TDLAS dual-gas concentration synchronous calibration and decoupling method in presence of spectral aliasing interference

Abstract

The present application relates to the technical field of laser spectrum gas detection, and particularly discloses a TDLAS double-gas concentration synchronous calibration and decoupling method in the presence of spectral aliasing interference, which comprises the following steps: acquiring a multi-frequency band harmonic signal optimized adaptively while collecting environmental parameters, and determining an optimal frequency band signal based on a signal quality index; calculating a spectral aliasing coefficient according to the characteristics of the optimal frequency band signal, dynamically adjusting the operation parameters of a search algorithm based on the numerical interval of the spectral aliasing coefficient, and determining an initial range of double-gas concentration in a search space. The present application combines an improved genetic algorithm based on the dynamic adjustment of the search strategy of the spectral aliasing coefficient with an incremental support vector regression model, realizes the hierarchical processing from coarse calibration to fine decoupling, can automatically balance the calculation efficiency and search accuracy when the degree of spectral aliasing changes, and can online correct the model in combination with the aging characteristics.

Description

Technical Field

[0001] This invention relates to the field of laser spectroscopy gas detection technology, specifically a method for simultaneous calibration and decoupling of TDLAS dual gas concentrations in the presence of spectral aliasing interference. Background Technology

[0002] Tunable semiconductor laser absorption spectroscopy (TDLAS) technology is widely used in industrial process control and environmental monitoring due to its high selectivity and fast response characteristics. However, when detecting multiple mixed gases in complex industrial environments, the absorption lines of different gases often overlap to varying degrees, resulting in severe spectral aliasing interference.

[0003] Existing decoupling algorithms typically employ static multiple linear regression or machine learning models with fixed parameters. While these methods perform reasonably well in ideal laboratory environments, in real-world conditions, static models cannot adapt to dynamic changes in aliasing caused by significant fluctuations in gas concentration ratios, or to center wavelength drift or intensity attenuation due to laser aging. This results in a sharp decline in decoupling accuracy, making it difficult to distinguish whether signal changes are caused by changes in gas concentration or by equipment aging or spectral interference.

[0004] More seriously, in real-world high-dynamic monitoring scenarios, existing detection systems face irreconcilable conflicts between the time and frequency domains:

[0005] On the one hand, in order to suppress electromagnetic noise in industrial sites, algorithms are usually configured with large filtering parameters or regression models with high penalty factors. While this ensures the stability of readings under steady state, it leads to a sluggish response of the system to step changes such as gas leaks, resulting in serious safety lag.

[0006] On the other hand, the TDLAS optical path senses pressure at the speed of light, while physical sensors used for environmental compensation, such as temperature and air pressure sensors, are limited by thermal inertia and mechanical structure, with only a response time in the order of seconds. When the monitored environment experiences rapid decompression or heating, the system may incorrectly use lagging environmental parameters to compensate for the real-time spectral signal. This spatiotemporal mismatch of the sensor can cause the compensation model to fail, resulting in huge spurious concentration spikes at the moment of change in operating conditions. Summary of the Invention

[0007] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a method for simultaneous calibration and decoupling of TDLAS dual-gas concentrations in the presence of spectral aliasing interference, thereby improving the real-time performance and accuracy of dual-gas concentration detection in complex and variable industrial environments.

[0008] To achieve the above objectives, a first aspect of the present invention proposes a method for simultaneous calibration and decoupling of TDLAS dual-gas concentrations in the presence of spectral aliasing interference, comprising the following steps:

[0009] The signal acquisition process involves acquiring environmental parameters, obtaining adaptively optimized multi-band harmonic signals, and selecting the optimal frequency band signal based on signal quality indicators.

[0010] Perform the coarse calibration step: calculate the spectral aliasing coefficient based on the characteristics of the optimal frequency band signal, dynamically adjust the running parameters of the search algorithm based on the numerical range of the spectral aliasing coefficient, and determine the initial range of the dual gas concentrations within the search space;

[0011] Perform the fine decoupling step: Construct an input matrix containing the initial range, the features of the optimal frequency band signal, and the equipment aging features. After dynamically determining the dimensionality reduction dimension based on the spectral aliasing coefficient, map the input matrix to a low-dimensional space and perform regression analysis to obtain the fine decoupling concentration.

[0012] Perform compensation and correction steps: dynamically compensate the fine decoupling concentration based on the environmental parameters, and output the final accurate concentration of the two gases;

[0013] The spectral aliasing coefficient represents the degree of attenuation of the current signal quality relative to the reference signal quality.

[0014] The operating parameters of the dynamically adjusted search algorithm include: when the spectral aliasing coefficient indicates an increased degree of aliasing, increasing the number of iterations of the search algorithm and reducing the mutation probability.

[0015] To achieve the above objectives, a second aspect of the present invention proposes a TDLAS dual-gas concentration synchronous calibration and decoupling system in the presence of spectral aliasing interference, the system comprising:

[0016] The signal acquisition module is used to acquire adaptively optimized multi-band harmonic signals while collecting environmental parameters, and to select the optimal frequency band signal based on signal quality indicators.

[0017] The coarse calibration module is used to calculate the spectral aliasing coefficient based on the characteristics of the optimal frequency band signal, dynamically adjust the running parameters of the search algorithm based on the numerical range of the spectral aliasing coefficient, and determine the initial range of the dual gas concentrations within the search space.

[0018] The fine decoupling module is used to construct an input matrix containing the initial range, the features of the optimal frequency band signal, and the equipment aging features. After dynamically determining the dimensionality reduction dimension based on the spectral aliasing coefficient, the input matrix is ​​mapped to a low-dimensional space and regression analysis is performed to obtain the fine decoupling concentration.

[0019] The compensation and correction module is used to dynamically compensate the fine decoupling concentration based on the environmental parameters and output the final accurate concentration of the two gases.

[0020] The spectral aliasing coefficient represents the degree of attenuation of the current signal quality relative to the reference signal quality; the dynamic adjustment of the search algorithm's operating parameters includes: when the spectral aliasing coefficient indicates an increase in the degree of aliasing, increasing the number of iterations of the search algorithm and reducing the mutation probability.

[0021] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described method for simultaneous calibration and decoupling of TDLAS dual gas concentrations in the presence of spectral aliasing interference.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] The TDLAS dual-gas concentration synchronous calibration and decoupling method in the presence of spectral aliasing interference, as described in this invention, effectively solves the technical problems in the background art through a multi-dimensional dynamic adaptive mechanism.

[0024] First, by combining an improved genetic algorithm based on a dynamic adjustment search strategy using spectral aliasing coefficients with an incremental support vector regression model, a hierarchical processing from coarse calibration to fine decoupling is achieved. This can automatically balance computational efficiency and search accuracy when the degree of spectral aliasing changes, and online correction of the model is performed by combining aging characteristics, which significantly extends the maintenance-free cycle of the equipment.

[0025] More importantly, this invention successfully resolves the contradiction between sensitivity and stability by employing a dual-core co-evolution strategy and intelligently switching between sensitive and steady-state kernels. This allows for the filtering out of steady-state noise and the capture of concentration mutations within seconds. Simultaneously, it introduces a sensor hysteresis correction mechanism based on spectral line shape inversion. By utilizing the physical characteristics of the spectral signal itself as a zero-delay virtual sensor, it eliminates the transient compensation error caused by the response hysteresis of the physical sensor, ensuring the real-time performance and accuracy of dual-gas concentration detection in complex and ever-changing industrial environments. Attached Figure Description

[0026] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0027] Figure 1 This is a flowchart illustrating the TDLAS dual-gas concentration synchronous calibration and decoupling method provided by the present invention when spectral aliasing interference exists.

[0028] Figure 2 This is a comparison chart of the adaptive fusion effect of multi-band harmonic signals under different signal-to-noise ratio conditions in the TDLAS dual-gas concentration synchronous calibration and decoupling method provided by this invention when there is spectral aliasing interference;

[0029] Figure 3 This is a graph showing the evolution of spectral aliasing and the dynamic change of the aliasing coefficient K value in the TDLAS dual-gas concentration synchronous calibration and decoupling method provided by this invention when spectral aliasing interference exists.

[0030] Figure 4 This is a comparison of the search convergence trajectories of the improved genetic algorithm based on adaptive adjustment of the aliasing coefficient in the TDLAS dual-gas concentration synchronous calibration and decoupling method provided by this invention when there is spectral aliasing interference.

[0031] Figure 5 This is a visualization of the LLE manifold dimensionality reduction and feature clustering effect of the TDLAS dual-gas concentration synchronous calibration and decoupling method provided by this invention when there is spectral aliasing interference;

[0032] Figure 6 The diagram shows the effect of wavelength drift caused by laser aging on the measurement results and the adaptive calibration correction effect in the TDLAS dual-gas concentration synchronous calibration and decoupling method provided by this invention when there is spectral aliasing interference.

[0033] Figure 7 This is a time-domain response deviation diagram of the physical sensor readings and the instantaneous pressure obtained from the spectral inversion during the rapid depressurization process in the TDLAS dual-gas concentration synchronous calibration and decoupling method provided by this invention when there is spectral aliasing interference;

[0034] Figure 8 This is a schematic diagram illustrating the implementation of the TDLAS dual-gas concentration synchronous calibration and decoupling system provided by the present invention in the presence of spectral aliasing interference;

[0035] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0036] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0037] The following description, with reference to the accompanying drawings, describes a method, system, and electronic device for simultaneous calibration and decoupling of TDLAS dual gas concentrations in the presence of spectral aliasing interference, according to embodiments of the present invention.

[0038] Example 1:

[0039] This embodiment details a method for simultaneous calibration and decoupling of dual gas concentrations in TDLAS (Tunable Laser Absorption Spectroscopy) under spectral aliasing interference. The method in this embodiment aims to solve the problems of dynamic fluctuations in the degree of spectral aliasing caused by changes in gas component concentration and measurement drift caused by device aging when facing complex industrial environments. By constructing a multi-dimensional, dynamically adaptive calibration and decoupling closed-loop system, high-precision gas concentration detection is achieved.

[0040] Specifically, the hardware foundation of this embodiment is a high-precision TDLAS detection system. This system mainly includes a tunable laser, a dual-bandpass wavelength selector, a mid-infrared detector, a signal acquisition and processing unit, and an environmental monitoring sensor array. The tunable laser serves as the light source, emitting a laser beam that covers the characteristic absorption peaks of the two gases being measured. The dual-bandpass wavelength selector is placed in the optical path to filter background noise and assist in wavelength calibration. The mid-infrared detector receives the optical signal after absorption by the gas and converts it into an electrical signal.

[0041] like Figure 1 As shown, the method in this embodiment includes the following steps:

[0042] S1: Perform the signal acquisition step.

[0043] In this embodiment, signal acquisition is not just a simple acquisition of data, but also includes synchronous perception of environmental parameters and adaptive optimization and selection of spectral signals, which is the basis for subsequent high-precision decoupling.

[0044] Specifically, after startup, the TDLAS detection system first enters the signal acquisition phase. The system controls the environmental monitoring sensor array to collect environmental parameters while simultaneously processing the spectral signals. These environmental parameters are multi-dimensional, including at least the temperature, pressure, humidity, and flow rate of the gas environment being measured. The acquisition frequency of these parameters is strictly synchronized with the scanning frequency of the spectral signals to ensure the time alignment of the subsequent compensation model.

[0045] For example, in acquiring adaptively optimized multi-band harmonic signals, the system does not use fixed driving parameters, but instead introduces device aging characteristics as feedback variables. These device aging characteristics include at least the drive current-wavelength mapping deviation, the wavelength selection device feature point offset, and the signal-to-noise ratio attenuation. The system first reads the device aging characteristics stored in the timing sample pool or register, particularly the recorded drive current-wavelength mapping deviation, which reflects the shift or change in slope of the laser's output wavelength response curve as a function of the injected current after long-term operation.

[0046] Specifically, based on the deviation between the driving current and wavelength recorded in the device aging characteristics, the system dynamically corrects the scanning range of the laser driving current. For example, if the deviation indicates a blue shift in the wavelength, the system will automatically increase the DC bias of the driving current to ensure that the laser scanning range once again covers the center of the characteristic absorption peak of the gas under test, thereby avoiding signal truncation or peak loss due to wavelength drift. After correcting the driving parameters, the system synchronously extracts signals from different harmonic frequency bands through a lock-in amplifier or a digital quadrature demodulation algorithm, typically including the third, fourth, and fifth harmonic signals.

[0047] It is also important to note that, in order to obtain the optimal signal-to-noise ratio (SNR) under different operating conditions, the system needs to select the optimal frequency band signal based on signal quality indicators. The system calculates the product of the SNR and correlation coefficient for each harmonic frequency band signal, and defines this product as the comprehensive quality indicator. Here, the SNR refers to the ratio of the peak-to-peak value of the signal to the standard deviation of the baseline noise, and the correlation coefficient refers to the Pearson correlation coefficient between the currently acquired waveform and the standard reference waveform; where peak-to-peak value represents the difference between the positive maximum value and the negative minimum value of the signal waveform.

[0048] For example, let the signal-to-noise ratio of the i-th harmonic be... The correlation coefficient is Then the overall quality index of this frequency band The calculation is as follows:

[0049] ;

[0050] The system will calculate the frequency bands 3f, 4f, and 5f in real time. value.

[0051] Optionally, the selection of the optimal frequency band signal follows the following logical judgment strategy: If the signal-to-noise ratio (SNR) and correlation coefficient of a single frequency band signal, such as the 4f signal, both exceed a preset validity threshold (e.g., SNR greater than 30 dB and correlation coefficient greater than 0.95), the system directly selects this single frequency band signal as the optimal frequency band signal, because the signal quality is sufficient to support high-precision concentration inversion, and the computational load for single-band processing is relatively small. If the SNR fluctuation of a single frequency band signal exceeds a preset stability threshold (e.g., fluctuation exceeding 10% over 10 consecutive periods), this usually indicates the presence of electromagnetic interference or optical path jitter at a specific frequency. In this case, the system will activate a frequency band fusion mechanism, that is, fusing the peak-to-peak values ​​of the two frequency band signals (e.g., 3f and 4f) with the highest overall quality index, and using weighted averaging or Kalman filtering, the fused signal is selected as the optimal frequency band signal. This method utilizes the differences in noise sensitivity among different frequency bands, effectively improving the robustness of the signal.

[0052] like Figure 2 The figure shows a comparison of the adaptive fusion effect of multi-band harmonic signals. The horizontal axis represents the scanning time of the laser in milliseconds, and the vertical axis represents the voltage amplitude of the acquired harmonic signals in volts.

[0053] Figure 2 The blue dashed curve represents a single high-sensitivity frequency band signal collected under strong interference in an industrial environment. It can be seen that although the signal has a high response amplitude at the gas absorption peak, its waveform is superimposed with dense random spikes and significant baseline fluctuations. This indicates that the signal-to-noise ratio of the single frequency band is greatly affected by environmental noise and the fluctuation amplitude exceeds the system's preset stability threshold.

[0054] Figure 2 The green dotted line curve represents the synchronously acquired high-stability frequency band signal, which is less affected by noise but has relatively low sensitivity to gas concentration. Figure 2 The solid red curve in the middle represents the adaptive fused signal generated after processing by the frequency band fusion mechanism described in this embodiment of the invention.

[0055] By comparison, it can be intuitively seen that the red curve perfectly inherits the morphological advantages of the high-sensitivity signal at the characteristic peak, while effectively smoothing the background noise and clutter interference by utilizing the characteristics of the high-stability signal, thus significantly improving the signal-to-noise ratio and smoothness of the final output waveform.

[0056] This processing method, based on the optimal fusion of signal quality indicators, effectively solves the technical problem of unstable signal quality in a single detection frequency band under harsh operating conditions. It ensures that the optimal frequency band signal can accurately reflect the true absorption characteristics of the gas, thus providing a high-quality data foundation for coarse calibration and fine decoupling operations in subsequent steps.

[0057] S2: Perform the coarse calibration procedure.

[0058] After determining the optimal frequency band signal, the system enters the coarse calibration stage. The core objective of this stage is to quickly pinpoint the approximate range of the two gas concentrations without requiring a large amount of computational resources, providing reliable initial values ​​for subsequent fine decoupling, thereby solving the problem of global search easily getting trapped in local optima.

[0059] Specifically, the coarse calibration step first requires evaluating the current spectral aliasing state. The system calculates the spectral aliasing coefficient based on the characteristics of the optimal frequency band signal. The spectral aliasing coefficient is a dimensionless parameter that characterizes the degree of attenuation of the current signal quality relative to the reference signal quality.

[0060] For example, spectral aliasing coefficient The definition formula is as follows:

[0061] ;

[0062] in, Represents the spectral aliasing coefficient; The real-time signal-to-noise ratio represents the current optimal frequency band signal; The correlation coefficient represents the correlation between the current optimal frequency band signal and the standard reference waveform; The reference signal-to-noise ratio represents the system's initial calibration state, typically the maximum signal-to-noise ratio against a clean background; The value represents the preset reference attenuation factor, which is 0.95 in this embodiment. The physical meaning of this formula is that when severe aliasing occurs in the spectrum, the signal waveform will be distorted, leading to a decrease in the correlation coefficient. Simultaneously, the effective signal amplitude will decrease, resulting in a decrease in the signal-to-noise ratio, thus... The value increases. Therefore, The larger the value, the more severe the spectral aliasing interference.

[0063] like Figure 3 A three-dimensional surface visually demonstrates the dynamic mapping relationship between the spectral aliasing coefficient and the concentration values ​​and ratios of the two gases in a dual-gas mixture detection scenario.

[0064] Figure 3 The two horizontal coordinate axes represent the real-time concentration values ​​of gas A and gas B, respectively, while the vertical coordinate axes and the color gradient of the surface itself intuitively represent the magnitude of the spectral aliasing coefficient.

[0065] from Figure 3The chromatographic distribution and peak morphology shown in the figure indicate that when the concentrations of the two gases are in a specific ratio range, especially when their concentrations are close, the surface shows a significant bulge accompanied by a warm color representing high values. This corresponds to a strong aliasing interference state, indicating that at this time, due to the severe overlap of the characteristic absorption spectra of the two gases on the wavelength axis, the superimposed signal waveform is drastically distorted relative to the standard reference waveform, and the product of the signal-to-noise ratio and the correlation coefficient decreases significantly.

[0066] Conversely, at the edges or troughs of the surface, the color displays cool tones representing low values, indicating that a single gas is dominant or the total concentration is low, and the spectral characteristics maintain high purity. This mapping diagram forms the core decision-making basis for the coarse calibration module. The system locates the coordinates of the concentration data on the surface based on real-time acquisition, accurately quantifying the severity of interference under the current operating conditions. This information is then used as a feedforward signal to dynamically adjust and improve the operating parameters of the genetic algorithm. Specifically, in the high aliasing regions indicated by the surface peaks, the algorithm automatically increases the number of generations and reduces the mutation probability, thereby ensuring that the algorithm can still accurately lock the initial range of the two gas concentrations through fine searching even in complex and highly interfering environments.

[0067] Specifically, the system is based on the spectral aliasing coefficient. The search algorithm's operating parameters are dynamically adjusted based on the numerical range in which the solution is located. In this embodiment, an improved genetic algorithm is used. This algorithm searches for the optimal concentration combination within the solution space by simulating selection, crossover, and mutation operations in biological evolution.

[0068] Specifically, in this embodiment, a dual-gas standard reference spectrum library is pre-established in the memory. During each generation of the improved genetic algorithm, for each individual (representing a set of potential concentration values)... and The system constructs a theoretically predicted spectrum based on the linear superposition principle of the Lambert-Beer law. :

[0069] ;

[0070] in, and This is a preset standard reference spectrum.

[0071] Subsequently, the system calculates theoretical predictions of the spectrum. Compared with the optimal frequency band signal actually collected The root mean square error (RMSE) between the two values ​​is calculated, and the reciprocal of this error is used as the fitness function value. Based on this physical model, the system further uses the spectral aliasing coefficients... The search algorithm's operating parameters are dynamically adjusted based on the numerical range it falls within.

[0072] Optionally, the dynamic adjustment of the search algorithm's operating parameters follows the strategy of: when the spectral aliasing coefficient indicates an increase in aliasing, i.e. As the value increases, it means that the surface of the solution space becomes more complex, with more local extrema. To prevent the algorithm from prematurely converging to an incorrect local optimum, the system automatically increases the number of iterations, i.e., the number of generations, for example, from 20 to 40, to ensure the population has sufficient time to evolve. Simultaneously, the system reduces the mutation probability, for example, from 0.1 to 0.05, and combines this with an elite retention strategy to maintain the stability of superior individuals and prevent random mutations from destroying existing search results in highly disruptive environments.

[0073] like Figure 4 The figure shows a comparison of the search convergence trajectory between the improved genetic algorithm based on adaptive adjustment of the aliasing coefficient and the traditional fixed-parameter genetic algorithm. The horizontal axis represents the number of evolutionary iterations performed by the algorithm, and the vertical axis represents the fitness function value, i.e., the error level of the dual-gas concentration inversion.

[0074] Figure 4 The blue dashed line represents the convergence process of the traditional genetic algorithm. It can be seen that when dealing with complex solution spaces with severe spectral aliasing, the algorithm lacks a dynamic parameter adjustment mechanism. Although the error decreases rapidly in the early stage of iteration, it then falls into a premature convergence state and eventually stagnates at a high error level, i.e., it falls into a local optimum.

[0075] In comparison Figure 4 The solid red line represents the improved genetic algorithm described in this embodiment of the invention. During the iteration process, the algorithm senses the degree of environmental interference in real time based on the calculated spectral aliasing coefficient. By automatically increasing the number of iterations and reducing the mutation probability when a high aliasing state is detected, combined with the elite retention strategy, the curve does not stagnate after the initial rapid decline, but continues to effectively search for and break through local extreme points in subsequent iterations, and finally converges to a significantly lower error level.

[0076] This comparison intuitively demonstrates that the dynamic strategy adopted in this invention can effectively overcome the multi-peak trap in the solution space caused by spectral aliasing, ensuring that the coarse calibration step can output a high-precision initial concentration range, laying a reliable foundation for subsequent fine decoupling calculations.

[0077] It is also important to note that the improved genetic algorithm employs a binary segmented encoding method. Specifically, the chromosome is divided into two segments: the first segment corresponds to the concentration value of the first gas, and the second segment corresponds to the concentration value of the second gas. During the coarse calibration step, the system also monitors the offset of wavelength-selective device feature points in the device aging characteristics in real time. If this offset exceeds a preset time drift threshold, such as 0.2 milliseconds, it indicates that wavelength drift has caused a significant change in the spectral position. In this case, the original encoding accuracy may not be sufficient to cover the drifted wavelength range. Therefore, the system will perform adaptive adjustment of the encoding length, increasing the number of bits in the binary segmented encoding, for example, adding 2 bits to each of the two encoding segments, thereby expanding the resolution and range of the search space.

[0078] Specifically, the initial search center in the coarse calibration step is not randomly generated, but determined based on the final precise concentration output from the previous time step. The system constructs a local search neighborhood centered on the precise concentration value obtained in the previous time step. If the spectral aliasing coefficient... If the value is smaller, the search range is reduced to one-third of the original range to improve the convergence speed; if... If the value is large, the search range is maintained or expanded to ensure robustness. Ultimately, the population-optimal solution output by the improved genetic algorithm becomes the initial range of the two gas concentrations.

[0079] S3: Perform fine decoupling steps.

[0080] The coarse calibration determines the approximate location of the concentration, while the fine decoupling step aims to achieve precise localization in the vicinity of this approximate location. This step utilizes manifold learning and nonlinear regression techniques to solve the nonlinear mapping challenge of TDLAS in high-dimensional data spaces.

[0081] Specifically, the system first needs to construct input data. The system vectorizes and concatenates the initial range (output from coarse calibration), the optimal frequency band signal (containing complete waveform data points), the environmental parameters (temperature, air pressure, etc.), and the equipment aging characteristics (deviation, offset, attenuation ratio) to construct a high-dimensional input data matrix. This matrix not only contains the original waveform data but also the residual vector calculated based on the coarse calibration results and the standard reference spectrum. The residual vector contains nonlinear characteristics that cannot be explained by the standard linear model, such as nonlinear distortion caused by gas crosstalk.

[0082] This matrix integrates spectral information, physical environment information, and equipment status information, and its dimensions are usually as high as thousands of dimensions, which is mainly determined by the number of spectral data points.

[0083] For example, to address the curse of dimensionality and computational time issues associated with high-dimensional data, the system utilizes the Local Linear Embedding (LLE) algorithm to perform manifold learning on the high-dimensional input data matrix. The LLE algorithm can map high-dimensional data to a low-dimensional space while preserving the local neighborhood structure of the data, thereby extracting essential features from the spectral data, such as the intensity, full width at half maximum (FWHM), and peak position difference of absorption peaks, while simultaneously eliminating noise redundancy.

[0084] like Figure 5 The figure shows a comparison of the spatial distribution of the constructed high-dimensional input data matrix before and after dimensionality reduction using the local linear embedding algorithm through manifold learning.

[0085] Figure 5 The left-hand subplot shows the distribution of the original spectral feature data in the 3D projection space before dimensionality reduction. Each scatter point represents a spectral sample collected at one time step, and the color intensity of the scatter point corresponds to the magnitude of the gas concentration. It can be seen that due to severe spectral aliasing interference and nonlinear drift caused by equipment aging, the original data exhibits a highly coiled and entangled manifold structure. This indicates that in the original input space with up to thousands of dimensions, it is difficult to extract the essential features reflecting gas concentration using a linear regression model directly.

[0086] Figure 5 The right-hand subplot shows the low-dimensional feature distribution obtained by dynamically selecting the target dimension based on the calculated spectral aliasing coefficients and processing it using a local linear embedding algorithm. It can be seen that after manifold learning, the originally convoluted and entangled data manifold is successfully unfolded and flattened into a low-dimensional space. Sample points with different concentration gradients exhibit a clear and continuous linear arrangement, effectively preserving the local neighborhood structure of the data and decoupling from nonlinear features through dimensionality reduction.

[0087] This transformation process intuitively demonstrates that the method described in this invention can effectively extract essential features such as absorption peak intensity and full width at half maximum (FWHM) from a complex high-dimensional matrix containing environmental parameters and aging characteristics, providing a low-redundancy and highly separable data foundation for establishing a high-precision concentration mapping relationship through an incremental support vector regression model.

[0088] It is important to note that the target dimension of the LLE algorithm, i.e., the dimension after dimensionality reduction, is crucial. Too low a dimension will result in the loss of useful information, while too high a dimension will fail to achieve the desired dimensionality reduction effect. In this embodiment, the target dimension of the locally linear embedding algorithm is not a fixed value, but rather an adaptively selected dimension based on the spectral aliasing coefficient and the deviation between the driving current and wavelength mapping. This strategy fully considers the impact of spectral aliasing degree and device aging on data complexity.

[0089] Specifically, the adaptive selection rule for the target dimension is as follows:

[0090] The system pre-sets a first aliasing threshold (e.g., 0.3) and a second aliasing threshold (e.g., 0.7), wherein the first aliasing threshold is less than the second aliasing threshold. Simultaneously, for the deviation between the driving current and wavelength mapping, a first deviation range (e.g., less than or equal to 3%), a second deviation range (e.g., greater than 3% and less than or equal to 5%), and a third deviation range (e.g., greater than 5%) are also set.

[0091] When the spectral aliasing coefficient is less than the first aliasing threshold, and the deviation between the driving current and the wavelength mapping is within the first deviation range, it indicates that the spectral interference is small, the device is in good condition, and the nonlinearity of the data is low. The system sets the target dimension to a first dimension value, such as 30 dimensions, to compress the data to the maximum extent and improve the calculation speed.

[0092] When the spectral aliasing coefficient is between the first aliasing threshold and the second aliasing threshold, or when the deviation between the driving current and the wavelength mapping is within the second deviation range, it indicates that the interference is moderate or the device is slightly aging. The system sets the target dimension to a second dimension value, such as 50 dimensions, to retain more feature information for distinguishing aliased signals.

[0093] When the spectral aliasing coefficient is greater than the second aliasing threshold, or the deviation between the driving current and the wavelength mapping is within the third deviation range, it indicates high aliasing or severe aging of the equipment, resulting in an extremely complex spectral waveform. The system sets the target dimension to a third dimension value, such as 70 dimensions.

[0094] The third dimension value is greater than the second dimension value, and the second dimension value is greater than the first dimension value. This step-by-step dimension adjustment strategy achieves an optimal balance between computing resources and decoupling accuracy.

[0095] Specifically, after dimensionality reduction, the system establishes a nonlinear mapping relationship between the dimensionality-reduced low-dimensional spatial data and the concentrations of the two gases using an incremental support vector regression (SVR) model. The SVR model employs a radial basis function (RBF), which can effectively fit the highly nonlinear spectrum-concentration relationship. "Incremental" means that the model supports online updates, utilizing a sliding window mechanism to cache the latest time-series data and continuously fine-tune the model's support vectors and weights, thereby adapting to the slow drift of the environment and equipment states to obtain precisely decoupled concentrations.

[0096] S4: Perform compensation and correction steps.

[0097] The concentration values ​​output by the fine decoupling step are primarily based on spectral characteristics, but the absorption intensity of the gas is also significantly affected by temperature and pressure. Therefore, a compensation correction step must be performed to obtain the final accurate concentration.

[0098] Specifically, the system first inputs the collected environmental parameters, mainly temperature, air pressure, and humidity, into a pre-trained Gradient Boosting Regression (GBR) model. This GBR model was pre-trained in a laboratory environment by testing standard gas concentrations under different temperature and pressure combinations. The output of the GBR model is an environmental impact compensation coefficient, which characterizes the rate of change of gas absorption line intensity relative to standard conditions under the current environmental conditions, particularly considering the broadening interference of polar gases (such as water vapor) on the spectral lines of non-polar analyte gases.

[0099] For example, the system calculates the final accurate concentration using temperature and pressure correction logic. This correction logic uses the finely decoupled concentration as a reference and performs the calculation using a corrected form of the ideal gas equation of state. Assume the concentration obtained from fine decoupling is... The environmental impact compensation coefficient is The environmental impact weight is (Automatically learned by the model), the current air pressure is The current temperature is Standard atmospheric pressure is (101.325 kPa), standard temperature is (298K). The final precise concentration... The calculation formula is as follows:

[0100] ;

[0101] ;

[0102] It is important to note here that the correction terms in the formula include not only the physical corrections to the ideal gas law, but also... The inverse relationship of the terms is also shown through intermediate variables. Nonlinear environmental disturbance compensation calculated by the GBR model was introduced, thus achieving dual correction.

[0103] S5: Equipment aging adaptive calibration process.

[0104] To address the issue of model inaccuracy caused by long-term equipment operation, this embodiment incorporates an adaptive calibration process for equipment aging in parallel with the aforementioned process. This process enables hierarchical management and targeted calibration of the equipment's health status.

[0105] Specifically, the system first needs to assess the aging condition of the equipment. The assessment criteria involve three core parameters: drive current and wavelength mapping deviation (denoted as...). ), wavelength selection device feature point offset (denoted as ) and signal-to-noise ratio attenuation ratio (denoted as The system has preset deviation thresholds (e.g., 3%), drift thresholds (e.g., 0.2ms), and quality thresholds (e.g., 0.8).

[0106] when Less than the deviation threshold, and Less than the drift threshold, and When the quality exceeds the stated quality threshold, the system determines that the device is in a state of mild aging. At this time, the device performance is basically normal, with only minor fluctuations. When any of the above parameters falls into a preset intermediate degradation range, for example... Between 3% and 5%, or Between 0.2ms and 0.5ms, the system determines it to be in a moderate aging state. At this point, measurement accuracy begins to be affected. When any of the above parameters exceeds the limit value indicated by the intermediate degradation range, for example... When the value exceeds 5%, the system determines it to be in a severely aged state, at which point the device may no longer be able to guarantee basic measurement functions.

[0107] Specifically, the system implements differentiated calibration strategies for different aging states:

[0108] If the system is in the aforementioned mild aging state, it primarily adapts through software parameter fine-tuning. Specifically, this involves adjusting the encoding range of the improved genetic algorithm (fine-tuning the search boundary) and the kernel parameters of the support vector regression model, such as adjusting the width coefficient of the RBF kernel. This is to compensate for minor feature drift.

[0109] If the system is in the moderate aging state, it indicates that the hardware characteristics have undergone visible changes, and simple model fine-tuning is no longer sufficient. At this point, the system triggers a rapid recalibration process. This process does not require full calibration; instead, it controls the laser to scan only specific characteristic wavelength points, such as the absorption peak center and the half-maximum. The data from these points is used to quickly update the driving current-wavelength mapping table, which takes very little time, thereby correcting the wavelength axis without interrupting the measurement.

[0110] If the device is in a severely aged state, it indicates that single-band measurements are no longer reliable. The system will forcibly switch to a dual-band fusion decoupling mode, i.e., the fusion mode mentioned in the aforementioned signal acquisition steps, forcibly fusing the 3f and 4f signals. This utilizes the information redundancy of multiple frequency bands to resist the signal-to-noise ratio degradation caused by device aging. Simultaneously, a maintenance reminder signal will be generated via the host computer or indicator lights to prompt manual intervention to replace the device.

[0111] like Figure 6The figure shows the effect of device aging drift on the spectral signal and the comparison of the effect before and after applying the adaptive calibration strategy of the present invention. The horizontal axis represents the scanning time of the laser or the corresponding driving current phase, and the vertical axis represents the normalized amplitude of the second harmonic signal after phase-locked amplification.

[0112] Figure 6 The black dashed curve in the middle represents the standard reference spectrum collected by the system during the initialization and calibration phase. Its characteristic absorption peak is precisely located at the center of the scan cycle, representing the ideal waveform when the device is in a healthy state.

[0113] Figure 6 The solid blue line curve represents the degraded spectrum caused by laser aging or wavelength selection device characteristic drift after long-term operation of the equipment. It can be seen that the characteristic peak of the curve has shifted significantly laterally relative to the reference position, accompanied by attenuation of signal amplitude and deterioration of signal-to-noise ratio. If this shift is not corrected, the concentration inversion algorithm will be unable to correctly capture the feature points, resulting in huge errors.

[0114] Figure 6 The solid red curve represents the repair spectrum output by the adaptive calibration process triggered after the system detects that the aging characteristic parameters have exceeded the limit. By dynamically correcting the scanning range of the laser driving current and updating the mapping relationship based on the aging characteristic vector, the system successfully pulls the drifting characteristic peak back to the center position of the scanning cycle and achieves high alignment with the reference spectrum. This proves that the device aging adaptive calibration mechanism described in this invention can effectively offset the measurement deviation caused by hardware performance degradation and ensure that the system maintains high-precision detection capability throughout its entire life cycle.

[0115] In summary, this embodiment constructs a TDLAS dual-gas concentration detection system that can resist spectral aliasing interference, adapt to dynamic environmental changes, and has self-diagnostic capabilities through five tightly coupled processes: signal acquisition, coarse calibration, fine decoupling, compensation correction, and graded aging calibration.

[0116] It should be further noted that, in order to ensure the real-time performance of the above method, the entire processing flow adopts a pipelined parallel architecture.

[0117] For example, the first-stage pipeline is responsible for multi-band signal acquisition and environmental parameter reading; the second-stage pipeline is responsible for optimal frequency band selection and LLE dynamic dimensionality reduction; and the third-stage pipeline is responsible for coarse calibration search and SVR incremental regression. This design enables the system to complete a full concentration output within milliseconds, meeting the stringent timeliness requirements for gas leak monitoring in industrial settings.

[0118] In practical applications, the method of this embodiment can run in embedded control systems, such as high-performance processors based on FPGA (Field Programmable Gate Array) or ARM architecture. The timing sample pool is stored in high-speed SRAM, while the pre-trained dynamic environment compensation model (GBR) and incremental SVR model reside in DDR memory.

[0119] It is also important to note that regarding spectral aliasing interference, in this embodiment, it specifically refers to the situation where two gases, such as methane and ethane, or carbon monoxide and carbon dioxide, are present in a mixture. Because the center wavelengths of the absorption lines of these two gases are extremely close (e.g., less than 1 nanometer), the spectral lines overlap under normal pressure, causing the characteristic peaks of a single gas to no longer be independent. Traditional methods often handle this with simple multiple linear regression (MLR), but this requires linear superposition of spectral lines without wavelength drift. This embodiment, by introducing a spectral aliasing coefficient K, essentially quantifies the degree of nonlinear distortion caused by this overlap and adjusts the search strategy of the genetic algorithm and the dimensionality reduction of LLE accordingly. Essentially, it is a cognitive computational process of perceiving interference and adjusting the strategy, which is one of the innovative points that distinguishes this method from existing technologies.

[0120] Specifically, regarding the incremental support vector regression model, its core algorithm logic is as follows: Let... The training sample set at time t is ,in The reduced-dimensional feature vectors are the result of dimensionality reduction. This corresponds to the true concentration. When... New samples at time Upon arrival, the model does not retrain the entire dataset. Instead, it uses KKT (Karush-Kuhn-Tucker) conditions to determine whether a new sample violates the current retention set conditions. If it does, the sample is added to the support vector set, and the oldest non-support vector sample or the support vector with the smallest contribution is removed. Simultaneously, the Lagrange multipliers are updated. and bias This incremental update mechanism allows the model to learn as it is tested. As the detection time progresses, the model becomes increasingly adaptable to the current device state and environmental characteristics, thereby achieving continuous self-optimization of accuracy.

[0121] Furthermore, regarding binary segmented encoding, the specific encoding structure in the embodiment is as follows: assuming the total chromosome length is... Position, of which the first The position represents the concentration of gas A, and then... The position represents the concentration of gas B. Let the measurement range for the concentration of gas A be... Then the binary value encoded in the front end Corresponding concentration value The calculation formula is:

[0122] ;

[0123] When the system detects wavelength selection device offset When the threshold is exceeded, it means that the wavelength axis has shifted, and the original concentration inversion correspondence may become invalid. At this time, the system will... Increasing the value, for example from 20 bits to 24 bits, essentially improves the resolution of concentration analysis, enabling the genetic algorithm to search for the true concentration corresponding to the small changes in spectral intensity caused by wavelength drift on a finer grid, thereby offsetting the error caused by hardware drift.

[0124] Finally, the method provided in this embodiment not only outputs a highly accurate final concentration but also exhibits extremely high robustness. Even in the presence of strong spectral aliasing interference (… and moderate aging of equipment ( Under harsh working conditions, experimental verification shows that this method can still control the decoupling accuracy error to an extremely low level, and the response delay is much lower than that of traditional offline calibration methods, which fully demonstrates the advanced nature and practicality of this technical solution.

[0125] Example 2:

[0126] Building upon Example 1, this embodiment focuses on a deeper refinement and expansion of the implementation logic of the incremental support vector regression model under complex and dynamic real-world conditions. Specifically, for common industrial scenarios involving sudden gas concentration changes, such as valve opening or pipeline leaks, this embodiment details a dual-core collaborative evolution strategy based on residual accumulation and detection. This strategy aims to resolve the inherent contradiction between steady-state accuracy and dynamic response speed in a single model, ensuring that the system can respond to sudden concentration changes within milliseconds while maintaining low-noise monitoring.

[0127] Specifically, the nonlinear mapping core algorithm used in this embodiment does not follow the traditional single SVR kernel architecture, but innovatively constructs a dual-core architecture with a main regression kernel and a sensitive tracking kernel running in parallel. In software logic, this architecture is equivalent to maintaining two regressors with different characteristics and tasks in memory at the same time. The two share the input data stream, but have significant differences in parameter configuration and activation state.

[0128] For example, the principal regression kernel is positioned as the steady-state benchmark of the system, primarily responsible for providing high-precision concentration output under normal monitoring conditions. To achieve this, the principal regression kernel is configured with a first penalty factor, which is set to a large value, such as 100. In the mathematical principles of support vector regression, the penalty factor represents the model's tolerance for training errors. A large first penalty factor means that the principal regression kernel imposes a severe penalty on any sample point that deviates from the regression hyperplane, forcing the model to fit the majority of samples in the historical data as accurately as possible, thereby constructing a smooth and stable regression curve. This configuration gives the principal regression kernel extremely strong noise resistance, effectively filtering out high-frequency random noise in the optical path and ensuring the smoothness of the output curve.

[0129] Optionally, a sensitive state tracking kernel, configured in parallel with the main regression kernel, is positioned as the system's mutation catcher, primarily responsible for quickly capturing data trends when concentrations change drastically. The sensitive state tracking kernel is configured with a second penalty factor, significantly smaller than the first penalty factor, for example, set to 1. A smaller second penalty factor means the kernel allows for a larger fitting error, thus giving the model greater flexibility and resilience, preventing it from being constrained by the inertia of historical data. Furthermore, the sensitive state tracking kernel is configured with a special Gaussian kernel width scaling factor. In the radial basis function (RBF), the kernel width determines the influence range of a single support vector on surrounding samples. The sensitive state tracking kernel, employing a smaller kernel width or a specific scaling factor, possesses a more acute local perception capability in the feature space, enabling a strong response to newly input data points that deviate from the old distribution, thereby improving transient sensitivity.

[0130] Specifically, to intelligently switch and schedule between these two kernels, this embodiment introduces a real-time detection mechanism based on the cumulative sum of residuals (CUSUM). During system operation, the predicted residuals of the main regression kernel are calculated in real time. This is defined as the deviation between the predicted output value of the master regression kernel at the current moment and the characteristics of the actual input signal (or a temporary reference value transformed through coarse calibration). Subsequently, the system updates the cumulative sum of the predicted residuals based on this residual.

[0131] For example, cumulative sum values The iterative calculation logic is as follows: Let The prediction residual at time is The cumulative sum value at the previous moment is The system's set drift tolerance value is Then the cumulative sum value at the current moment The calculation formula is:

[0132] ;

[0133] The physical meaning of this formula is that as long as the current prediction residual... Less than the allowable drift value , accumulation and The residuals will tend towards zero or remain at zero; however, if the residuals consistently exceed the allowable range, it indicates that a non-random systematic bias has occurred in the data stream. It will accumulate rapidly in a positive direction. This mechanism can effectively distinguish instantaneous random noise spikes (without causing...) (continuous increase) and a real gas concentration step (which would lead to) (Continued rapid increase).

[0134] Specifically, based on the calculated cumulative sum, the system will operate in different modes:

[0135] When the accumulated sum value is lower than a preset drift threshold At this point, the system determines that it is currently in steady-state mode. In this mode, changes in the external environment and gas concentration are gradual, and the predictive performance of the principal regression kernel is good. At this time, the system keeps the parameter update mechanism of the principal regression kernel enabled, only using newly added data to fine-tune the support vector weights of the principal regression kernel to adapt to slow baseline drift. Meanwhile, to save computational resources, the sensitive tracking kernel is in a dormant state or only performs background data caching without performing regression calculations.

[0136] It should also be noted that when the cumulative sum value exceeds the drift threshold... When this happens, the system immediately determines that it has entered a concentration mutation mode. This situation usually corresponds to the beginning of a gas leak. At this time, due to its large inertia, the main regression kernel's predicted value often remains at a low concentration level, leading to a sharp increase in the prediction residual.

[0137] Once the mutation mode is confirmed, the system performs the following key operations: First, it freezes the parameter updates of the master regression kernel. This is to protect the kernel's memory from being contaminated by unstable transitional data during the mutation period, preventing the destruction of its support vector set. Second, it immediately activates the sensitive tracking kernel, utilizing its high sensitivity to quickly fit the current low-dimensional spatial data. During the mutation mode, the system directly outputs the output of the sensitive tracking kernel as a temporary decoupling result of the current dual-gas concentration to the host computer or alarm unit, thereby eliminating the potential response lag of several seconds that might occur with the master regression kernel.

[0138] Optionally, after the gas concentration completes a step change and stabilizes at a new level, such as when the concentration stabilizes at a high value after the valve is fully opened, although the prediction residual of the main regression kernel is large, the fitting of the sensitive state tracking kernel gradually tends to stabilize. At this time, the cumulative sum value... The sum will fall back due to the stabilization of the residuals or the reset logic. This will occur when the accumulated sum value falls back to a preset reset threshold. When the mutation process ends, the system needs to switch back to steady-state mode. However, directly switching back to the principal regression kernel will cause the output value to jump back to the old value before the mutation, so a support vector transfer operation must be performed.

[0139] Specifically, the support vector transfer operation is a key step in achieving co-evolution in this embodiment. When performing this operation, the system first extracts the boundary support vectors identified during the concentration mutation mode from the memory of the sensitive tracking kernel. These boundary support vectors outline the new data distribution profile in the feature space, representing the latest concentration state knowledge. Subsequently, the system forcibly injects these boundary support vectors into the support vector set of the master regression kernel. This process is equivalent to giving the master regression kernel a knowledge indoctrination, enabling it to instantly learn the new concentration state without undergoing a lengthy, gradual learning process.

[0140] It's also important to note that to prevent the support vector set of the main regression kernel from expanding indefinitely and causing a decrease in computational efficiency, the system removes redundant old support vectors from the main regression kernel based on linear correlation while injecting new support vectors. Specifically, the system calculates the linear correlation matrix within the support vector set, or, based on the sample timestamps, prioritizes removing old vectors that have a very high linear correlation with the newly injected vectors (i.e., those containing duplicate information), or removes the oldest non-critical vectors. Through this metabolic mechanism, the main regression kernel maintains the compactness of the model size while completing state updates.

[0141] For example, the entire process of dual-core co-evolution can be summarized as follows: when there are no accidents, the main core takes the lead and ensures accuracy; when an accident occurs, the sensitive core takes over and ensures speed; and when there are no accidents again, the sensitive core transfers the new support vectors to the main core, and the main core continues to maintain high-precision monitoring on the new high ground.

[0142] As can be seen from the detailed description of the above embodiments, the present invention cleverly resolves the contradiction between sensitivity and stability by introducing a dual-core co-evolution strategy based on residual accumulation and detection into the incremental SVR model. In actual testing, when faced with a step change in methane concentration from 0 to 1000 ppm, the response time (T90) of the traditional single-core SVR model is usually more than 5 seconds. However, after adopting the dual-core strategy described in this embodiment, the system response time is shortened to less than 200 milliseconds, and the measurement standard deviation under steady state remains at a low level. This technical solution has significant safety value for industrial field applications that require rapid gas source cutoff or triggering of safety interlocks.

[0143] Example 3:

[0144] This embodiment, building upon Embodiments 1 and 2, further focuses on in-depth optimization of the compensation and correction steps for extreme transient conditions. Specifically, for high-speed fluid environments or pressure vessel monitoring scenarios common in industrial settings, physical sensors, such as piezoresistive pressure sensors or thermocouple temperature sensors, are limited by thermal inertia and mechanical response characteristics, with response times often on the order of seconds or hundreds of milliseconds. In contrast, TDLAS optical detection, based on the interaction of photons with gas molecules, can achieve response times on the order of nanoseconds. This temporal mismatch in the physical detection mechanism leads to significant computational errors when the system compensates for real-time spectral signals using lagging environmental parameters during drastic fluctuations in environmental parameters, such as a sudden depressurization of a pipeline.

[0145] To resolve this deep-seated technical contradiction, this embodiment details a sensor hysteresis correction mechanism based on spectral line shape inversion. This mechanism utilizes the physical information carried by the spectral signal itself as a zero-delay virtual sensor, taking over and correcting the readings of the physical sensor during transient processes.

[0146] Specifically, the sensor hysteresis correction mechanism in this embodiment is embedded in the preprocessing stage of the compensation correction step. When the system enters this stage, although the fine decoupling concentration has been obtained through the method of Embodiment 1 and the environmental parameters at the current moment have been collected through physical sensors, the system does not directly call these data for final calculation, but first starts the hysteresis verification process.

[0147] For example, the first step in this process is to extract the spectral line shape factor from the optimal frequency band signal. In the principle of TDLAS technology, the shape of gas absorption spectral lines is mainly affected by both Doppler broadening (temperature-dependent) and Lorentz broadening (pressure-dependent). Under normal or high-pressure industrial environments, Lorentz broadening, dominated by molecular collisions, is the main factor determining the spectral linewidth. Therefore, there is a strict physical correspondence between the waveform characteristics of harmonic signals and ambient air pressure.

[0148] Specifically, the spectral line shape factor is defined as a dimensionless parameter characterizing the Lorentz broadening of gas absorption lines. To reduce computational complexity and improve robustness against amplitude fluctuations, this embodiment defines the spectral line shape factor as the ratio of the dominant peak value of the harmonic signal to the zero-crossing width. Let the dominant peak value of the current optimal frequency band signal, such as a second harmonic (2f) signal or a fourth harmonic (4f) signal, be... The spectral scan time width (or corresponding wavelength width) between the nearest zero-crossing points on both sides of the main peak is The measured spectral line type factor is... The calculation formula is as follows:

[0149] ;

[0150] It's also important to note that the formula defined here uses the form of width divided by amplitude, or the reciprocal, to construct a quantity positively correlated with pressure. When air pressure increases, increased collisions lead to spectral broadening, i.e. As the peak height increases, it may decrease or remain the same due to peak normalization or absorption conservation, thus making... Significant changes occur. This feature quantity extracted from the original waveform is completely synchronized with the spectral acquisition process, and therefore has zero delay relative to the actual pressure change on the time axis.

[0151] Optionally, after extracting the measured features, the system needs to determine the reliability of the current physical sensor data. This requires performing a second step: calculating the theoretical linearity factor. The system reads the air pressure values ​​reported by the current physical barometric sensor. and the temperature value reported by the temperature sensor The system internally stores a table of gas collision broadening coefficients fitted from the HITRAN database or experimental calibration data. The system then adjusts the parameters based on the current environmental parameters. and By combining the gas collision broadening coefficients in the spectral database and using the Voigt line shape simulation algorithm or simplified empirical formulas, the theoretically observable line shape characteristics, i.e., the theoretical line shape factor, can be calculated under the current sensor readings. Its functional relationship can be expressed as:

[0152] ;

[0153] in The air widening factor is a known physical constant.

[0154] Specifically, after obtaining the measured and theoretical values, the system performs the third step: deviation calculation and decision. The system calculates the spectral line type factor. With the theoretical linearity factor The morphological deviation. The morphological deviation Defined as the absolute value of the relative error between the two:

[0155] ;

[0156] The system has a preset hysteresis threshold. For example, it can be set to 0.15, which is a deviation of 15%.

[0157] When the morphological deviation Less than or equal to the hysteresis determination threshold At this point, the system determines that the physical sensors are working normally and without significant hysteresis, meaning the current environment is in a steady state or a low-speed changing state. The system then continues to use the conventional procedure described in Example 1. Subsequent compensation calculations are performed when the morphological deviation... Exceeding the preset hysteresis threshold At this time, the system determines that there is a lag in the acquisition of environmental parameters. This usually occurs when a pipeline valve is suddenly opened, causing a momentary drop in air pressure. At this point, the spectral linewidth has already narrowed, i.e. It gets smaller, but the physical barometer reading has not yet decreased. Still high value, leading to It remains a large value, thus producing a huge deviation.

[0158] It is also important to note that once a lag is detected, the system immediately initiates a transient reconstruction process. The core idea of ​​this process is to discard unreliable physical sensors and utilize photons as virtual sensors. The system uses a preset linewidth-pressure mapping function to perform the inversion operation. This linewidth-pressure mapping function is a mathematical model established during the system's factory calibration phase, using polynomial fitting or a neural network to record spectral characteristics and standard pressure gauge data at high frequency in a high-dynamic pressure test chamber. This function describes the monotonic correspondence between the spectral linetype factor and actual air pressure within a specific temperature range. Let this mapping function be... The instantaneous equivalent air pressure obtained by inversion The calculation is as follows:

[0159] ;

[0160] It should be noted here that although physical temperature is still used in the formula... However, since the rate of pressure change is usually much faster than the rate of temperature conduction, except for adiabatic expansion, even in adiabatic processes, linewidth is much more sensitive to pressure than to temperature.

[0161] Therefore, in short-term stress changes, the following approach should be adopted. The resulting error is far less than that of using the lag. This introduces errors. Alternatively, in a more advanced implementation, a joint inversion model of the linear factor for temperature and pressure can be established, simultaneously reconstructing pressure and temperature.

[0162] Specifically, in calculating the instantaneous equivalent air pressure Then, the system performs the most crucial step: parameter replacement and recalculation. The system utilizes the instantaneous equivalent air pressure. Temporarily replace the physical air pressure value in the environmental parameter data packet. Subsequently, the system will use this reconstructed set of environmental parameters (including...) and The stress value is then re-input into the pre-trained Gradient Boosting Regression (GBR) model. Based on this stress value, which more closely approximates the actual physical state, the GBR model outputs an accurate environmental impact compensation coefficient. Furthermore, this instantaneous equivalent pressure is also used in the calculations when executing the final temperature and pressure correction formula:

[0163] ;

[0164] In this way, the system can sense the pressure change through the spectral signal that travels at the speed of light before the physical sensor readings are transmitted to the processor through the signal conditioning circuit, and complete the real-time pressure correction at the algorithm level.

[0165] For example, in an experiment simulating a natural gas pipeline leak—that is, a pressure drop from 0.5 MPa to 0.1 MPa within 100 milliseconds: without the method of this embodiment, due to the approximately 500 millisecond response delay of the physical pressure sensor, the system would still use a high pressure of 0.5 MPa to compensate for the weak spectral signal actually generated at 0.1 MPa during the initial 500 milliseconds of the leak. According to the ideal gas law, this would cause the calculated concentration value to be incorrectly amplified by a factor of 5, resulting in a huge false concentration spike on the monitoring curve, which could easily trigger a false alarm. However, with the sensor hysteresis correction mechanism described in this embodiment, the system detects the leak as early as the 10th millisecond after it occurs (this is mainly limited by the spectral scanning frequency). The mutations, and calculated If the threshold is exceeded, the refactoring process is triggered. (Inversely derived) The concentration curve closely followed the actual pressure curve. The final output concentration curve was stable without any spurious peaks, demonstrating the effectiveness of this mechanism under extreme transient conditions.

[0166] like Figure 7 The figure shows a comparative analysis of the pressure reconstruction performance and the resulting concentration compensation error of a traditional physical sensor and a virtual sensor based on spectral line inversion proposed in this invention under transient fluid dynamic conditions simulating a rapid pressure relief fault in a high-pressure pipeline network.

[0167] Figure 7 The horizontal axis represents the system's high-frequency monitoring time axis, the vertical axis on the left represents the absolute value of the ambient air pressure, and the vertical axis on the right represents the relative error in concentration calculation caused by pressure compensation deviation.

[0168] Figure 7The solid black line depicts the actual fluid pressure change trajectory inside the pipe, showing the nonlinear pressure drop triggered at 200 milliseconds. The dashed blue curve clearly demonstrates that traditional piezoresistive sensors, limited by inherent thermal inertia and mechanical damping characteristics, exhibit significant phase lag and amplitude attenuation in their response trajectory, failing to keep up with the actual pressure drop rate.

[0169] In stark contrast, the red dotted line curve represents the virtual sensor pressure reconstruction curve constructed by the present invention using the spectral line factor in real time. This curve closely follows the real pressure change trajectory thanks to the nanosecond-level response characteristics of the interaction between photons and gas molecules, compressing the system response delay to the limit of the data processing cycle.

[0170] Figure 7 The semi-transparent green filling area visually quantifies the consequences of physical sensor hysteresis, namely, the generation of false positive concentration deviations with amplitudes as high as several times during the pressure transient window, which can easily lead to misjudgments and malfunctions in industrial safety systems. The thin purple line indicates that after applying the transient reconstruction mechanism of this invention for dynamic correction, the residual error of concentration compensation is effectively converged to an extremely narrow noise floor range, fully verifying the technical advancement and measurement reliability of this method in solving the sensor time-domain mismatch problem.

[0171] Furthermore, it should be noted that the construction of the linewidth-pressure mapping function should have adaptive update capabilities. During the system's steady-state operation, i.e. At that time, the system will automatically collect the current physical air pressure. and measured linear factor As new training samples The coefficients of the function are fine-tuned. This online learning mechanism ensures that the calibration curve of the virtual sensor can be automatically calibrated to follow the aging of the wavelength selection device or small changes in the focal length of the optical path, preventing inaccurate inversions caused by the drift of the optical system itself.

[0172] In summary, this embodiment establishes a second-dimensional environmental perception channel independent of physical sensors by deeply mining the frequency domain morphological characteristics of spectral signals. This not only resolves the fundamental conflict between optical and physical detection in terms of time response, but also essentially installs a light-speed barometer on the TDLAS system, greatly expanding the application potential of this method in extremely high-dynamic fields such as supersonic flow field diagnosis, detonation process monitoring, and high-pressure pipeline safety monitoring.

[0173] Example 4:

[0174] like Figure 8As shown, this fourth embodiment, based on the above method embodiments, further provides a hardware system architecture capable of executing the above methods. This embodiment details a TDLAS dual-gas concentration synchronous calibration and decoupling system in the presence of spectral aliasing interference. This system adopts a modular design, with each functional module interacting via a high-speed data bus to collaboratively complete the entire process from signal sensing to final concentration output. This system aims to solve the decoupling distortion problem caused by severe spectral overlap, as mentioned in the background art, and the transient error problem caused by the lag in physical sensor response. Through deep integration of hardware and algorithms, high-precision monitoring under complex industrial conditions is achieved.

[0175] Specifically, the system described in this embodiment mainly includes a signal acquisition module, a coarse calibration module, a fine decoupling module, and a compensation and correction module. These modules do not exist in isolation, but correspond to the core steps in the above method embodiments, together forming a closed-loop intelligent detection system.

[0176] First, the system includes a signal acquisition module. This module is the front-end sensing unit of the entire system, and its physical layer is connected to a tunable laser, a dual-bandpass wavelength selection device, and a mid-infrared detector. The main function of the signal acquisition module is to acquire adaptively optimized multi-band harmonic signals while collecting environmental parameters. Environmental parameter acquisition is achieved by connecting to external temperature and pressure sensors via an interface, while harmonic signal acquisition is achieved by controlling the laser's drive current and extracting it using lock-in amplification technology. To address wavelength drift caused by equipment aging, the module integrates adaptive optimization logic, which dynamically adjusts the scanning range based on the feedback of the drive current and wavelength mapping deviation. More importantly, the module possesses intelligent decision-making capabilities, capable of calculating the signal-to-noise ratio and correlation coefficient of each frequency band in real time, and selecting the optimal frequency band signal based on these signal quality indicators. For example, when a drastic fluctuation in the signal-to-noise ratio of a single frequency band is detected, the module automatically performs a frequency band fusion operation, weighting and combining the features of multiple frequency bands to provide a high-quality data source for subsequent processing.

[0177] Secondly, the system includes a coarse calibration module. This module, as the system's primary processing unit, is responsible for rapidly narrowing down the solution space. It receives the optimal frequency band signal from the signal acquisition module and calculates the spectral aliasing coefficient based on the characteristics of the optimal frequency band signal. This coefficient is a core indicator for the system to determine the complexity of the current operating condition, representing the degree of attenuation of the current signal quality relative to the reference signal quality. Based on the numerical range of the spectral aliasing coefficient, the coarse calibration module dynamically adjusts the operating parameters of the search algorithm. In this embodiment, the search algorithm specifically employs an improved genetic algorithm. When the calculated spectral aliasing coefficient indicates a deepening of aliasing, it means that the spectral waveform distortion is severe and there are many local extrema in the solution space. At this time, the coarse calibration module automatically increases the number of iterations of the search algorithm and reduces the mutation probability. Through this strategy, the system can ensure population diversity while giving the algorithm more time to evolve and escape local optima, thereby accurately determining the initial range of the dual gas concentrations within a broad search space. This process strictly corresponds to the coarse calibration step in the first embodiment of the method, laying a solid foundation for subsequent fine processing.

[0178] Next, the system is equipped with a fine decoupling module. This module is the core computing unit of the system, responsible for achieving accurate concentration inversion within the initial range determined by coarse calibration. The fine decoupling module first constructs an input matrix, which is a high-dimensional data structure that integrates multi-source information, including the initial range, the characteristics of the optimal frequency band signal, and the aging characteristics of the equipment. To address the computational burden caused by high-dimensional data, this module incorporates a manifold learning algorithm unit, which can dynamically determine the dimensionality reduction dimension based on the spectral aliasing coefficient. For example, when aliasing is severe, it automatically retains a higher dimension to maintain feature resolution. Subsequently, the module maps the input matrix to a low-dimensional space and performs regression analysis. It is particularly worth mentioning that when performing regression analysis, this module integrates the dual-core co-evolution strategy described in Method Example 2, internally running the main regression kernel and the sensitive state tracking kernel in parallel. This allows the fine decoupling module to not only adapt to the slow aging of the equipment through incremental learning to obtain finely decoupled concentrations, but also to respond quickly using the sensitive state kernel when gas concentration undergoes a step change, thus resolving the technical contradiction of not being able to simultaneously achieve sensitivity and stability.

[0179] Finally, the system also includes a compensation and correction module. Located at the end of the signal processing link, this module eliminates interference from environmental factors on the measurement results. Based on the environmental parameters, the compensation and correction module dynamically compensates for the finely decoupled concentration, outputting the final accurate concentration of the two gases. This module internally stores a pre-trained gradient boosting regression model, capable of calculating nonlinear environmental impact compensation coefficients based on temperature and pressure data. More importantly, to address the sensor spatiotemporal mismatch problem mentioned in the background art, this module integrates the sensor hysteresis correction mechanism from Method Embodiment 3. When hysteresis is detected in the physical sensor data, this module can invert the instantaneous equivalent pressure using the spectral line type factor and use this equivalent value to replace the physical reading for compensation calculation. This design essentially gives the system the capability of a virtual sensor sensing at the speed of light, ensuring the reliability of the output results during sudden pressure changes.

[0180] In summary, the TDLAS dual-gas concentration synchronous calibration and decoupling system provided in this embodiment, under the presence of spectral aliasing interference, organically combines the calculation of spectral aliasing coefficients, dynamic adjustment of search strategies, dual-core regression analysis, and environmental compensation based on photophysical properties through close collaboration among its modules. This system not only achieves real-time monitoring and feedback of equipment aging characteristics at the hardware level, but also constructs a comprehensive processing logic at the algorithm level, from coarse to fine and from steady-state to transient states. This effectively overcomes the shortcomings of existing detection equipment, such as low accuracy, slow response, and poor anti-interference capabilities in complex dynamic environments, fully demonstrating the significant progress and beneficial effects of this invention compared to existing technologies.

[0181] Example 5:

[0182] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0183] like Figure 9 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0184] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0185] Bus 102 may include a path for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0186] The memory 103 stores a computer program corresponding to the TDLAS dual-gas concentration synchronization calibration and decoupling method in the presence of spectral aliasing interference as described in the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0187] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 9 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0188] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for simultaneous calibration and decoupling of TDLAS dual-gas concentrations in the presence of spectral aliasing interference, characterized in that, The method includes: The signal acquisition process involves acquiring environmental parameters, obtaining adaptively optimized multi-band harmonic signals, and selecting the optimal frequency band signal based on signal quality indicators. Perform the coarse calibration step: calculate the spectral aliasing coefficient based on the characteristics of the optimal frequency band signal, dynamically adjust the running parameters of the search algorithm based on the numerical range of the spectral aliasing coefficient, and determine the initial range of the dual gas concentrations within the search space; Perform the fine decoupling step: Construct an input matrix containing the initial range, the features of the optimal frequency band signal, and the equipment aging features. After dynamically determining the dimensionality reduction dimension based on the spectral aliasing coefficient, map the input matrix to a low-dimensional space and perform regression analysis to obtain the fine decoupling concentration. Perform compensation and correction steps: dynamically compensate the fine decoupling concentration based on the environmental parameters, and output the final accurate concentration of the two gases; The spectral aliasing coefficient represents the degree of attenuation of the current signal quality relative to the reference signal quality. The operating parameters of the dynamically adjusted search algorithm include: when the spectral aliasing coefficient indicates an increased degree of aliasing, increasing the number of iterations of the search algorithm and reducing the mutation probability.

2. The method according to claim 1, characterized in that, The process of acquiring adaptively optimized multi-band harmonic signals and determining the optimal frequency band signal based on signal quality indicators includes: Based on the deviation between driving current and wavelength recorded in the aging characteristics of the equipment, the scanning range of the laser driving current is corrected, and signals of different harmonic frequency bands are extracted simultaneously. The signal-to-noise ratio of each harmonic frequency band and the correlation coefficient between the signal in that frequency band and the preset standard reference spectrum are calculated respectively, and the product of the two is used as the comprehensive quality index. If the signal-to-noise ratio and correlation coefficient of a single frequency band signal both exceed the preset validity threshold, then the frequency band signal is selected as the optimal frequency band signal. If the signal-to-noise ratio fluctuation of a single frequency band signal exceeds a preset stability threshold, the peak-to-peak values ​​of the two frequency band signals with the highest overall quality index are fused, and the fused signal is taken as the optimal frequency band signal.

3. The method according to claim 1, characterized in that, The device aging characteristics include at least the deviation between drive current and wavelength mapping, the offset of characteristic points of wavelength selection devices, and the signal-to-noise ratio attenuation ratio. The fine decoupling step includes: vectorizing and concatenating the initial range, the optimal frequency band signal, the environmental parameters and the equipment aging characteristics to construct a high-dimensional input data matrix; The high-dimensional input data matrix is ​​learned by using a local linear embedding algorithm, wherein the target dimension of the local linear embedding algorithm is adaptively selected based on the spectral aliasing coefficient and the driving current and wavelength mapping deviation. A nonlinear mapping relationship between the dimensionality-reduced low-dimensional spatial data and the concentrations of the two gases is established using an incremental support vector regression model.

4. The method according to claim 3, characterized in that, The rules for adaptively selecting the target dimension based on the spectral aliasing coefficient and the deviation between the driving current and the wavelength mapping include: Set a first aliasing threshold and a second aliasing threshold, wherein the first aliasing threshold is less than the second aliasing threshold; When the spectral aliasing coefficient is less than the first aliasing threshold and the deviation between the driving current and the wavelength is within the first deviation range, the target dimension is set to the first dimension value. When the spectral aliasing coefficient is between the first aliasing threshold and the second aliasing threshold, or when the deviation between the driving current and the wavelength is in the second deviation range, the target dimension is set to the second dimension value. When the spectral aliasing coefficient is greater than the second aliasing threshold, or when the deviation between the driving current and the wavelength is in the third deviation range, the target dimension is set to the third dimension value. Wherein, the third dimension value is greater than the second dimension value, and the second dimension value is greater than the first dimension value.

5. The method according to claim 1, characterized in that, The compensation and correction steps include: The environmental parameters are input into a pre-trained gradient boosting regression model to obtain the environmental impact compensation coefficients. The final accurate concentration is calculated using temperature and pressure correction logic, which includes: using the accurate decoupled concentration as a reference, performing a product operation using the ratio of standard atmospheric pressure to current atmospheric pressure, the ratio of current temperature to standard temperature, and the environmental impact compensation coefficient.

6. The method according to claim 3, characterized in that, The method also includes performing an adaptive calibration process for device aging: Based on the numerical range of the equipment aging characteristics, the equipment status is divided into mild aging, moderate aging, and severe aging. If it is in the mild aging state, adjust the encoding range of the search algorithm and the kernel parameters of the support vector regression model; If the device is in the moderate aging state, a rapid recalibration process is triggered, and the mapping relationship between the driving current and the wavelength is updated by collecting characteristic wavelength points. If the device is in the severely aged state, switch to the dual-band fusion decoupling mode and generate a maintenance reminder signal.

7. The method according to claim 6, characterized in that, The criteria for classifying the equipment status include: setting a deviation threshold, a drift threshold, and a quality threshold; When the deviation between the driving current and the wavelength is less than the deviation threshold, the feature point offset is less than the drift threshold, and the signal-to-noise ratio attenuation ratio is greater than the quality threshold, it is determined to be the mild aging state. When any of the above parameters falls into the preset intermediate degradation range, it is determined to be the moderate aging state; When any of the above parameters exceeds the limit value pointed to by the intermediate degradation range, it is determined to be the severe aging state.

8. The method according to claim 1, characterized in that, A reference spectral library containing standard reference spectral data of the two gases to be tested is pre-constructed; the search algorithm adopts an improved genetic algorithm, which uses binary segmented encoding; the fitness function of the improved genetic algorithm is determined based on the concentration combination corresponding to the current population individuals, the fitting residual between the theoretical superimposed waveform constructed from the standard reference spectral data and the optimal frequency band signal; In the coarse calibration step, the offset of the feature point of the wavelength selection device is monitored. If the offset exceeds the preset time drift threshold, the number of bits of the binary segmented code is increased. The initial search center of the coarse calibration step is determined based on the final accurate concentration output at the previous time step.

9. The method according to claim 3, characterized in that, In the process of establishing a nonlinear mapping relationship through an incremental support vector regression model, a dual-core co-evolution strategy is adopted to respond to step abrupt changes in gas concentration. This strategy includes: The main regression kernel and the sensitive state tracking kernel are run in parallel, wherein the main regression kernel is configured with a first penalty factor and the sensitive state tracking kernel is configured with a second penalty factor, and the first penalty factor is greater than the second penalty factor; Real-time monitoring of the cumulative sum of the prediction residuals of the main regression kernel; When the accumulated sum is below a preset drift threshold, maintain steady-state mode, update only the master regression kernel, and freeze the sensitive tracking kernel; When the accumulated sum exceeds the preset drift threshold, it is determined that a mutation mode has been entered, the sensitive state tracking kernel is activated for fitting, and the output of the sensitive state tracking kernel is used as a temporary result; When the accumulated value falls back to the reset threshold, the boundary support vectors generated by the sensitive tracking kernel during the mutation mode are extracted and injected into the support vector set of the main regression kernel to complete the model state update.

10. The method according to claim 5, characterized in that, The compensation and correction step also includes a sensor hysteresis correction mechanism, specifically including: The spectral line shape factor is extracted from the optimal frequency band signal, and the spectral line shape factor characterizes the Lorentz broadening characteristics of the gas absorption spectral lines; Calculate the theoretical line-type factor based on the current environmental parameters, and calculate the morphological deviation between the spectral line-type factor and the theoretical line-type factor; When the morphological deviation exceeds the preset hysteresis determination threshold, the spectral line shape factor is inverted into instantaneous equivalent air pressure using a preset mapping relationship. The instantaneous equivalent air pressure is used to replace the air pressure value in the environmental parameters, and the compensation correction step is executed again.

Images