An overlapping spectrum separation device and method based on a new evolutionary deep learning model

Abstract

The application is suitable for the technical field of gas sensing, and provides an overlapping spectrum separation device and method based on a new evolutionary deep learning model. First, elite reverse learning, golden sine algorithm and dynamic nonlinear inertia weight are combined to form an improved whale optimization algorithm. Then, the improved whale optimization algorithm is used to automatically optimize the core hyperparameters of the deep learning framework, so as to reduce the time cost of manual adjustment according to experience. Finally, the normalized overlapping gas absorption spectrum is separated through the optimal deep learning framework after training, so as to guarantee the separation accuracy of each gas. The method can effectively improve the measurement accuracy, the model response time can reach the millisecond level, and when the gas composition, concentration, temperature and pressure change, the corresponding absorption spectrum training data set can be replaced, so that a new method is provided for the simultaneous measurement of multi-component gases in different environments.

Description

Technical Field

[0001] This invention belongs to the field of gas sensing technology, and particularly relates to an overlapping spectral separation device and method based on a novel evolutionary deep learning model. Background Technology

[0002] During tunnel construction and operation, CH4, when mixed with air at specific concentrations, is highly likely to cause an explosion. To ensure a safe working environment, the CH4 concentration inside tunnels is typically limited to ≤1%. Similarly, as a colorless and odorless gas, CO poses a serious health risk even at relatively low concentrations due to its high toxicity. Therefore, the CO concentration in the tunnel operating environment usually needs to be controlled to no more than 24 ppm. Thus, achieving highly accurate real-time monitoring of CH4 and CO gases is a key technical requirement for ensuring the safe operation of tunnels.

[0003] Compared with traditional gas measurement techniques, tunable diode laser absorption spectroscopy offers advantages such as high sensitivity, strong selectivity, and rapid response, and is widely used in environmental monitoring, industrial process control, and combustion diagnostics. Systems using multiple lasers to measure multi-component gases typically suffer from drawbacks such as complex control, high cost, and difficulties in optical alignment. To address these issues, a method using a single laser to scan the absorption spectra of multi-component gases within the same wavelength band has been proposed. This method improves the stability of the measurement system and reduces signal crosstalk.

[0004] However, effectively resolving the cross-interference and even complete overlap between absorption lines of multi-component gases remains a challenging problem to be solved. Currently, many effective methods have been introduced to separate spectral lines. Dang et al. chose 4284.51 cm⁻¹. -1 The CH4 absorption line and 4285.0089 cm⁻¹ -1 The CO absorption lines were measured, and a novel spectral interference cancellation method was employed. This method utilizes second-harmonic 2f wavelength modulation spectroscopy combined with approximation techniques to effectively resolve the interference between closely packed absorption lines under atmospheric pressure, with a relative error of less than ±3.17% for CO concentration. Shao et al. adjusted the driving current of a 2.3µm distributed feedback laser under low pressure, selecting 4297.56 cm⁻¹ as the target. -1 and 4297.70 cm -1The CH4 and CO absorption lines were used to achieve simultaneous measurement of CH4 and CO in the atmosphere. While this method improves measurement accuracy, there is a high linearity between the gas concentrations of CO (0.046–4.6 ppm) and CH4 (0.487–48.7 ppm) and the 2f / 1f normalized signals. These methods effectively separate overlapping spectra and improve system performance, but introduce additional complexity. Sun et al. introduced a neural separator to address the challenge of ultraspectral overlap, selecting a spacing as narrow as 0.015 cm⁻¹. -1 (4291.5147cm respectively) -1 and 4291.4994 cm -1 The method, which uses CH4 and CO absorption lines, has been successfully applied to the measurement of dual gas concentrations in coal mines. The predicted concentrations show excellent consistency with the preset concentrations, and the correlation coefficient R for CH4 is high. 2 =0.99960, correlation coefficient R of CO 2 =0.99301, indicating a significant improvement in measurement accuracy. While this method simplifies the system architecture, its ability to accurately separate multiple gas components simultaneously remains limited. Furthermore, the neural network parameters require manual tuning based on empirical knowledge, leading to increased time costs. Summary of the Invention

[0005] The purpose of this invention is to provide an overlapping spectral separation device and method based on a novel evolutionary deep learning model, which aims to solve the problems mentioned in the background art.

[0006] The present invention is implemented as follows: an overlapping spectral separation device based on a novel evolutionary deep learning model includes a multi-pass cell, an internal temperature control device, and an external temperature control device. The external temperature control device is connected to the inlet and outlet pipes of the multi-pass cell. An inlet proportional valve and a pressure sensor are installed on the inlet pipe, and an outlet proportional valve and a vacuum pump are installed on the outlet pipe. A primary temperature sensor is installed on the external temperature control device, and a tertiary temperature sensor is installed on the multi-pass cell.

[0007] The light inlet of the multi-pass cell is optically connected to the light outlet of the laser. The light outlet of the multi-pass cell is optically connected to one end of the photodetector. The other end of the photodetector is electrically connected to the photodetector amplifier. The photodetector amplifier is also electrically connected to one end of the data acquisition card. The other end of the data acquisition card is electrically connected to one end of the microcomputer. The other end of the microcomputer is electrically connected to the driving temperature control circuit board. The other end of the driving temperature control circuit board is electrically connected to the laser.

[0008] Both the multi-channel cell and the photodetector are housed in an internal temperature control device, and a secondary temperature sensor is installed on the internal temperature control device.

[0009] Another objective of this invention is to provide an overlapping spectral separation method based on a novel evolutionary deep learning model, which, based on the aforementioned apparatus, includes the following steps:

[0010] Step 1: Use the first-level temperature sensor, the second-level temperature sensor and the third-level temperature sensor to maintain the stability of the measurement environment temperature. Turn on the vacuum pump, the inlet proportional valve and the outlet proportional valve to fill the multi-pass cell with the gas to be measured. Close the inlet proportional valve and use the pressure sensor to control the pressure inside the multi-pass cell. Then close the outlet proportional valve and the vacuum pump.

[0011] Step 2: Set the drive current output of the temperature control circuit board through the microcomputer to drive the laser to start working, and at the same time control the working temperature of the laser.

[0012] Step 3: Turn on the photodetector to start the measurement, use the acquisition card to collect the direct absorption signal after it is amplified by the photodetector amplifier, and upload it to the microcomputer for normalization processing;

[0013] Step 4: Combine elite reverse learning, the golden sine algorithm, and dynamic nonlinear inertial weights to form an improved whale optimization algorithm;

[0014] Step 5: Utilize the improved whale optimization algorithm to automatically optimize the core hyperparameters of the deep learning framework, thereby reducing the time cost of manually trying to adjust them based on experience;

[0015] Step 6: Separate the normalized overlapping gas absorption spectra using the optimized deep learning framework to ensure the separation accuracy of each gas;

[0016] Step 7: Calibrate the linear relationship between gas concentration and peak absorbance according to Beer-Lambert's law;

[0017] Step 8: Obtain the gas separation concentration value after inversion.

[0018] In a further technical solution, in step 4, elite reverse learning is integrated into the whale pod. During the random initialization process, the search capability and stability of the algorithm are improved according to formulas (1) and (2);

[0019] Assume that the extreme point in the group corresponding to the ordinary individual is the elite individual. Then its reverse solution of A dimensional vector can be represented as:

[0020] (1);

[0021] in, It is an elite individual 3D search space It is an elite reverse solution 3D search space It is a random number in the range [0,1]. , and yes The dynamic boundary of the dimensional search space. If the boundary is exceeded, a random method is used for correction:

[0022] (2);

[0023] in, This represents a random function.

[0024] In a further technical solution, in step 4, the golden sine algorithm is added to the spiral bubble network attack stage of the whale optimization algorithm. Formulas (3), (4) and (5) are used to control the displacement distance and direction of the individual whale, thereby strengthening the communication between the individual whale and the optimal individual.

[0025] When updating their position, individual whales also search their neighborhood. The whale spiral update method, which incorporates the golden sine algorithm, is specifically represented as follows:

[0026] (3);

[0027] in, It is the current iteration number and the displacement distance. It is a random number between [0, 2π], and the displacement direction is... It is a random number in the range [0, π]. This is the best position for the whale at present. This is the current location of the whale. The location of the whale in the next iteration; and By introducing the golden ratio The obtained coefficients are defined as follows:

[0028] (4);

[0029] (5).

[0030] A further technical solution involves setting adaptive dynamic nonlinear inertia weights in step 4 according to formula (6). The adjustment mechanism combines the inverse incomplete gamma function and Cauchy distribution to balance the algorithm's global search and local development capabilities.

[0031] (6);

[0032] in, and Representing the lower and upper bounds of inertia weights, random variables It needs to be greater than or equal to zero. It is the maximum number of iterations. This is the inertia weight adjustment factor, with a value of 0.01. This represents a random number that follows a Cauchy distribution.

[0033] In a further technical solution, in step 5, the improved whale optimization algorithm of step 4 is used to automatically optimize the core hyperparameters of the deep learning framework, and the inherent local optima problem in the traditional whale optimization algorithm is avoided according to formulas (7) to (13).

[0034] The search behavior of an individual whale around the optimal solution is abstracted as a dynamic combination of a shrinking encirclement pattern and a spiral path update pattern. This dual-mode strategy utilizes random probability parameters. Dynamic switching, assuming each mode has a 50% probability, with random probability parameters. It follows a uniform distribution in the range [0,1], if The whale's position update follows the following shrinking and wrapping pattern:

[0035] (7);

[0036] (8);

[0037] (9);

[0038] (10);

[0039] in, and There are two control parameters. It is the convergence factor, which decreases linearly from 2 to 0 during the iteration process. and It is a random number uniformly distributed between [0,1]; if The spiral update equation is expressed as:

[0040] (11);

[0041] When an individual whale fails to find prey close enough during the iteration process, the algorithm switches to a random prey search mode, simulating the behavior of a whale population dispersing to search for other potential prey. At this point, the whale no longer moves around the current best individual, but randomly selects an individual from the population as a reference for the global search. During the random prey search, the position is updated according to the following formula:

[0042] (12);

[0043] (13);

[0044] in, This indicates the location of a whale randomly selected from the current population. express and The distance between them.

[0045] A further technical solution involves constructing a fitness function based on formula (14) in step 6 to evaluate and optimize the trained deep learning framework.

[0046] (14);

[0047] in, This represents the root mean square error. It is the total number of training samples. It is the actual value. This is a predicted value.

[0048] In a further technical solution, in step 7, the concentration of the gas in tunable semiconductor laser absorption spectroscopy (TDLAS) is determined by measuring the intensity of its characteristic absorption line. According to Beer-Lambert law, the intensity of the emitted light... and incident light intensity The following relationship must be satisfied:

[0049] (15);

[0050] in, It is the optical frequency. Indicates gas pressure. It is the absorption path length. Indicates the target gas concentration; absorption coefficient Depends on spectral line intensity and normalized linear function The product of:

[0051] (16);

[0052] When there is spectral overlap between multiple gas components, the total absorption coefficient It is the sum of the absorption coefficients of each component:

[0053] (17);

[0054] in, It is the first The spectral line intensities of the components, It is the first Normalized linear functions of the components.

[0055] This invention provides an overlapping spectrum separation device and method based on a novel evolutionary deep learning model. When separating overlapping spectra, the method uses an improved whale optimization algorithm to balance global search and local exploitation capabilities, improving measurement accuracy while achieving millisecond-level model response time. Furthermore, when the gas composition, concentration, temperature, and pressure change, the corresponding absorption spectrum training dataset can be replaced, providing a new method for the simultaneous measurement of multi-component gases under different environments. Attached Figure Description

[0056] Figure 1 A schematic diagram of an overlapping spectral separation device based on a novel evolutionary deep learning model provided in an embodiment of the present invention;

[0057] Figure 2 A flowchart illustrating an overlapping spectral separation method based on a novel evolutionary deep learning model, provided for an embodiment of the present invention.

[0058] In the attached diagram: 1. Primary temperature sensor; 2. Vacuum pump; 3. Inlet proportional valve; 4. Outlet proportional valve; 5. Pressure sensor; 6. Secondary temperature sensor; 7. Drive temperature control circuit board; 8. Laser; 9. Multi-pass cell; 10. Photodetector; 11. Photodetector amplifier; 12. Tertiary temperature sensor; 13. Internal temperature control device; 14. Microcomputer; 15. Data acquisition card; 16. External temperature control device. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0060] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0061] like Figure 1 As shown, an overlapping spectral separation device based on a novel evolutionary deep learning model is provided in an embodiment of the present invention. It includes a multi-pass cell 9, an internal temperature control device 13, and an external temperature control device 16. The external temperature control device 16 is connected to the inlet pipe and the outlet pipe of the multi-pass cell 9. An inlet proportional valve 3 and a pressure sensor 5 are provided on the inlet pipe. An outlet proportional valve 4 and a vacuum pump 2 are provided on the outlet pipe. A primary temperature sensor 1 is provided on the external temperature control device 16, and a tertiary temperature sensor 12 is provided on the multi-pass cell 9.

[0062] The light inlet of the multi-pass cell 9 is optically connected to the light outlet of the laser 8. The light outlet of the multi-pass cell 9 is optically connected to one end of the photodetector 10. The other end of the photodetector 10 is electrically connected to the photodetector amplifier 11. The photodetector amplifier 11 is also electrically connected to one end of the data acquisition card 15. The other end of the data acquisition card 15 is electrically connected to one end of the microcomputer 14. The other end of the microcomputer 14 is electrically connected to the driving temperature control circuit board 7. The other end of the driving temperature control circuit board 7 is electrically connected to the laser 8.

[0063] The multi-channel cell 9 and the photodetector 10 are both located in the internal temperature control device 13, and the internal temperature control device 13 is equipped with a secondary temperature sensor 6.

[0064] like Figure 2 As shown, an embodiment of the present invention provides an overlapping spectral separation method based on a novel evolutionary deep learning model, which, based on the above-described apparatus, includes the following steps:

[0065] Step 1: Use the primary temperature sensor 1, the secondary temperature sensor 6, and the tertiary temperature sensor 12 to maintain a stable ambient temperature. Turn on the vacuum pump 2, the inlet proportional valve 3, and the outlet proportional valve 4 to fill the multi-pass cell 9 with the gas to be measured. Close the inlet proportional valve 3 and use the pressure sensor 5 to control the pressure inside the multi-pass cell 9. Then, close the outlet proportional valve 4 and the vacuum pump 2.

[0066] Step 2: The microcomputer 14 sets the drive current output of the temperature control circuit board 7 to drive the laser 8 to start working, and at the same time controls the working temperature of the laser 8.

[0067] Step 3: Turn on the photodetector 10 to start the measurement, use the acquisition card 15 to acquire the direct absorption signal amplified by the photodetector amplifier 11, and upload it to the microcomputer 14 for normalization processing.

[0068] Step 4: Combine elite reverse learning, the golden sine algorithm, and dynamic nonlinear inertial weights to form an improved whale optimization algorithm;

[0069] Step 5: Utilize the improved whale optimization algorithm to automatically optimize the core hyperparameters of the deep learning framework, thereby reducing the time cost of manually trying to adjust them based on experience;

[0070] Step 6: Separate the normalized overlapping gas absorption spectra using the optimized deep learning framework to ensure the separation accuracy of each gas;

[0071] Step 7: Calibrate the linear relationship between gas concentration and peak absorbance according to Beer-Lambert's law;

[0072] Step 8: Obtain the gas separation concentration value after inversion.

[0073] In a preferred embodiment of the present invention, in step 4, elite reverse learning is integrated into the whale pod. During the random initialization process, the search capability and stability of the algorithm are improved according to formulas (1) and (2);

[0074] Assume that the extreme point in the group corresponding to the ordinary individual is the elite individual. Then its reverse solution of A dimensional vector can be represented as:

[0075] (1);

[0076] in, It is an elite individual 3D search space It is an elite reverse solution 3D search space It is a random number in the range [0,1]. , and yes The dynamic boundary of the dimensional search space. If the boundary is exceeded, a random method is used for correction:

[0077] (2);

[0078] in, This represents a random function.

[0079] As a preferred embodiment of the present invention, in step 4, the golden sine algorithm is added to the spiral bubble network attack stage of the whale optimization algorithm, and the displacement distance and direction of the individual whale are controlled by formulas (3), (4) and (5), which strengthens the communication between the individual whale and the optimal individual.

[0080] When updating their position, individual whales also search their neighborhood. The whale spiral update method, which incorporates the golden sine algorithm, is specifically represented as follows:

[0081] (3);

[0082] in, It is the current iteration number and the displacement distance. It is a random number between [0, 2π], and the displacement direction is... It is a random number in the range [0, π]. This is the best position for the whale at present. This is the current location of the whale. The location of the whale in the next iteration; and By introducing the golden ratio The obtained coefficients are defined as follows:

[0083] (4);

[0084] (5).

[0085] In a preferred embodiment of the present invention, in step 4, an adaptive dynamic nonlinear inertial weight is set according to formula (6). The adjustment mechanism combines the inverse incomplete gamma function and Cauchy distribution to balance the algorithm's global search and local development capabilities.

[0086] (6);

[0087] in, and Representing the lower and upper bounds of inertia weights, random variables It needs to be greater than or equal to zero. It is the maximum number of iterations. This is the inertia weight adjustment factor, with a value of 0.01. This represents a random number that follows a Cauchy distribution.

[0088] As a preferred embodiment of the present invention, in step 5, the improved whale optimization algorithm of step 4 is used to automatically optimize the core hyperparameters of the deep learning framework, and the local optimum problem inherent in the traditional whale optimization algorithm is avoided according to formulas (7) to (13).

[0089] The search behavior of an individual whale around the optimal solution is abstracted as a dynamic combination of a shrinking encirclement pattern and a spiral path update pattern. This dual-mode strategy utilizes random probability parameters. Dynamic switching, assuming each mode has a 50% probability, with random probability parameters. It follows a uniform distribution in the range [0,1], if The whale's position update follows the following shrinking and wrapping pattern:

[0090] (7);

[0091] (8);

[0092] (9);

[0093] (10);

[0094] in, and There are two control parameters. It is the convergence factor, which decreases linearly from 2 to 0 during the iteration process. and It is a random number uniformly distributed between [0,1]; if The spiral update equation is expressed as:

[0095] (11);

[0096] When an individual whale fails to find prey close enough during the iteration process, the algorithm switches to a random prey search mode, simulating the behavior of a whale population dispersing to search for other potential prey. At this point, the whale no longer moves around the current best individual, but randomly selects an individual from the population as a reference for the global search. During the random prey search, the position is updated according to the following formula:

[0097] (12);

[0098] (13);

[0099] in, This indicates the location of a whale randomly selected from the current population. express and The distance between them.

[0100] In a preferred embodiment of the present invention, in step 6, a fitness function is constructed according to formula (14) to evaluate the optimized deep learning framework:

[0101] (14);

[0102] in, This represents the root mean square error. It is the total number of training samples. It is the actual value. This is a predicted value.

[0103] In a preferred embodiment of the present invention, in step 7, the concentration of the gas in the TDLAS is determined by measuring the intensity of its characteristic absorption line. According to Beer-Lambert law, the intensity of the emitted light... and incident light intensity The following relationship must be satisfied:

[0104] (15);

[0105] in, It is the optical frequency. Indicates gas pressure. It is the absorption path length. Indicates the target gas concentration; absorption coefficient Depends on spectral line intensity and normalized linear function The product of:

[0106] (16);

[0107] When there is spectral overlap between multiple gas components, the total absorption coefficient It is the sum of the absorption coefficients of each component:

[0108] (17);

[0109] in, It is the first The spectral line intensities of the components, It is the first Normalized linear functions of the components.

[0110] As a preferred embodiment of the present invention, taking the measurement of standard gases (4400ppm CH4 and 13ppm CO) at 25°C using the aforementioned device as an example, the implementation process of the above method is illustrated, including the following steps:

[0111] Step 1: Use the primary temperature sensor 1, the secondary temperature sensor 6, and the tertiary temperature sensor 12 to maintain a stable ambient temperature. Turn on the vacuum pump 2, the inlet proportional valve 3, and the outlet proportional valve 4 to fill the multi-pass cell 9 with the CH4 and CO mixture to be measured. Close the inlet proportional valve 3 and use the pressure sensor 5 to control the pressure inside the multi-pass cell 9. Then close the outlet proportional valve 4 and the vacuum pump 2.

[0112] Step 2: Set the drive current output of the temperature control circuit board 7 through the microcomputer 14 to drive the laser 8 to start working, and at the same time control the working temperature of the laser 8 to 25°C.

[0113] Step 3: Turn on the photodetector 10 to start the measurement, use the acquisition card 15 to acquire the direct absorption signal amplified by the photodetector amplifier 11, and upload it to the microcomputer 14 for normalization processing.

[0114] Step 4: Combine the formulas (1) and (2) of elite reverse learning, the formulas (3) to (5) of the golden sine algorithm, and the formula (6) of dynamic nonlinear inertial weight to form an improved whale optimization algorithm.

[0115] Step 5: The core hyperparameters of the deep learning framework are automatically optimized using the improved whale optimization algorithm derived from formulas (7) to (13). This algorithm is configured with a maximum of 50 iterations and 10 search agents, with an optimal learning rate of 0.0054 and a regularization coefficient set to 10. -3 There are 50 nodes in the hidden layer;

[0116] Step 6: Evaluate the optimal deep learning framework after training using formula (14) and separate the normalized overlapping gas absorption spectra;

[0117] Step 7: Calibrate the linear relationship between gas concentration and peak absorbance according to Beer-Lambert's law;

[0118] Step 8: The concentrations of CH4 and CO after inversion are 4415.88 ppm and 12.89 ppm, respectively.

[0119] The separation accuracy of CH4 and CO can reach 99.992% and 99.982%, respectively, which shows that the above method can achieve high-accuracy measurement of multiple gas components simultaneously when separating overlapping spectra.

[0120] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for separating overlapping spectra based on a novel evolutionary deep learning model, characterized in that, Based on the overlapping spectral separation device, the overlapping spectral separation device includes a multi-pass cell, an internal temperature control device, and an external temperature control device. The external temperature control device is connected to the inlet pipe and the outlet pipe of the multi-pass cell. The inlet pipe is equipped with an inlet proportional valve and a pressure sensor, and the outlet pipe is equipped with an outlet proportional valve and a vacuum pump. The external temperature control device is equipped with a primary temperature sensor, and the multi-pass cell is equipped with a tertiary temperature sensor. The light inlet of the multi-pass cell is optically connected to the light outlet of the laser. The light outlet of the multi-pass cell is optically connected to one end of the photodetector. The other end of the photodetector is electrically connected to the photodetector amplifier. The photodetector amplifier is also electrically connected to one end of the data acquisition card. The other end of the data acquisition card is electrically connected to one end of the microcomputer. The other end of the microcomputer is electrically connected to the driving temperature control circuit board. The other end of the driving temperature control circuit board is electrically connected to the laser. Both the multi-pass cell and the photodetector are housed in an internal temperature control device, and a secondary temperature sensor is installed on the internal temperature control device. The method includes the following steps: Step 1: Use the first-level temperature sensor, the second-level temperature sensor and the third-level temperature sensor to maintain the stability of the measurement environment temperature. Turn on the vacuum pump, the inlet proportional valve and the outlet proportional valve to fill the multi-pass cell with the gas to be measured. Close the inlet proportional valve and use the pressure sensor to control the pressure inside the multi-pass cell. Then close the outlet proportional valve and the vacuum pump. Step 2: Set the drive current output of the temperature control circuit board through the microcomputer to drive the laser to start working, and at the same time control the working temperature of the laser. Step 3: Turn on the photodetector to start the measurement, use the acquisition card to collect the direct absorption signal after it is amplified by the photodetector amplifier, and upload it to the microcomputer for normalization processing; Step 4: Combine elite reverse learning, the golden sine algorithm, and dynamic nonlinear inertial weights to form an improved whale optimization algorithm; Step 5: Utilize the improved whale optimization algorithm to automatically optimize the core hyperparameters of the deep learning framework, thereby reducing the time cost of manually trying to adjust them based on experience; Step 6: Separate the normalized overlapping gas absorption spectra using the optimized deep learning framework to ensure the separation accuracy of each gas; Step 7: Calibrate the linear relationship between gas concentration and peak absorbance according to Beer-Lambert's law; Step 8: Obtain the inverted gas separation concentration value; In step 4, elite back-learning is integrated into the whale pod. During the random initialization process, the Golden Sine Algorithm was added to the spiral bubble network attack phase of the Whale Optimization Algorithm; In step 4, the adaptive dynamic nonlinear inertial weights are set according to formula (6). The adjustment mechanism combines the inverse incomplete gamma function and Cauchy distribution to balance the algorithm's global search and local development capabilities. （6）； in, and Representing the lower and upper bounds of inertia weights, random variables It needs to be greater than or equal to zero. It is the maximum number of iterations. This is the inertia weight adjustment factor, with a value of 0.

01. This represents a random number that follows a Cauchy distribution. In step 5, the improved whale optimization algorithm from step 4 is used to automatically optimize the core hyperparameters of the deep learning framework. According to formulas (7) to (13), the inherent local optima problem in the traditional whale optimization algorithm is avoided. The search behavior of an individual whale around the optimal solution is abstracted as a dynamic combination of a shrinking encirclement pattern and a spiral path update pattern. This dual-mode strategy utilizes random probability parameters. Dynamic switching, assuming each mode has a 50% probability, with random probability parameters. It follows a uniform distribution in the range [0,1], if The whale's position update follows the following shrinking and wrapping pattern: （7）； （8）； （9）； （10）； in, and There are two control parameters. It is the convergence factor, which decreases linearly from 2 to 0 during the iteration process. and It is a random number uniformly distributed between [0,1]; if The spiral renewal equation is expressed as: （11）； When an individual whale fails to find prey during the iteration process, the algorithm switches to a random prey search mode, simulating the behavior of a whale population dispersing to search for other potential prey. At this point, the whale no longer moves around the current best individual, but randomly selects an individual from the population as a reference for the global search. During the random prey search, the position is updated according to the following formula: （12）； （13）； in, This indicates the location of a whale randomly selected from the current population. express and The distance between them.

2. The overlapping spectral separation method based on a novel evolutionary deep learning model according to claim 1, characterized in that, In step 4, the search capability and stability of the algorithm are improved according to formulas (1) and (2); Assume that the extreme point in the group corresponding to the ordinary individual is the elite individual. Then its reverse solution of A dimensional vector can be represented as: （1）； in, It is an elite individual 3D search space It is an elite reverse solution 3D search space It is a random number in the range [0,1]. , and yes The dynamic boundary of the dimensional search space; if the boundary is exceeded, a random method is used for correction: （2）； in, This represents a random function.

3. The overlapping spectral separation method based on a novel evolutionary deep learning model according to claim 2, characterized in that, In step 4, formulas (3), (4) and (5) are used to control the displacement distance and direction of the individual whale, thereby enhancing the communication between the individual whale and the optimal individual. When updating their position, individual whales also search their neighborhood. The whale spiral update method, which incorporates the golden sine algorithm, is specifically represented as follows: （3）； in, It is the current iteration number and the displacement distance. It is a random number between [0, 2π], and the displacement direction is... It is a random number in the range [0, π]. This is the best position for the whale at present. This is the current location of the whale. The location of the whale in the next iteration; and By introducing the golden ratio The obtained coefficients are defined as follows: （4）； （5）。 4. The overlapping spectral separation method based on a novel evolutionary deep learning model according to claim 3, characterized in that, In step 6, a fitness function is constructed according to formula (14) to evaluate the optimized deep learning framework: （14）； in, This represents the root mean square error. It is the total number of training samples. It is the actual value. This is a predicted value.

5. The overlapping spectral separation method based on a novel evolutionary deep learning model according to claim 4, characterized in that, In step 7, the concentration of the gas in the TDLAS is determined by measuring the intensity of its characteristic absorption line, according to Beer-Lambert's law, the intensity of the emitted light... and incident light intensity The following relationship must be satisfied: （15）； in, It is the optical frequency. Indicates gas pressure. It is the absorption path length. Indicates the target gas concentration; absorption coefficient Depends on spectral line intensity and normalized linear function The product of: （16）； When there is spectral overlap between multiple gas components, the total absorption coefficient It is the sum of the absorption coefficients of each component: （17）； in, It is the first The spectral line intensities of the components, It is the first Normalized linear functions of the components.

Images