Combustion state inversion method, system and device based on neural network and absorption spectrum
By constructing a neural network model based on U-Net and Inception-ResNet, effective features are directly extracted from the absorption spectrum of an unknown baseline under high pressure, solving the error problem of combustion state parameter inversion under high pressure and realizing stable and accurate combustion state monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU INTERNATIONAL INNOVATION INSTITUTE OF BEIHANG UNIVERSITY
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-09
Smart Images
Figure CN122171470A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of combustion diagnostics and optical measurement technology, specifically to a combustion state inversion method, system, and apparatus based on neural networks and absorption spectroscopy. Background Technology
[0002] High-precision diagnostics of combustion processes are of great significance in fields such as energy, aerospace, and industrial reaction equipment, especially for real-time, in-situ monitoring of combustion temperature, pressure, and component concentration. Tunable diode laser absorption spectroscopy (TDLS) is widely used in combustion diagnostics due to its advantages of being non-contact, highly sensitive, and having a fast response. Its basic principle is based on Beer-Lambert's law; by analyzing the absorption spectrum of the laser beam passing through the gas being tested, parameters such as temperature, pressure, and concentration of the gas can be deduced.
[0003] Currently, most high-precision inversion methods still require relatively accurate baseline information. Due to the severe broadening of gas absorption spectral lines under high pressure, the absorption spectral lines overlap, making it difficult to identify baseline regions without absorption. Once the unreliable baseline under high pressure is directly introduced into the systematic error, the inversion method will have large errors or be difficult to implement under high pressure. Summary of the Invention
[0004] The present invention aims to at least partially solve one of the technical problems in the above-mentioned technologies.
[0005] To this end, this invention utilizes an envelope correction network based on the U-Net structure and a temperature inversion network and a pressure and concentration inversion network based on the Inception-ResNet structure. This allows for the direct extraction of effective absorption features from the original absorption spectrum containing unknown and complex baselines, avoiding dependence on independent baseline measurements or estimates, and making the model extremely robust to baseline fluctuations.
[0006] To solve the above problems, the present invention is implemented using the following technical solution.
[0007] In a first aspect, the present invention provides a combustion state inversion method based on neural networks and absorption spectra, comprising the following steps: Data acquisition and preprocessing steps: Acquire the raw absorption spectrum signal of the combustion region to be tested. (ν), for the original absorption spectrum signal (ν) is obtained through preprocessing. ,pass The apparent absorbance was calculated. Spectral inversion step: The apparent absorbance... The data is input into a pre-trained spectral inversion neural network model, which directly outputs the inverted combustion state parameters, including at least temperature, pressure, and the concentration of the gas component to be measured. The spectral inversion neural network model is an end-to-end model, which includes three sub-networks: An envelope correction network, built on a U-Net architecture, is used to obtain the apparent absorbance from the input. Extract the lower envelope ,pass The envelope-corrected absorbance spectrum was calculated. A temperature inversion network, built on an Inception-ResNet architecture, is used to correct the absorbance spectrum using the normalized envelope. The temperature is output as input; Pressure and concentration inversion network, which is constructed based on the Inception-ResNet structure, is used to correct the absorbance spectrum using the envelope. As input, the pressure and concentration are output.
[0008] Furthermore, the training method for the spectral inversion neural network model includes the following steps: A basic experimental spectral dataset is obtained, which contains absorption spectra measured under different known combustion state parameters. The absorption spectra are then subjected to envelope correction processing to obtain "pure" envelope-corrected spectra and their corresponding lower envelope datasets. Enhance the basic experimental spectral dataset to generate an enhanced training dataset that includes a simulated baseline; The spectral inversion neural network model is trained using a phased training strategy.
[0009] Furthermore, the step of enhancing the basic experimental spectral dataset includes: randomly superimposing one or more envelope lines selected from the lower envelope dataset onto each basic spectrum to form an input spectrum with different baseline morphologies.
[0010] Furthermore, the spectral inversion step supports organizing multiple continuously acquired spectra into batches and inputting them into the spectral inversion neural network model for batch parallel inversion.
[0011] Furthermore, in the data acquisition and preprocessing steps, the preprocessing is background subtraction, where the background is the dark current when the laser is turned off.
[0012] Furthermore, the combustion state parameters in the acquired basic experimental spectral dataset include temperature, pressure, and target wavelength, wherein the temperature is 100-1000°C, the pressure is 1-75 bar, and the target wavelength is 7323 cm⁻¹-7483 cm⁻¹.
[0013] Furthermore, the phased training strategy includes: Train the envelope correction network separately; With the envelope correction network fixed, the temperature inversion network and the pressure and concentration inversion network are trained respectively. The entire spectral inversion neural network model is jointly fine-tuned.
[0014] Secondly, the present invention provides a combustion state inversion system based on neural networks and absorption spectra, comprising the following modules: Spectral acquisition module: Used to acquire the raw absorption spectral signal of the combustion region under test. (ν); Data preprocessing module: used for processing the raw absorption spectrum signal (ν) is obtained by dark current subtraction. ,pass The apparent absorbance was calculated and processed into a 233-dimensional fixed-length vector using a bilinear interpolation algorithm. Neural network computing module: Stores and runs the spectral inversion neural network model trained by the first aspect method, receives preprocessed vector data, and outputs estimated values of temperature, pressure and concentration after forward propagation; Results output module: used to display and record the inversion parameters corresponding to single or batch spectra in real time, supporting closed-loop control applications of combustion state.
[0015] Furthermore, it also includes a model training module for executing a training method for the spectral inversion neural network model to generate or update the spectral inversion neural network model.
[0016] Thirdly, the present invention provides a combustion state inversion device based on neural networks and absorption spectra, comprising: The laser emitting and receiving unit consists of a tunable diode laser, a photodetector, and a data acquisition card. It is used to emit scanning lasers toward the combustion area to be tested and to receive the transmitted or reflected laser signals. The signal processing unit is used to convert the received laser signal into an absorption spectrum signal. ,pass The apparent absorbance was calculated. The computational control unit stores or can load a pre-trained spectral inversion neural network model for receiving the surface absorbance. It then performs inversion calculations and outputs combustion state parameters.
[0017] The advantages of this invention compared to existing technologies are: 10. This invention provides a combustion state inversion method, system, and apparatus based on neural networks and absorption spectra. By using an envelope correction network based on a U-Net structure and a temperature inversion network and a pressure and concentration inversion network based on an Inception-ResNet structure, it is possible to directly extract effective absorption features from the original absorption spectrum containing unknown and complex baselines, avoiding dependence on independent baseline measurements or estimates, and making the model extremely robust to baseline fluctuations. 11. This invention provides a combustion state inversion method, system, and apparatus based on neural networks and absorption spectra. The neural network, by learning from a large amount of experimental data, can adaptively compensate for parameter errors in spectral databases (such as HITRAN, HITEMP, BT2, etc.) under high pressure. 12. This invention provides a combustion state inversion method, system, and apparatus based on neural networks and absorption spectra. It adopts an architecture design that separates the temperature inversion network from the pressure and concentration inversion networks. This makes the temperature inversion mainly rely on the line intensity ratio information, which is not sensitive to broadening, while the pressure and concentration inversion focuses on utilizing the broadening and absolute absorption intensity information. This decoupling design reduces the coupling interference between parameters. Combined with the global optimization characteristics of neural networks, it effectively avoids the problems of sensitivity to initial values and easy divergence in traditional methods, ensuring the stability of the inversion process and the accuracy of the results. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the structure of the spectral inversion neural network of the present invention.
[0019] Figure 2 This is a structural diagram of the envelope correction network constructed based on U-Net in this invention.
[0020] Figure 3 This is a structural diagram of the temperature inversion network and pressure and concentration inversion network constructed based on Inception-ResNet in this invention.
[0021] Figure 4 yes Figure 3 The structure diagram of Inception_A.
[0022] Figure 5 This is a comparison chart of the effects of traditional envelope correction methods and neural network envelope correction methods.
[0023] Figure 6 This is a statistical distribution and linear fitting graph of the inversion results of this invention on a large test set. Detailed Implementation
[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.
[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0026] Example 1 like Figure 1 The embodiment shown provides a combustion state inversion method based on neural networks and absorption spectra, including the following steps: Step 1: Data Acquisition and Preprocessing: Acquire the raw absorption spectrum signal of the combustion region to be tested. (ν), for the original absorption spectrum signal (ν) is obtained through preprocessing. ,pass Calculate the apparent absorbance.
[0027] Furthermore, in the data acquisition and preprocessing steps, the preprocessing step involves background subtraction, where the background is the dark current when the laser is turned off.
[0028] It should be noted that in the high-pressure combustion experimental setup, a tunable diode laser is deployed, whose scanning range covers the absorption wave of water vapor from 7323 cm⁻¹ to 7483 cm⁻¹. After the laser passes through the combustion area under test, it is received by a photodetector and converted into a digital signal by a data acquisition card to obtain the original transmitted light intensity sequence. (ν), to eliminate common-mode interference such as laser power fluctuations, simple background subtraction can be performed (e.g., subtracting the dark current when the laser is turned off) to obtain the relative light intensity. ,pass The apparent absorbance is calculated. This spectrum is the raw data containing baseline, noise, and gas absorption information. No further baseline fitting or filtering is required. The spectral data is resampled or interpolated into a fixed-length vector using a bilinear interpolation algorithm so that it can be input into the neural network model.
[0029] To clearly illustrate the previous embodiment, in one embodiment of the present invention, as follows: Figures 2 to 4As shown, step 2, the apparent absorbance The data is input into a pre-trained spectral inversion neural network model, which directly outputs the inverted combustion state parameters, including at least temperature T, pressure P, and the concentration X of the gas component to be measured.
[0030] The construction of the spectral inversion neural network model: The spectral inversion neural network model is an end-to-end model, which includes three sub-networks: Envelope correction networks, built on the U-Net architecture, are used to adjust the apparent absorbance of the input. Extract the lower envelope Through formula The envelope-corrected absorbance spectrum was calculated.
[0031] It should be noted that the envelope correction network uses, for example, Figure 2 The U-Net structure shown has an input layer that receives 233-dimensional data. The vector is expanded to 256 channels by a fully connected layer, followed by a 3-layer U-Net encoder-decoder structure, and finally restored to 233 channels by a fully connected layer. The output is a lower envelope estimate with the same dimension as the input. .
[0032] Specifically, each layer of the encoding part contains two one-dimensional convolutional layers with a kernel size of 3 (followed by ReLU activation) and a max pooling layer with a 2x downsampling; each layer of the decoding part contains a deconvolutional layer with a 2x upsampling, concatenation with the feature map of the corresponding encoding layer, and two one-dimensional convolutional layers. All of the above convolutional layers use ReLU activation.
[0033] The temperature inversion network, built on the Inception-ResNet architecture, is used to correct absorbance spectra using the normalized envelope. As input, the output temperature T.
[0034] It should be noted that the temperature inversion network uses, for example... Figure 3 The Inception-ResNet variant shown has an output layer dimension of 1, and the input is a pair. The vector after maximum value normalization, i.e., each element divided by the maximum value of the vector, is obtained by using the formula for maximum value normalization: ,in, It is an envelope-corrected spectrum. This is the maximum value of the envelope-corrected spectrum. The formula for calculating the envelope-corrected spectrum is: ,in, It measures the spectrum. It is the lower envelope of the measured spectrum, and the formula for calculating the measured spectrum is: ,in, Apparent absorbance, Minimum apparent absorbance Maximum apparent absorbance.
[0035] Specifically, the Inception-ResNet network architecture includes: an initial convolutional layer, two cascaded Inception_A modules, and three fully connected layers. The Inception_A module structure is as follows: Figure 3 As shown, ReLU activation is used in both the Inception_A module and the fully connected layers, ultimately outputting the inversion temperature. .
[0036] Pressure and concentration inversion network, built on the Inception-ResNet architecture, is used to correct absorbance spectra with unnormalized envelopes. As input, the outputs are pressure P and concentration X.
[0037] It should be noted that the pressure and concentration inversion network uses, for example... Figure 3 The Inception-ResNet variant shown has an output layer dimension of 2, but the input is an undefined layer. After maximum value normalization, the Inception-ResNet network structure in the pressure and concentration inversion network is the same as the Inception-ResNet network structure in the temperature inversion network mentioned above, and the final output is the inverted pressure. and concentration .
[0038] Specifically, such as Figure 5 and Figure 6 As shown, by using the envelope correction network based on the U-Net structure and the temperature inversion network and pressure and concentration inversion network based on the Inception-ResNet structure, effective absorption features can be extracted directly from the original absorption spectrum containing unknown and complex baselines, avoiding dependence on independent baseline measurements or estimates, and making the model extremely robust to baseline fluctuations.
[0039] To clearly illustrate the above embodiment, in one embodiment of the present invention, the training method for the spectral inversion neural network model includes the following steps: Obtain the basic experimental spectral dataset, which contains absorption spectra measured under different known combustion state parameters. Perform envelope correction on the absorption spectra to obtain "pure" envelope-corrected spectra and their corresponding lower envelope datasets.
[0040] Furthermore, the combustion state parameters obtained from the basic experimental spectral dataset include temperature, pressure, and target wavelength range: temperature 100-1000°C, pressure 1-75 bar, and target wavelength range 7323 cm⁻¹-7483 cm⁻¹.
[0041] It should be noted that, using a high-pressure combustion experimental apparatus, under controlled temperature (100-1000°C) and pressure (1-75 bar) conditions, the absorption spectra of water vapor in the target wavelength band were collected, resulting in a total of 7903 sets of basic experimental spectra. Each set of data contains multiple repeated measurements. The detailed structure of the dataset is shown below: For each experimental spectrum in the basic dataset, the corresponding "true" envelope-corrected spectrum is obtained by processing it using the traditional difference envelope correction method. and lower envelope All and their corresponding experimental measurement parameters This constitutes the basic training sample library.
[0042] To supplement the data and compensate for the lack of experimental concentration points, linear interpolation was performed on measured spectra at the same pressure and temperature but different concentrations to generate virtual spectra of intermediate concentrations. This expanded the concentration coverage of the sample library to 0.2%. 40%.
[0043] Furthermore, the basic experimental spectral dataset is enhanced to generate an enhanced training dataset that includes simulated baselines.
[0044] Specifically, the steps to enhance the basic experimental spectral dataset include: randomly superimposing one or more envelope lines selected from the lower envelope dataset onto each basic spectrum to form an input spectrum with different baseline shapes.
[0045] It should be noted that during the training phase of the envelope neural network dataset, for each training cycle, for each basic envelope correction spectral sample... Perform the following operations to generate the input data: Randomly generate an integer ; Envelope from all bases From the set, randomly select n envelopes ; pass Calculate the average envelope; pass Generate an input spectrum with baseline, where To add tiny Gaussian noise, These are the spectral lines after correction of the true spectral envelope. These are the training labels for the envelope neural network.
[0046] The above method can generate hundreds of millions of unique "baseline-spectrum" combinations in each training cycle, greatly enhancing the generalization ability of the envelope neural network.
[0047] Specifically, by learning from a large amount of experimental data, the neural network can adaptively compensate for parameter errors in spectral databases (such as HITRAN, HITEMP, BT2, etc.) under high pressure, reduce the sensitivity of the inversion results to parameter errors in existing spectral databases, and improve the inversion accuracy under high pressure conditions, especially the inversion accuracy of pressure and concentration. Experiments show that the present invention is significantly better than the traditional fitting method based on the same database in terms of pressure and temperature inversion accuracy. This is a direct result of the temperature inversion network and the pressure and concentration inversion network learning the real physical mapping relationship from the data.
[0048] Furthermore, a phased training strategy is adopted to train the spectral inversion neural network model. The phased training strategy includes: Train the envelope correction network separately; A fixed envelope correction network was used to train the temperature inversion network and the pressure and concentration inversion network, respectively. The entire spectral inversion neural network model is jointly fine-tuned.
[0049] It should be noted that the envelope correction network is trained separately using the dynamically augmented dataset, and the percentage mean squared error is calculated using the loss function, which is: ,in The envelope output of the envelope neural network is used for training with the Adam optimizer and an initial learning rate lr0 set to [value missing]. A stepwise exponential decay strategy with a learning rate lr was adopted, and training was performed until the loss on the validation set converged. This training strategy is expressed by the following function: The temperature inversion network and the pressure and concentration inversion network are trained independently. First, the parameters of the pre-trained envelope correction network are frozen. Then, the envelope correction network is connected to the initialized temperature and pressure / concentration inversion networks. During the temperature inversion network training, the input is the normalized spectrum output from the envelope correction network, labeled as follows. In the pressure and concentration inversion network, the input is the unnormalized spectrum from the envelope correction network, labeled as... The percentage mean squared error is calculated using the loss function. The loss function of the temperature inversion network is as follows: ,in, The inversion result output by the temperature inversion network is given by the loss function of the pressure and concentration inversion networks. ,in, The pressure inversion result is output by the pressure and concentration inversion network. The concentration result is obtained from the concentration inversion output of the pressure and concentration inversion network.
[0050] The initial learning rate lr0 for both the temperature inversion network and the pressure and concentration inversion network is 1. Magnitude.
[0051] The entire spectral inversion neural network model is jointly fine-tuned by combining the trained envelope correction network, temperature inversion network, and pressure and concentration inversion network, and unlocking all network parameters using the complete original input. and triplet tags Fine-tuning is performed. The loss function is as follows: and The weighted sum. A learning rate of... Training a small number of epochs can smooth the interfaces between subnetworks and improve overall performance.
[0052] By adopting a separate architecture for temperature inversion network and pressure and concentration inversion network, temperature inversion mainly relies on the line intensity ratio information that is insensitive to broadening, while pressure and concentration inversion focus on utilizing broadening and absolute absorption intensity information. This decoupled design reduces coupling interference between parameters. Combined with the global optimization characteristics of neural networks, it effectively avoids the problems of sensitivity to initial values and easy divergence in traditional methods, ensuring the stability of the inversion process and the accuracy of the results.
[0053] To clearly illustrate the previous embodiment, in one embodiment of the present invention, the spectral inversion step supports organizing multiple continuously acquired spectra into batches and inputting them into the spectral inversion neural network model for batch parallel inversion.
[0054] By employing a feedforward neural network for computation, a single inversion involves only one forward propagation process, eliminating the need for iterative optimization. This reduces the inversion time for a single spectrum from seconds (approximately 10 seconds) using traditional methods to milliseconds (<10 ms), and supports batch processing, providing a feasible technical solution for real-time, online monitoring of combustion processes. This achievement is attributed to transforming a complex nonlinear fitting problem into efficient neural network feedforward computation.
[0055] Example 2 This embodiment provides a combustion state inversion system based on neural networks and absorption spectroscopy, including the following modules: Spectral acquisition module: Used to acquire the raw absorption spectral signal of the combustion region under test. (ν); Data preprocessing module: used for processing the raw absorption spectrum signal (ν) is obtained by dark current subtraction. ,pass The apparent absorbance was calculated and processed into a 233-dimensional fixed-length vector using a bilinear interpolation algorithm. Neural network computing module: Stores and runs the spectral inversion neural network model trained by the above training method, receives preprocessed vector data, and outputs estimated values of temperature T, pressure P and concentration X through forward propagation; Results output module: used to display and record the inversion parameters corresponding to single or batch spectra in real time, supporting closed-loop control applications of combustion state.
[0056] Furthermore, in the data acquisition and preprocessing steps, the preprocessing step involves background subtraction, where the background is the dark current when the laser is turned off.
[0057] Example 3 This embodiment provides a combustion state inversion device based on neural networks and absorption spectroscopy, including a laser emitting and receiving unit, a signal processing unit, and a signal processing unit. The laser emitting and receiving unit consists of a tunable diode laser, a photodetector, and a data acquisition card, used to emit scanning laser light into the combustion area to be tested and receive the transmitted or reflected laser signal. The signal processing unit is used to convert the received laser signal into an absorption spectral signal. (ν) and perform dark current subtraction to obtain ,pass The apparent absorbance is calculated, and the calculation control unit stores or can load the spectral inversion neural network model trained by the above training method, which is used to receive the apparent absorbance. It then performs inversion calculations and outputs combustion state parameters.
[0058] It should be noted that by organizing multiple continuously acquired spectra into batches and inputting them into the spectral inversion neural network model, and utilizing the parallel computing capabilities of the neural network framework, the inversion parameter matrix corresponding to all spectra can be output at once, which greatly improves the throughput and is suitable for high-speed monitoring scenarios.
[0059] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
[0061] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A combustion state inversion method based on neural networks and absorption spectroscopy, characterized in that, Includes the following steps: Data acquisition and preprocessing steps: acquire the original absorption spectrum signal of the combustion area to be tested, perform background subtraction on the original absorption spectrum signal to obtain the preprocessed spectrum signal, and calculate the apparent absorbance from the preprocessed spectrum signal; Spectral inversion step: The apparent absorbance is input into a pre-trained spectral inversion neural network model. The spectral inversion neural network model directly outputs the inverted combustion state parameters, which include temperature, pressure and concentration of the gas component to be measured. The spectral inversion neural network model is an end-to-end model, which includes three sub-networks: An envelope correction network, which is built based on a U-Net structure, is used to extract the lower envelope from the input apparent absorbance and calculate the envelope-corrected absorbance spectrum using the lower envelope. A temperature inversion network, which is constructed based on the Inception-ResNet structure, takes the normalized envelope-corrected absorbance spectrum as input and outputs the temperature; A pressure and concentration inversion network, which is constructed based on the Inception-ResNet structure, takes the envelope-corrected absorbance spectrum as input and outputs the pressure and the concentration.
2. The combustion state inversion method based on neural networks and absorption spectroscopy according to claim 1, characterized in that: In the data acquisition and preprocessing steps, the background is the dark current when the laser is off, and the apparent absorbance calculation formula is: ,in, To preprocess the spectral signal, The apparent absorbance is calculated using the envelope-corrected absorbance spectral formula as follows: ,in, This is the lower envelope.
3. The combustion state inversion method based on neural networks and absorption spectroscopy according to claim 2, characterized in that, The training method for the spectral inversion neural network model includes the following steps: A basic experimental spectral dataset is obtained, which contains absorption spectra measured under different known combustion state parameters. The absorption spectra are then subjected to envelope correction processing to obtain the envelope-corrected spectra and their corresponding lower envelope datasets. Enhance the basic experimental spectral dataset to generate an enhanced training dataset that includes a simulated baseline; The spectral inversion neural network model is trained using a phased training strategy.
4. The combustion state inversion method based on neural networks and absorption spectra according to claim 3, characterized in that: The combustion state parameters in the acquired basic experimental spectral dataset include temperature, pressure, and target wavelength. The temperature is 100-1000°C, the pressure is 1-75 bar, and the target wavelength is 7323 cm⁻¹-7483 cm⁻¹.
5. The combustion state inversion method based on neural networks and absorption spectra according to claim 4, characterized in that: The step of enhancing the basic experimental spectral dataset includes: randomly superimposing one or more envelope lines selected from the lower envelope dataset onto each basic spectrum to form an input spectrum with different baseline shapes.
6. The combustion state inversion method based on neural networks and absorption spectroscopy according to claim 5, characterized in that, The phased training strategy includes: Train the envelope correction network separately; With the envelope correction network fixed, the temperature inversion network and the pressure and concentration inversion network are trained respectively. The entire spectral inversion neural network model is jointly fine-tuned.
7. The combustion state inversion method based on neural networks and absorption spectroscopy according to claim 1, characterized in that: The spectral inversion step supports organizing multiple continuously acquired spectra into batches and inputting them into the spectral inversion neural network model for batch parallel inversion.
8. A combustion state inversion system based on neural networks and absorption spectroscopy, characterized in that, Includes the following modules: Spectral acquisition module: used to acquire the raw absorption spectral signal of the combustion region under test; Data preprocessing module: used to perform dark current subtraction on the original absorption spectrum signal to obtain a preprocessed spectrum signal, calculate the apparent absorbance from the preprocessed spectrum signal, and process it into a 233-dimensional fixed-length vector using a bilinear interpolation algorithm; Neural network computing module: stores and runs the spectral inversion neural network model trained by the method of any one of claims 1-7, receives preprocessed vector data, and outputs estimated values of temperature, pressure and concentration through forward propagation; Results output module: used to display and record the inversion parameters corresponding to single or batch spectra in real time, supporting closed-loop control applications of combustion state.
9. The combustion state inversion system based on neural networks and absorption spectroscopy according to claim 8, characterized in that: It also includes a model training module for executing the training method of the spectral inversion neural network model according to claim 6, so as to generate or update the spectral inversion neural network model.
10. A combustion state inversion device based on neural networks and absorption spectroscopy, characterized in that, include: The laser emitting and receiving unit consists of a tunable diode laser, a photodetector, and a data acquisition card. It is used to emit scanning lasers toward the combustion area to be tested and to receive the transmitted or reflected laser signals. The signal processing unit is used to convert the received laser signal into an absorption spectrum signal and perform dark current subtraction to obtain a preprocessed spectrum signal, and calculate the apparent absorbance from the preprocessed spectrum signal. The calculation control unit stores or can load a spectral inversion neural network model trained by the method according to any one of claims 1-7, for receiving the surface absorbance and performing inversion calculations, and outputting combustion state parameters.