNN) employs different loss functions for classification and regression tasks due to their different learning mechanisms. Classification tasks use cross-entropy loss, while regression tasks use mean squared error (MSE), the most common loss function in regression problems, which is computationally simple and has good convergence. The overall loss function of the CNN is a linear sum of the loss functions for both tasks. Adjusting the weights of the classification and regression loss functions balances the network's focus on both tasks.

[0056] S402: Save the trained convolutional neural network model as a gas classification and concentration model. Extract and identify features from the separated absorption spectrum signal obtained in step three within the gas classification and concentration model. Output the results through the convolutional neural network model. Based on the results (Y=(y cla ,y reg_NH3 ,y reg_CO2 ,y reg_CH4 ,y reg_H2S ,y reg_HBr The type and concentration of the mixed gas can then be obtained.

[0057] Finally, the type and concentration of the mixed gas are displayed on the industrial control computer, enabling the classification and concentration detection of the mixed gas.

[0058] Although preferred embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these are within the scope of protection of the present invention.

Claims

1. A fiber optic cavity gas detection system based on a convolutional neural network, characterized in that, It includes a pump light source (1), wavelength division multiplexer (2), thulium-doped fiber (3), isolator (4), tunable filter (5), test gas chamber (6), fiber optic circulator (7), photodetector (8), data acquisition card (9), and industrial control computer (10). The pump source (1) is a pump source with a center wavelength around 1570 nm; The thulium-doped fiber (3) can generate amplified spontaneous emission light from 1750 nm to 2100 nm; The test chamber (6) is used to hold at least two mixed gases to be tested, and the length of the test chamber (6) is less than 20 cm; The output end of the fiber optic circulator (7) with a splitting ratio of 10% is connected to the other input end of the wavelength division multiplexer (2), and the output end with a splitting ratio of 90% is connected to the input end of the photodetector (8), so that the laser output by the pump source (1) is repeatedly amplified in the fiber optic loop and passes through the gas to be tested multiple times. The output end of the photodetector (8) is connected to the input end of the data acquisition card (9), and the output end of the data acquisition card (9) is connected to the industrial control computer (10). The industrial control computer (10) receives data from the data acquisition card (9), performs photoelectric demodulation on the interference signal, and uses a pre-trained gas classification and concentration model to extract features and identify gas classification and concentration detection. The photoelectric demodulation step includes: Step 1: The industrial control computer (10) performs preliminary processing on the received electrical signal, extracts the absorption spectrum signal according to the absorption wavelength of the test gas, and after extracting the spectrum, uses a low-pass filter to remove high-frequency noise, and uses Haar wavelet transform and empirical mode decomposition to denoise the original absorption spectrum, thereby obtaining the absorption spectrum signal. Step 2: Further process the noise-reduced absorption spectrum signal obtained in Step 1, perform baseline removal to subtract the baseline signal of light that was not absorbed by the test gas, and then perform Lorentz line fitting; obtain the spectral distribution of each component gas in the mixture through multi-peak fitting, separate the absorption lines, and obtain the separated spectral signal. Step 3: Input the spectral signal processed by multi-peak fitting and spectral line separation in Step 2 into the gas classification and concentration model based on convolutional neural network for feature extraction and recognition, thereby outputting the type of mixed gas and the concentration of each gas; wherein the gas classification and concentration model is pre-trained with gas samples using convolutional neural network; The gas classification and concentration model uses a one-dimensional convolutional neural network to extract spatial features and introduces a long short-term memory network to build a multi-task learning model.

2. The fiber optic cavity gas detection system based on convolutional neural networks according to claim 1, characterized in that, The mixed gas in the test chamber (6) is selected from at least two of the following gases: ammonia, hydrogen sulfide, methane, carbon dioxide, and hydrogen bromide.

3. The fiber optic cavity gas detection system based on convolutional neural networks according to claim 2, characterized in that, In step one, the absorption spectrum of methane gas is extracted at 1965 nm, the absorption spectrum of ammonia gas is extracted at 1959 nm, the absorption spectrum of hydrogen sulfide gas is extracted at 1945 nm, the absorption spectrum of hydrogen bromide gas is extracted at 1970 nm, and the absorption spectrum of carbon dioxide gas is extracted at 2004 nm.

Images