Alternating current driven limited space toxic gas spectrum detection system and method

Abstract

The application discloses an alternating current driven limited space toxic gas spectrum detection system and method, which modulates a wide spectrum infrared light source through different energy levels alternating current, expands the detection characteristic dimension by using spectral nonlinear evolution, and relieves the multi-component gas spectrum line cross interference; relies on a long light path sealed absorption cell and a multi-channel narrow band filter light detector to collect the transmitted spectrum, and completes digitization through hardware filtering and synchronous sampling; adopts interval difference method to extract pulse amplitude characteristics, combines one-dimensional steady-state Kalman filtering denoising and constructs a channel ratio, collects environmental parameters in real time, combines the multi-current channel ratio, derived expansion characteristics and environmental parameters into a multi-dimensional feature vector, and inputs a gradient boosting regression tree model to realize synchronous inversion of multi-component gas concentration. The application significantly improves the detection precision of complex toxic gas in a complex temperature and pressure environment and environmental adaptation ability while reducing the hardware cost, and is suitable for online monitoring of limited spaces such as mines and urban pipe networks.

Description

Technical Field

[0001] This invention relates to the field of optical gas sensing and photoelectric detection technology, specifically to a confined space spectral detection system and method for toxic gases driven by alternating current. Background Technology

[0002] In confined spaces such as underground mines, urban pipe networks, septic tanks, and sealed containers, the decomposition of organic matter or poor ventilation often leads to the generation and accumulation of various complex toxic gases, including methane, carbon monoxide, hydrogen sulfide, and sulfur dioxide. This not only seriously threatens the lives of workers but also easily triggers acute poisoning and secondary accidents. These confined spaces are often small, highly dangerous, and hidden. Traditional manual sampling and contact electrochemical detection technologies are limited by issues such as sensor poisoning, high maintenance frequency, and personnel safety costs, making timely and comprehensive continuous online monitoring difficult. Differential absorption spectroscopy, however, utilizes the characteristic absorption lines of gases for non-contact measurement. It offers advantages such as fast response, intrinsic safety, and resistance to poisoning, making it an effective means of early warning of toxic gases in confined spaces.

[0003] Existing differential absorption spectroscopy detection techniques mainly employ fixed-current driven infrared light sources. The signal characteristics output by the infrared detector are singular, and when faced with overlapping spectral lines of multi-component toxic gases, the lack of multi-dimensional modulation features easily leads to serious misjudgments in component concentration retrieval. For example, US Patent No. 7358489B2 discloses an NDIR gas sensor using two infrared light sources driven by different powers, which calculates gas concentration by the ratio of high and low period outputs. However, this scheme only aims to simulate the effect of traditional dual-beams to eliminate light source fluctuations, without utilizing the nonlinear evolution of the spectrum under different currents to expand the feature dimensions, thus failing to solve the problem of multi-component cross-interference.

[0004] Furthermore, to eliminate common-mode interference such as light source aging and window contamination, existing solutions (such as Chinese invention patent applications with publication numbers CN116818699A and CN117250166A) often require independent configuration of optical paths and reference channels for each target gas. This not only significantly increases the hardware complexity and manufacturing cost of multi-channel sensors but also limits their integrated application in compact, confined spaces. Simultaneously, the variable temperature and pressure within a confined space easily cause nonlinear changes in the gas absorption coefficient, resulting in poor concentration inversion accuracy and environmental adaptability for traditional static algorithms (such as linear fitting and mechanistic models). Although existing technologies also apply gradient boosting regression trees to gas detection, the input features are all spectral data under a single current, without incorporating dynamic features from multiple current levels, resulting in limited temperature and pressure compensation effects.

[0005] Therefore, the research direction of this invention is to provide a new spectral detection system and method for complex toxic gases in a confined space that can effectively enrich the detection feature dimensions and reduce the cost of hardware channels, while eliminating the influence of nonlinear temperature and pressure changes on the inversion results, so as to achieve accurate inversion and detection of complex toxic gas concentrations in complex and harsh environments. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a confined space toxic gas spectral detection system and method driven by alternating current. It combines time-division multiplexed alternating current light source driving, dynamic pulse amplitude feature extraction, one-dimensional steady-state Kalman time-domain filtering, and gradient boosting regression tree model that integrates environmental temperature and pressure parameters. This effectively enriches the feature dimensions of spectral detection and eliminates the influence of multi-component cross-absorption and environmental temperature and pressure nonlinear drift. While eliminating the need for an independent hardware reference channel and reducing hardware costs, it achieves accurate inversion and detection of complex toxic gas concentrations in complex and harsh environments.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: On the one hand, an alternating current driven spectral detection system for toxic gases in a confined space is provided, including an optical unit, a circuit unit, and a computing unit.

[0008] The optical unit includes an infrared light source, an infrared detector, an absorption cell, and an environmental parameter detector. The composite toxic gas to be tested is injected into the absorption cell, and the infrared light generated by the infrared light source enters the absorption cell, is transmitted through the composite toxic gas to be tested, and is received by the infrared detector. The environmental parameter detector is installed in the absorption cell and is used to monitor the temperature and pressure parameters in the absorption cell.

[0009] The circuit unit includes a light source driving circuit, a filtering circuit, and a sampling circuit. The light source driving circuit periodically and alternately outputs currents of different energy levels to drive the infrared light source, causing the spectrum emitted by the infrared light source to undergo nonlinear evolution and reconstruction. The infrared detector collects the transmitted light intensity signal and converts it into an electrical signal. After noise reduction by the filtering circuit, the signal is synchronously sampled into a digital signal by the sampling circuit.

[0010] The computing unit includes a dynamic feature extraction module, a Kalman filter module, and a concentration inversion module. The computing unit extracts pulse amplitude features through the dynamic feature extraction module, and after noise reduction by the Kalman filter module, constructs a multi-current channel ratio signal. Then, the multi-current channel ratio signal, the derived extended features, and the temperature and pressure parameters are combined into a multi-dimensional feature vector, which is input into the concentration inversion module to synchronously output the concentration of each component gas.

[0011] Furthermore, the infrared light source is a broadband light source; the infrared detector is a multi-channel detector equipped with multiple narrowband filters of different types, and the transmission wavelength of each narrowband filter covers the characteristic absorption spectrum of the target gas.

[0012] Furthermore, the absorption cell adopts a long optical path sealed gas chamber with an inner wall resistant to corrosion by toxic gases and without adsorption, catalysis, or non-chemical reaction activity with the target gas; the environmental parameter detector includes a temperature detector and a pressure detector, used to collect the temperature and pressure of the composite toxic gas to be tested in the absorption cell in real time.

[0013] Furthermore, the light source driving circuit is a programmable constant current source driving circuit, which can alternately output current signals of at least two different energy levels to drive the infrared light source to emit infrared spectra with different light intensity distribution states in different energy level states, thereby causing the infrared spectrum generated by the infrared light source to undergo nonlinear evolution and reconstruction with blue shift, broadening and amplitude enhancement.

[0014] Furthermore, the filtering circuit is a low-pass or band-pass filter circuit, used to suppress noise and limit the bandwidth of the weak electrical signal output by the infrared detector; the sampling circuit performs synchronous frequency doubling sampling based on the frequency of the current-driven signal, and at the same time converts the collected analog voltage signal into a digital signal that can be processed by the computing unit.

[0015] On the other hand, a method for spectroscopic detection of toxic gases in a confined space driven by alternating current is provided, comprising the following steps: Step 1: Completely replace the composite toxic gas to be tested in the long optical path sealed absorption cell using free diffusion or pump suction.

[0016] Step 2: The light source driving circuit outputs the first energy level current to drive the infrared light source, which then emits infrared spectrum in the first energy level state.

[0017] Step 3: The infrared detector acquires the analog electrical signals of the transmitted light intensity of each detection channel. After noise suppression by the filtering circuit, the sampling circuit synchronously samples and converts the analog electrical signals into digital signals. At the same time, the environmental parameter detector collects temperature and pressure parameters in real time.

[0018] Step 4: The computing unit stores the multi-channel digital signal data and environmental parameter data corresponding to the first energy level state in real time.

[0019] Step 5: The light source driving circuit switches to output the second energy level current to drive the infrared light source; repeat steps 2 to 4 to obtain the multi-channel digital signal data and environmental parameter data corresponding to the second energy level state; the light source driving circuit alternately outputs the first energy level and second energy level current to drive the infrared light source in turn, and continues to repeat steps 2 to 4 to obtain the multi-channel digital signal data and environmental parameter data corresponding to different energy level states.

[0020] Step 6: The computing unit calls the dynamic feature extraction algorithm and the digital filtering algorithm to perform interval averaging on the digital signals sampled under the first current energy level state and the second current energy level state, extracts the pulse amplitude features under the two energy level states to generate a periodic differential voltage mean sequence, and then inputs it into a one-dimensional steady-state Kalman filter model for time-domain smoothing and noise reduction, and constructs the channel ratio signal.

[0021] Step 7: The calculation unit calls the concentration inversion intelligent algorithm to construct a gradient boosting regression tree model. The channel ratios of the two energy level states, the extended features derived from the channel ratios, real-time temperature, and real-time pressure are combined as a multi-dimensional feature input vector. The input is fed into the gradient boosting regression tree model, which simultaneously inverts and outputs the concentration values ​​of each component gas in the complex toxic gas.

[0022] Furthermore, the interval difference formula for the dynamic feature extraction algorithm is:

[0023] In the formula, The average value of the periodic differential voltage. and These are the arithmetic mean voltage values ​​for the high-level steady-state interval and the low-level background interval within a single modulation cycle, respectively.

[0024] The Kalman filter smoothing recursive formula of the digital filtering algorithm includes: State-one-step prediction equation:

[0025] One-step prediction equation for covariance:

[0026] Kalman gain calculation equation:

[0027] State update estimation equation:

[0028] Covariance update equation:

[0029] In the formula, This is the state estimate of the channel ratio signal. For the systematic error covariance, For Kalman gain, This is the ratio of the actual observations currently being input. For process noise covariance, To measure the noise covariance, It is the identity matrix. This represents the current discrete time step.

[0030] Furthermore, in step seven, the multidimensional feature input vector The expression is:

[0031] In the formula, For the first Under what energy level current state, the first Selected channel ratio characteristics; This is an extended feature derived from the channel ratios of different current energy level states and different detection channels through mathematical transformation; The real-time temperature of the composite toxic gas to be tested in the absorption cell; This represents the real-time pressure of the composite toxic gas to be tested within the absorption cell.

[0032] The architecture of the gradient boosting regression tree model includes: Weak classifier initialization:

[0033] The calculation of negative gradient residuals for alternating currents at different energy levels and multi-channel characteristic states is introduced:

[0034] Optimal fit in the leaf node region:

[0035] Strong classifier state update:

[0036] In the formula, For multidimensional feature input vectors; This represents the actual labeled value of the concentration of the complex toxic gas components; The loss function; The total number of samples; The fitted value is a constant; This is the index of the current regression tree; This is the sequence number of the current sample; For the first tree to the first The negative gradient residuals of each sample; This is the index of the current leaf node; For the first The first tree A leaf node region; This represents the optimal output value for the corresponding leaf node. This represents the total number of leaf nodes in a single tree. For indicator functions; This is the learning rate.

[0037] After the gradient boosting regression tree model is established, it is iterated step by step in the forward direction. Finally, the strong classifier synchronously inverts and outputs the true concentration values ​​of each component gas of methane and carbon monoxide in the compound toxic gas, and automatically compensates for the cross-absorption spectral drift caused by nonlinear disturbances of real-time temperature and real-time pressure.

[0038] Compared with the prior art, the present invention has the following advantages: 1. This invention employs at least two different energy levels of current to drive an infrared light source. By utilizing the nonlinear evolution of the spectrum under different currents (blue shift, broadening, amplitude enhancement), it enriches the spectral feature information in both the current energy level and time domain dimensions, expands the spectral feature dimensions, improves the distinguishability of complex toxic gases, and effectively alleviates the problem of inversion misjudgment caused by the cross-over of multi-component spectral lines.

[0039] 2. This invention uses a time-division multiplexing method with alternating current driving modes at different energy levels, combined with a dynamic feature extraction algorithm, to directly calculate the interval difference and channel ratio through software calculation. While ensuring anti-common-mode interference performance, it eliminates the need for an independent hardware reference channel required by traditional solutions, simplifies the optical path and circuit structure, and reduces hardware costs by more than 30%.

[0040] 3. This invention combines the multi-current channel ratio, derived extended features, and real-time temperature and pressure parameters into a multi-dimensional feature vector. By utilizing the gradient to enhance the nonlinear fitting ability of the regression tree model, it achieves adaptive compensation for the gas cross-absorption law of temperature and pressure, thereby improving environmental adaptability. In a wide temperature and pressure range of -10℃ to 50℃ and 80kPa to 120kPa, the inversion accuracy is improved by more than 40%. Attached Figure Description

[0041] Figure 1 This is a hardware block diagram of the alternating current drive detection system according to Embodiment 1 of the present invention.

[0042] Figure 2 The graphs are relative light intensity variation curves of the infrared light source under different alternating characteristic driving currents according to Embodiment 1 of the present invention; where graph a is the CH4 channel and graph b is the CO channel.

[0043] Figure 3 This is a timing control flowchart of the spectral detection method based on alternating current driving according to Embodiment 1 of the present invention.

[0044] Figure 4 This is a diagram of the gradient boosting regression tree concentration inversion model architecture of Embodiment 1 of the present invention.

[0045] Figure 5This is a comparison chart of the concentration inversion results of the gradient boosting regression tree model in Embodiment 1 of the present invention. Detailed Implementation

[0046] The present invention will be further described below.

[0047] Example 1: Detection of a combined toxic gas mixture of methane (CH4) and carbon monoxide (CO) This embodiment addresses the most common methane and carbon monoxide complex gas found in confined spaces of mines, and constructs a spectroscopic detection system based on alternating current driving, as specifically implemented below: like Figure 1 and Figure 3 As shown, the hardware components of the detection system in this embodiment include an optical unit, a circuit unit, and a computing unit.

[0048] Optical Unit: Employs a broadband infrared light-emitting diode (IR-LED) with a wavelength range of 2~5μm; an InGaAs photodiode multi-channel detector with three narrowband filters, the center wavelengths of which are 1657nm (CH4 characteristic absorption peak), 2331nm (CO characteristic absorption peak), and 2000nm (reference channel); the inner wall is made of stainless steel with a long optical path sealed absorption cell, resistant to corrosive gases such as hydrogen sulfide and carbon monoxide, with an effective optical path of 400cm, and a gold-plated reflector on one side; the absorption cell has an air inlet and an air outlet; the environmental parameter detector uses an integrated temperature and pressure sensor, with a measurement range of -20℃~60℃ and 80kPa~120kPa, and an accuracy of ±0.1℃ and ±0.1kPa.

[0049] Circuit unit: Programmable constant current source drive circuit, output current range 0~500mA, accuracy ±1mA, can output 500Hz periodic alternating current signal; fourth-order RC low-pass filter circuit, cutoff frequency 2kHz; high-frequency synchronous sampling circuit based on AD7606 chip, sampling rate 20kHz, 16-bit resolution.

[0050] Computing Unit: Employs an STM32H743IIT6 microcontroller chip with a main frequency of 400MHz, running the STM32 software system, and internally deploys dynamic feature extraction algorithm, one-dimensional steady-state Kalman filter algorithm, and gradient boosting regression tree concentration inversion algorithm.

[0051] like Figure 2 As shown, under the alternating drive of the first energy level current (200mA) and the second energy level current (300mA), the energy distribution of the emission spectrum of the infrared light source exhibits a significant nonlinear evolution: the relative light intensity of the CH4 channel (1657nm) increases from 0.6 to 0.9, and the relative light intensity of the CO channel (2331nm) increases from 0.5 to 0.8, with a slight blue shift in the spectral peaks, providing a data basis for the subsequent extraction of the differential features of the two energy level states.

[0052] The detection timing control process in this embodiment includes the following steps: Step 1: The methane and carbon monoxide composite gas to be tested is introduced into the absorption cell using a pump, with the flow rate controlled at 0.5 L / min, for 30 seconds to completely replace the original gas in the cell.

[0053] Step 2: The STM32H743IIT6 microcontroller controls the programmable constant current source drive circuit to output the first energy level current (200mA) drive signal to the infrared light source, so that it emits infrared spectrum in the first energy level state.

[0054] Step 3: The multi-channel infrared detector acquires the analog electrical signal of the transmitted light intensity of each detection channel. After noise suppression by the fourth-order RC low-pass filter circuit, the AD7606 sampling circuit synchronously samples the signal at a frequency of 20kHz and converts it into a 16-bit digital signal. At the same time, the temperature and pressure sensors collect the temperature T and pressure P parameters in the absorption cell in real time.

[0055] Step 4: The microcontroller enables the DMA double-buffered transfer mechanism to store the multi-channel digital signal corresponding to the first energy level current state and the temperature and pressure data timing sequence into the internal SRAM buffer.

[0056] Step 5: The microcontroller controls the light source driving circuit to switch the output of the second energy level current (300mA) driving signal, repeating steps 2 to 4 to obtain and store the multi-channel digital signal and environmental parameter data under the second energy level state; then, the microcontroller controls the light source driving circuit to alternately output the first energy level and second energy level current to drive the infrared light source in turn, and continues to repeat steps 2 to 4 to obtain the multi-channel digital signal data and environmental parameter data corresponding to different energy level states.

[0057] Step Six: The microcontroller invokes the dynamic feature extraction algorithm and digital filtering algorithm to perform interval averaging on the sampled data in the first and second energy levels, extracting the voltage arithmetic mean of the high voltage steady-state interval (40% duty cycle) and the low-level background interval (60% duty cycle) respectively. and The mean value of the periodic differential voltage is calculated using the interval difference formula. Generate a periodic differential voltage mean sequence.

[0058] The periodic differential voltage mean sequence is input into a one-dimensional steady-state Kalman filter model, and the process noise covariance is set. Measure noise covariance Temporal smoothing and denoising are performed.

[0059] Constructing the channel ratio signal: , , , The subscripts 1 and 2 correspond to the first energy level state and the second energy level state, respectively.

[0060] Step 7: The microcontroller invokes the gradient boosting regression tree concentration inversion algorithm to establish the gradient boosting regression tree concentration inversion model architecture as follows: Figure 4 As shown, after completion, a multi-dimensional feature input vector is constructed. ,in and These are derived and extended features.

[0061] Load the pre-trained gradient boosting regression tree model, and set the model parameters as follows: number of weak learners M=100, number of leaf nodes per tree K=8, and learning rate. The loss function used is mean squared error (MSE). The model simultaneously inverts and outputs the concentration values ​​of methane and carbon monoxide.

[0062] To verify the detection effect of this embodiment, a mixed gas of 0-5% VOL methane and 0-500 ppm carbon monoxide was prepared in a laboratory environment, and tests were conducted within a temperature range of -10℃ to 50℃ and a pressure range of 80 kPa to 120 kPa. Figure 5 As shown, the concentration prediction values ​​output by the model of this invention are highly consistent with the standard values ​​measured by the gas chromatograph. The absolute error of methane inversion is ≤0.05%VOL, and the absolute error of carbon monoxide inversion is ≤5ppm, which is far superior to the traditional fixed current driving scheme (methane error ≤0.12%VOL, carbon monoxide error ≤15ppm).

[0063] Example 2: Detection of the combined toxic gases of hydrogen sulfide (H2S) and sulfur dioxide (SO2) This embodiment addresses the common hydrogen sulfide and sulfur dioxide complex gases found in urban pipe networks and septic tanks by adapting the detection system as follows: The system architecture of this embodiment is basically the same as that of Embodiment 1, with the main difference being the adaptation of optical unit parameters and algorithm model: Optical Unit: Employs a mid-infrared broadband light source with a wavelength range of 3~10μm; a PbS photodiode multi-channel detector with three narrowband filters, the center wavelengths of which are 3900nm (characteristic absorption peak of H2S), 7400nm (characteristic absorption peak of SO2), and 5000nm (reference channel); a long optical path sealed absorption cell with an inner wall coated with polytetrafluoroethylene, with an effective optical path of 600cm, further improving the detection sensitivity of low-concentration gases.

[0064] Circuit unit: The programmable constant current source drive circuit outputs alternating signals of the first energy level current (150mA) and the second energy level current (250mA), with a modulation frequency of 300Hz; the sampling circuit sampling rate is adjusted to 12kHz.

[0065] Algorithm Model: The gradient boosting regression tree model was retrained. The training set contained feature data of a mixture of hydrogen sulfide (0-100 ppm) and sulfur dioxide (0-50 ppm) under different temperature and pressure conditions. The model parameters were set as follows: number of weak learners M=120, number of leaf nodes in a single tree K=10, and learning rate... .

[0066] The detection process in this embodiment is consistent with that in Embodiment 1. Test results show that within the temperature range of 0℃ to 40℃ and the pressure range of 90kPa to 110kPa, the absolute error of hydrogen sulfide inversion is ≤0.5ppm, and the absolute error of sulfur dioxide inversion is ≤1ppm, meeting the detection accuracy requirements for confined spaces in urban pipe networks. Compared with traditional electrochemical sensors, this system has no poisoning failure issues, and the maintenance cycle is extended from 3 months to more than 2 years.

Claims

1. A confined space toxic gas spectroscopic detection system driven by alternating current, characterized in that, It includes optical units, circuit units, and computing units; The optical unit includes an infrared light source, an infrared detector, an absorption cell, and an environmental parameter detector; the gas to be tested is placed inside the absorption cell, and the environmental parameter detector is installed inside the absorption cell to monitor the temperature and pressure parameters inside the absorption cell. The circuit unit includes a light source driving circuit, a filtering circuit, and a sampling circuit; the light source driving circuit alternately outputs different energy level currents to drive the infrared light source, causing the spectrum emitted by the infrared light source to undergo a nonlinear evolution, enter the absorption cell, and be transmitted to the infrared detector through the gas to be measured. The infrared detector collects the transmitted light intensity signal and converts it into an electrical signal. After noise reduction by the filtering circuit, it is synchronously sampled into a digital signal by the sampling circuit. The computing unit includes a dynamic feature extraction module, a Kalman filtering module, and a concentration inversion module; The computing unit extracts pulse amplitude features through a dynamic feature extraction module, and after noise reduction by a Kalman filter module, constructs a multi-current channel ratio signal. Then, the multi-current channel ratio signal and its derived extended features are combined with temperature and pressure parameters to form a multi-dimensional feature vector, which is input to the concentration inversion module to synchronously output the concentration of each component gas.

2. The system according to claim 1, characterized in that, The infrared light source is a broadband light source; the infrared detector is a multi-channel detector equipped with multiple narrowband filters of different types, and the transmission wavelength of each narrowband filter covers the characteristic absorption spectrum of the target gas.

3. The system according to claim 1, characterized in that, The absorption cell is a long-optical-path sealed gas chamber with an inner wall resistant to corrosion by toxic gases and without adsorption, catalysis, or chemical reaction activity with the target gas; the environmental parameter detectors include a temperature detector and a pressure detector, used to collect the temperature and pressure of the gas to be measured in the absorption cell in real time.

4. The system according to claim 1, characterized in that, The light source driving circuit is a programmable constant current source driving circuit, which can alternately output current signals of at least two different energy levels to drive the infrared light source to emit infrared spectra with different light intensity distribution states in different energy level states.

5. The system according to claim 1, characterized in that, The filtering circuit is a low-pass or band-pass filter circuit, used to suppress noise and limit the frequency band of the weak electrical signal output by the infrared detector; the sampling circuit performs synchronous frequency doubling sampling based on the frequency of the current-driven signal, and at the same time converts the collected analog voltage signal into a digital signal that can be processed by the computing unit.

6. A method for spectroscopic detection of toxic gases in a confined space driven by alternating current, characterized in that, The system described in any one of claims 1 to 5 is implemented by comprising the following steps: Step 1: Completely replace the gas to be tested in the absorption cell using free diffusion or pump suction. Step 2: The light source driving circuit outputs the first energy level current to drive the infrared light source, which then emits infrared spectrum in the first energy level state. Step 3: The infrared detector acquires the analog electrical signal of the transmitted light intensity of each detection channel. After noise suppression by the filtering circuit, the sampling circuit synchronously samples and converts the analog electrical signal into a digital signal. At the same time, the environmental parameter detector collects temperature and pressure parameters in real time. Step 4: The computing unit stores the multi-channel digital signal data and environmental parameter data corresponding to the first energy level state in real time; Step 5: The light source driving circuit switches to output the second energy level current to drive the infrared light source; repeat steps 2 to 4 to obtain the multi-channel digital signal data and environmental parameter data corresponding to the second energy level state; the light source driving circuit alternately outputs the first energy level and second energy level current to drive the infrared light source in turn, and continues to repeat steps 2 to 4 to obtain the multi-channel digital signal data and environmental parameter data corresponding to different energy level states. Step 6: The computing unit calls the dynamic feature extraction algorithm and the digital filtering algorithm to perform interval averaging on the digital signals sampled from the first and second energy levels respectively, extract the pulse amplitude features of the two energy level states to generate a periodic differential voltage mean sequence, and then input it into a one-dimensional steady-state Kalman filter model for time-domain smoothing and noise reduction, and construct the channel ratio signal. Step 7: The calculation unit calls the concentration inversion intelligent algorithm to construct a gradient boosting regression tree model. The channel ratios of the two energy level states, the extended features derived from the channel ratios, the real-time temperature, and the real-time pressure are combined as a multi-dimensional feature input vector and input into the gradient boosting regression tree model. The concentration values ​​of each component gas in the compound toxic gas are simultaneously inverted and output.

7. The method according to claim 6, characterized in that, The interval difference formula for the dynamic feature extraction algorithm is: In the formula, The average value of the periodic differential voltage. and These are the arithmetic mean voltage values ​​for the high-voltage steady-state interval and the low-level background interval, respectively, within a single modulation cycle. The Kalman filter smoothing recursive formula of the digital filtering algorithm includes: State-one-step prediction equation: One-step prediction equation for covariance: Kalman gain calculation equation: State update estimation equation: Covariance update equation: In the formula, This is the state estimate of the channel ratio signal. For the system error covariance, For Kalman gain, This is the ratio of the actual observations currently being input. For process noise covariance, To measure the noise covariance, It is the identity matrix. This represents the current discrete time step.

8. The method according to claim 6, characterized in that, In step seven, the multidimensional feature input vector The expression is: In the formula, For the first Under what energy level current state, the first The ratio characteristics of selected channels; This is an extended feature derived from the channel ratios of different energy level current states and different detection channels through mathematical transformation; The real-time temperature of the gas to be measured in the absorption cell; This represents the real-time pressure of the gas to be measured within the absorption cell. The architecture of the gradient boosting regression tree model includes: Weak classifier initialization: The calculation of negative gradient residuals for alternating currents at different energy levels and multi-channel characteristic states is introduced: Optimal fit in the leaf node region: Strong classifier state update: In the formula, For multidimensional feature input vectors; This represents the actual labeled value of the concentration of the complex toxic gas components; The loss function; The total number of samples; The fitted value is a constant; This is the index of the current regression tree; This is the sequence number of the current sample; For the first tree to the first The negative gradient residuals of each sample; This is the index of the current leaf node; For the first The first tree A leaf node region; This represents the optimal output value for the corresponding leaf node. This represents the total number of leaf nodes in a single tree. For indicator functions; This is the learning rate.

Images