Nonlinear deconvolution multi-gas concentration retrieval method based on spectral prior constraint

Abstract

The present application relates to the technical field of industrial gas concentration inversion, and specifically discloses a nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraint, comprising the following steps: obtaining original multi-source data of a target area and preprocessing to obtain a standardized data set, wherein the original multi-source data comprises spectral data, tracer gas concentration data and environmental parameter data; performing data quality grading on the standardized data set to obtain a data quality grading result. The present application constructs an elastic spectral prior model through data quality grading, realizes dynamic adjustment of the strength of physical constraint, can improve precision by using tight constraint under high-quality data, and can prevent solving failure by elastic relaxation under low-quality data; the present application further introduces a dynamic model switching mechanism based on working condition complexity and a hysteresis buffer and resource compensation strategy, effectively solving the resource conflict problem between high-precision full-quantity model and limited edge computing power.

Description

Technical Field

[0001] This invention relates to the field of industrial gas concentration inversion technology, specifically to a nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints. Background Technology

[0002] Currently, in the field of emission monitoring in industries such as thermal power generation and petrochemicals, multi-gas concentration inversion technology based on spectral analysis has become the mainstream approach. Traditional methods typically rely on standard nonlinear deconvolution algorithms to infer gas concentrations by fitting measured spectra to a standard database. However, in actual industrial settings, monitoring equipment is constantly exposed to complex and extreme environments with high temperatures (typically exceeding 70°C), high humidity (relative humidity reaching 99% RH or even condensation), and high dust levels (particulate matter concentration exceeding 50 mg / m³), inevitably leading to baseline drift and nonlinear noise interference in sensor signals. Existing inversion methods often employ fixed-strength physical prior constraints, such as fixed spectral line positions or intensity limits, to assist in the solution. This static constraint approach has significant limitations: when data quality is good, overly broad constraints cannot effectively filter out noise; while when severe environmental interference causes signal distortion, overly strict constraints can lead to divergent solutions or getting trapped in local optima, making it difficult to guarantee the robustness of the inversion under complex and variable industrial conditions.

[0003] More importantly, existing technical solutions suffer from an irreconcilable conflict between inversion accuracy and real-time response capability. For example, to cope with complex nonlinear interference, theoretically, a high-dimensional full physical model is required for refined iterative solutions. However, this consumes enormous computing resources and time, which can easily lead to data delays or even system crashes in edge computing devices due to computing power exhaustion under sudden operating conditions.

[0004] Conversely, if the model is simplified in pursuit of real-time performance, it will lose accuracy in scenarios involving trace gas monitoring or complex gas mixtures, leading to missed or false alarms.

[0005] Therefore, how to meet the model complexity requirements under extreme working conditions and ensure millisecond-level response under sudden accidents at the edge where computing power is limited, while solving the problem of physical constraint failure caused by data quality fluctuations, is the core problem that current industrial-grade multi-gas monitoring technology urgently needs to solve. Summary of the Invention

[0006] 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 nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints, thereby improving the reliability of industrial safety monitoring.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a nonlinear deconvolution method for multi-gas concentration inversion based on spectral prior constraints, applied in a data processing terminal, comprising:

[0008] The raw multi-source data of the target area is acquired and preprocessed to obtain a standardized dataset. The raw multi-source data includes spectral data, tracer gas concentration data and environmental parameter data.

[0009] The standardized dataset is graded for data quality to obtain the data quality grading results.

[0010] Based on the data quality classification results, an elastic spectral prior model is constructed. The elastic spectral prior model is formed by combining physical prior information from a high-precision spectral database and dynamically adjusting the constraint strength of the physical prior information on the inversion process.

[0011] Calculate the operating condition complexity index, and select the corresponding nonlinear deconvolution inversion model from the preset model library accordingly;

[0012] Based on the dual constraints of the elastic spectral prior model and the tracer gas concentration data, the standardized dataset is iteratively solved using the nonlinear deconvolution inversion model to obtain multi-gas concentration inversion results.

[0013] To achieve the above objectives, a second aspect of the present invention proposes a nonlinear deconvolution multi-gas concentration inversion system based on spectral prior constraints, the system comprising:

[0014] The data acquisition and preprocessing module is used to acquire raw multi-source data of the target area and preprocess it to obtain a standardized dataset. The raw multi-source data includes spectral data, tracer gas concentration data and environmental parameter data.

[0015] The data quality grading module is used to grade the data quality of the standardized dataset and obtain the data quality grading results.

[0016] The elastic spectral prior construction module is used to construct an elastic spectral prior model based on the data quality classification results. The elastic spectral prior model is formed by combining physical prior information from a high-precision spectral database and dynamically adjusting the constraint strength of the physical prior information on the inversion process.

[0017] The working condition adaptation and model selection module is used to calculate the working condition complexity index and select the corresponding nonlinear deconvolution inversion model from the preset model library accordingly.

[0018] The dual-constraint inversion module is used to iteratively solve the standardized dataset using the nonlinear deconvolution inversion model based on the dual constraints formed by the elastic spectral prior model and the tracer gas concentration data, to obtain multi-gas concentration inversion results.

[0019] 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 nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] The nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints in this invention first constructs an elastic spectral prior model through data quality grading, realizing dynamic adjustment of the physical constraint strength. It can improve accuracy with tight constraints under high-quality data, and prevent solution failure by elastic loosening under low-quality data.

[0022] Secondly, this method introduces a dynamic model switching mechanism based on the complexity of the working conditions, as well as a hysteresis buffer and resource compensation strategy, which effectively solves the resource conflict problem between high-precision full model and limited edge computing power while ensuring computational stability.

[0023] Furthermore, by introducing the physical anchoring effect of tracer gas to form a dual constraint, and combining it with the direct-through mechanism of spectral reconstruction residuals, this method can not only use physical truth values ​​to calibrate inversion results in routine monitoring, but also skip the cumbersome iterative process and directly output high-confidence results in the event of sudden conditions such as high-concentration leaks. This achieves a perfect balance between high accuracy and low latency under all operating conditions, significantly improving the reliability of industrial safety monitoring. Attached Figure Description

[0024] 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:

[0025] Figure 1 This is a flowchart illustrating the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints provided by the present invention.

[0026] Figure 2 This is a simulation comparison of the effects of dynamic drift compensation of spectral signals before and after in the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints provided by this invention.

[0027] Figure 3This is a schematic diagram of the elastic prior constraint intervals under different data quality levels in the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints provided by the present invention.

[0028] Figure 4 This is a diagram of the nonlinear deconvolution iterative convergence trajectory under dual constraints in the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints provided by this invention.

[0029] Figure 5 This is a simulation diagram of the dynamic switching time of the model based on the hysteresis comparator in the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints provided by this invention.

[0030] Figure 6 This is a schematic diagram of the solution space compression under the prior compensation degradation mode in the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints provided by the present invention.

[0031] Figure 7 This is a comparison chart of spectral reconstruction residuals for high-confidence burst conditions and computational errors in the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints provided by this invention.

[0032] Figure 8 This is a schematic diagram illustrating the implementation of the nonlinear deconvolution multi-gas concentration inversion system based on spectral prior constraints provided by the present invention.

[0033] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0034] 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.

[0035] The following description, with reference to the accompanying drawings, describes an embodiment of the present invention: a nonlinear deconvolution method, system, and electronic device for multi-gas concentration inversion based on spectral prior constraints.

[0036] Example 1:

[0037] See Figure 1This embodiment details a nonlinear deconvolution method for multi-gas concentration inversion based on spectral prior constraints. This method is primarily applied to data processing terminals with edge computing capabilities, such as explosion-proof industrial computers or embedded smart gateways installed in industrial sites like thermal power plants and chemical industrial parks. This embodiment, through a series of rigorous logical steps, aims to resolve the contradiction between gas concentration monitoring accuracy and real-time performance caused by data quality fluctuations, complex and variable operating conditions, and limited edge computing power in extreme industrial environments.

[0038] Step S1: Obtain the original multi-source data of the target area and preprocess it to obtain a standardized dataset.

[0039] The data processing terminal first needs to collect multi-dimensional raw data from the target area. This raw, multi-source data forms the basis for subsequent inversion calculations, and specifically includes, but is not limited to, spectral data, tracer gas concentration data, and environmental parameter data.

[0040] For example, spectral data is acquired in real time by high-precision spectrometers deployed on-site, such as Fourier Transform Infrared (FTIR) or Tunable Diode Laser Absorption Spectrometers (TDLAS). These spectrometers capture the absorption spectra of gas mixtures at specific wavelengths, containing characteristic fingerprint information of target gases such as sulfur dioxide, nitrogen oxides, and carbon monoxide. Tracer gas concentration data refers to the concentration values ​​detected by mass spectrometers or dedicated sensors after a chemically inert gas (such as sulfur hexafluoride or perfluorocarbons) is injected into a specific emission port using an industrial-grade corrosion-resistant injection device. The introduction of tracer gases provides an absolute reference anchor for subsequent source apportionment and physical constraints. Environmental parameter data is acquired by equipment such as temperature and humidity sensors, air pressure sensors, and dust concentration detectors to characterize the current monitoring environment.

[0041] It is also important to note that in extreme industrial environments, the raw signals acquired by sensors often contain various noises and drift, which can lead to inversion failure if used directly. Therefore, preprocessing of the raw multi-source data is necessary. In this embodiment, the core step of preprocessing includes dynamic drift compensation.

[0042] The acquisition and preprocessing of raw multi-source data of the target area includes dynamic drift compensation, which specifically includes the following sub-steps:

[0043] First, the system determines the baseline drift correction based on a pre-defined environmental and baseline drift correlation model. This model, trained on long-term historical data, quantitatively describes the nonlinear effects of temperature, humidity, and air pressure changes on the spectrometer baseline. The system reads real-time environmental parameters, inputs them into the model, and outputs the baseline drift correction for the current moment, denoted as [reference needed]. .

[0044] Secondly, the system acquires real-time dust concentration values ​​and calculates dust interference terms using preset dust correction coefficients. High concentrations of dust in industrial environments can cause Mie or Rayleigh scattering in the optical path, leading to an overall decrease in spectral intensity and a tilted baseline. The system reads the real-time dust concentration value through a dust sensor and records it as follows: Combined with a pre-calibrated dust correction factor, it is denoted as... The specific amount of interference caused by dust to the spectral signal was calculated.

[0045] Finally, the system subtracts the baseline drift correction and the dust interference term from the original spectral signal to obtain the corrected spectral signal. The mathematical expression of this process is shown below:

[0046] ;

[0047] In the formula, This represents the standardized spectral signal after correction, which is the direct input to the subsequent inversion algorithm; This represents the raw spectral signal collected by the sensor, which includes real gas absorption signals as well as various interference signals; This represents the amount of baseline drift correction caused by environmental factors; The dust correction factor reflects the ability of a unit concentration of dust to attenuate the spectral signal, and is usually determined experimentally. This represents the real-time dust concentration value. Through this calculation, the system effectively eliminates interference from the environment and dust, generating a standardized dataset.

[0048] like Figure 2 The simulation diagram shows the effect comparison before and after dynamic drift compensation of the spectral signal in this embodiment. Figure 2 The horizontal axis represents the spectral wavelength in micrometers, corresponding to the sensor's acquisition range in the 2-15 micrometer band; the vertical axis represents the spectral intensity, reflecting the degree of absorption of light energy by gas molecules.

[0049] Figure 2The red curve at the top center represents the unprocessed raw spectral signal, simulating the actual conditions captured by the monitoring equipment under extreme conditions of 70 degrees Celsius, 99% relative humidity, and high concentrations of dust. It can be seen that due to the influence of the high-temperature environment on the energy level transitions of gas molecules, the characteristic absorption peaks at 4.5 μm, 7.2 μm, and 11.5 μm exhibit significant thermal broadening, with attenuation of peak intensity and a slight irregular frequency shift in the center wavelength. Simultaneously, due to the combined effects of Mie scattering caused by high-concentration dust and water vapor interference, the raw spectral signal is superimposed with a large amount of random, sharp noise spikes, and the overall baseline exhibits severe nonlinear upward sag and distortion, causing the weak effective signal to be almost completely masked by background interference.

[0050] Figure 2 The blue curve at the bottom center represents the standardized spectral signal obtained after processing by the dynamic drift compensation algorithm described in this invention and correcting for environmental parameters. A comparison clearly shows that the corrected signal successfully eliminates the nonlinear baseline drift component in the 1 to 1.5 intensity unit range and effectively filters out random spikes caused by dust. At this point, the full width at half maximum (FWHM) and center position of the characteristic absorption peaks have been corrected and returned to a standard Gaussian distribution. The signal baseline is flat and stable near zero, resulting in a significant improvement in the signal-to-noise ratio. This provides a reliable data foundation for subsequent high-precision nonlinear deconvolution inversion.

[0051] Step S2: Perform data quality grading on the standardized dataset to obtain the data quality grading results.

[0052] After obtaining a standardized dataset, it cannot be blindly fed directly into the inversion model, because data of different qualities have completely different degrees of dependence on prior constraints. Therefore, the system needs to perform a fine-grained quality assessment of the data.

[0053] The process of classifying the data quality of the standardized dataset includes the following detailed steps:

[0054] The system first calculates five key dimensions of the standardized dataset: signal-to-noise ratio (SNR), spectral overlap, proportion of interference peaks, environmental interference level, and operational stability level. The SNR reflects the ratio of the effective signal to the background noise, directly determining the detection limit of trace gases. Spectral overlap reflects the degree of obstruction between absorption peaks of different components in the gas mixture; higher overlap increases the difficulty of deconvolution. The proportion of interference peaks reflects the distribution of non-target gases, such as water vapor absorption peaks, within the analytical band. The environmental interference level is an indicator derived from the degree of rapid changes in temperature and humidity. The operational stability level is calculated based on the fluctuation rate of industrial production load.

[0055] For example, the system compares each calculated indicator with a preset first-level threshold set, which is a multi-dimensional logical judgment process.

[0056] Optionally, when the following stringent conditions are met: the signal-to-noise ratio is greater than a preset signal-to-noise ratio threshold, and the spectral overlap, interference peak ratio, environmental interference level, and operating condition stability level are all less than their respective evaluation thresholds, the system determines the data quality grading result to be the first quality level. The first quality level represents extremely pure data, stable operating conditions, and minimal interference; it is also known as superior data.

[0057] Conversely, when the data quality is poor, the system employs a more lenient judgment logic. Specifically, if any one of the signal-to-noise ratio, spectral overlap, interference peak ratio, environmental interference level, or operating condition stability level meets a preset poor quality condition, the data quality classification result is determined to be the second quality level. The second quality level represents data with significant noise, severe overlap, or under drastic fluctuation conditions; it is also referred to as inferior or poor-quality data. Through this binary or multi-dimensional classification mechanism, the system accurately identifies the difficulty level of the current processing object.

[0058] Step S3: Construct an elastic spectral prior model based on the data quality grading results.

[0059] Traditional inversion methods often use fixed physical constraints, which become rigid when faced with varying data quality. The elastic spectral prior model proposed in this embodiment is formed by combining physical prior information from a high-precision spectral database and dynamically adjusting the constraint strength of the physical prior information on the inversion process.

[0060] High-precision spectral databases typically refer to authoritative databases such as the High Resolution Transmission Molecular Absorption Database (HITRAN) or GEISA, which contain the true physical values ​​of gas molecules, such as spectral line positions, relative intensities, pressure broadening coefficients, and temperature dependence coefficients.

[0061] The process of constructing an elastic spectral prior model based on the data quality grading results specifically includes adopting drastically different constraint strategies for different quality levels:

[0062] In response to the data quality grading result being at the first quality level, the system sets a first constraint interval. Since the data quality is excellent and the signal-to-noise ratio is high, the system has reason to believe that the subtle features of the measured spectrum reflect the true physical condition; therefore, strict physical constraints are required to approximate the true value. The first constraint interval limits the deviations in spectral line position and intensity to a preset first deviation threshold. For example, the system will enforce that the drift of the spectral line center position during the inversion process must not exceed an extremely small nanometer value, and the rate of change in spectral line intensity must not exceed an extremely small percentage. This tight constraint strategy can significantly improve inversion accuracy and prevent overfitting under high-quality data.

[0063] In response to the data quality grading result being at the second quality level, the system sets a second constraint interval. At this point, the data is filled with noise and interference. If strict constraints are continued, the inversion algorithm may fail to converge or forcibly fit the noise. Therefore, the system adopts a flexible loosening strategy. The second constraint interval restricts spectral line position and intensity deviations by a greater range than the first deviation threshold, wherein the data quality at the second quality level is lower than that at the first quality level, thus achieving flexible loosening of the constraints on physical prior information. This means the system allows the inversion result to have a larger fluctuation range near the physical truth value, sacrificing the solution space for the algorithm's convergence and robustness, and avoiding solution failure due to data distortion.

[0064] like Figure 3 This is a schematic diagram of the constraint interval when constructing the elastic spectrum prior model in this embodiment. Figure 3 The horizontal axis represents the wavelength position of the spectrum, and the vertical axis represents the normalized spectral intensity. The solid black line curve in the center of the graph represents the standard physical spectral line retrieved from a high-precision spectral database, which serves as the absolute physical truth reference for the inversion process.

[0065] Figure 3 Around the spectral peaks, two rectangular shaded regions of different ranges are shown to visually represent the flexible solution spaces set for different data quality levels. The inner, dark rectangular region represents the first constraint interval set for the first quality level, i.e., superior data. This interval has extremely narrow boundaries in both wavelength position and intensity dimensions, forming a tight constraint that forces the inversion results to closely approximate the physical truth, thus fully utilizing the signal-to-noise ratio advantage of high-quality data to improve accuracy. The outer, light-colored rectangular region represents the second constraint interval set for the second quality level, i.e., poor data. Compared to the inner region, this interval has significantly expanded tolerances for both wavelength drift and intensity fluctuations, allowing for a flexible loosening of physical prior information.

[0066] This design allows the algorithm to allow the inversion results to fluctuate within a wider physically reasonable range when processing data containing a lot of noise or interference. This avoids problems such as non-convergence of iteration or forced fitting of noise caused by overly rigid constraints, and effectively ensures the robustness of the system under complex working conditions.

[0067] Step S4: Calculate the operating condition complexity index, and select the corresponding nonlinear deconvolution inversion model from the preset model library accordingly.

[0068] Because computing resources in industrial settings, i.e., edge computing power, are limited, and the complexity and accuracy of inversion models are usually directly proportional, this embodiment introduces a dynamic model switching mechanism to achieve optimal results with limited resources.

[0069] The computational operating condition complexity index includes a weighted summation of the number of industrial target gas types, spectral overlap, proportion of low-concentration components, environmental interference level, and operating condition stability level using preset weighting coefficients. This index quantifies the difficulty of the inversion task at the current moment. Mathematically, the operating condition complexity index can be expressed as the sum of the products of each sub-index and its corresponding weight.

[0070] Subsequently, the system selects a model from a pre-set model library based on the calculated metrics. The pre-set model library stores three different levels of models:

[0071] In the first scenario, if the complexity index of the operating condition is less than the first complexity threshold, it indicates that the current operating condition is very simple, and the system invokes the ultra-lightweight model. The ultra-lightweight model is configured to retain only the prior information of the core spectral line positions and intensities.

[0072] In the second scenario, if the operational complexity index falls between the first and second complexity thresholds, it indicates that the operational condition is at a normal level of complexity, and the system invokes the standard model. The standard model is configured to include complete prior knowledge and simplified multi-parameter coupling.

[0073] In the third scenario, if the complexity index of the operating condition exceeds the second complexity threshold, it indicates that the operating condition is extremely complex, and the system invokes the full model. The full model is configured to include complete prior knowledge and joint inversion of all parameters.

[0074] Step S5: Based on the dual constraints of the elastic spectral prior model and the tracer gas concentration data, the standardized dataset is iteratively solved using the nonlinear deconvolution inversion model to obtain the multi-gas concentration inversion results.

[0075] This is the calculation step for generating the final concentration data. The nonlinear deconvolution inversion model is not only constrained by the aforementioned elastic spectral prior (physical constraint), but also by the tracer gas data (source resolution constraint), forming a powerful dual constraint system.

[0076] The dual constraint based on the elastic spectral prior model and the tracer gas concentration data is implemented as follows:

[0077] The system utilizes tracer gas data to decouple mixed emission sources. First, a tracer gas emission contribution rate model is constructed based on the tracer gas injection concentration of each unit and the emission contribution rate to be solved. The principle is that the total tracer gas volume at the main outlet equals the sum of the emissions from each branch unit. Mathematically, the total tracer gas concentration at the main outlet is represented as the sum of the products of the injection concentration of each unit and its corresponding emission contribution rate. The formula can be expressed as:

[0078] ;

[0079] in, This represents the total concentration of tracer gas detected at the main discharge outlet; Represents the total number of generating units; The index number representing the generator set; Representing the Known tracer gas concentrations injected into each unit; Represents the first to be solved The emission contribution rate of each unit.

[0080] Next, the system uses constrained least squares to solve for the emission contribution rate of each unit. During the solution process, a hard constraint in a physical sense must be satisfied, namely, the sum of all emission contribution rates must be 1. And the contribution rate of a single emission is between 0 and 1. .

[0081] After calculating the contribution rate of each unit Subsequently, the system uses the obtained emission contribution rate to correct the target gas concentration for each unit. The system uses the corrected concentration as the input constraint condition for the nonlinear deconvolution inversion model, substitutes it into the iterative algorithm, and finally outputs the accurate multi-gas concentration inversion results for each unit and the total exhaust outlet.

[0082] like Figure 4 This example demonstrates a comparison of convergence performance when using a nonlinear deconvolution model for iterative solution. Figure 4 The horizontal axis represents the number of iterations performed by the algorithm, reflecting the time cost and resource consumption of the computation; the vertical axis represents the residual value of the objective function, and the lower the value, the higher the degree of fit between the inversion result and the physical truth.

[0083] Figure 4The solid red line represents the convergence trajectory after employing the dual constraints of the elastic spectral prior model and tracer gas concentration data described in this invention. It can be seen that, thanks to the decoupling effect of the tracer gas's contribution to the emission source and the effective compression of the solution space by the elastic physical prior, the curve exhibits a rapid decreasing trend in the first five to ten iterations, and quickly stabilizes at an extremely low objective function residual level. This is consistent with the technical characteristic of setting the number of iterations in the fine-tuning stage within the range of ten to twenty.

[0084] In comparison, Figure 4 The blue dashed curve represents the method using only traditional single constraints or static constraints, which has a significantly slower convergence speed and is accompanied by oscillations. The final convergence residual value is also significantly higher than that of the method of this invention.

[0085] This comparison demonstrates that the dual constraint mechanism can significantly reduce unnecessary computational overhead when solving complex nonlinear inversion problems, while greatly improving the final gas concentration inversion accuracy, ensuring the efficient operation of the algorithm on edge computing devices.

[0086] Step S6: In large-scale scenarios such as industrial parks, multiple edge data processing terminals and one core node are typically deployed. To further improve the inversion accuracy, this embodiment employs a federated learning collaborative mechanism.

[0087] The iterative solution of the standardized dataset using the nonlinear deconvolution inversion model includes performing federated learning collaborative steps, as follows:

[0088] First, each data processing terminal at the edge performs a local coarse inversion based on the ultra-lightweight model to obtain preliminary results of gas concentration. This step quickly completes the feature extraction of local data.

[0089] Next, the edge terminal receives global model parameters from the core node. These global model parameters are obtained by the core node assigning weights based on the data quality classification results of each data processing terminal and then fusing them.

[0090] Subsequently, the edge terminal performs a collaborative consistency check. The system calculates the deviation between the local coarse inversion result and the global optimal solution determined based on the global model parameters.

[0091] When the deviation exceeds a preset consistency threshold, the data processing terminal is triggered to load the standard model and recalculate the standardized dataset. This coarse-to-fine calculation mechanism leverages collective wisdom while ensuring accuracy at individual points.

[0092] Step S7: Multi-dimensional anomaly verification and closed-loop adjustment. After obtaining the multi-gas concentration inversion results, the process also includes multi-dimensional anomaly verification and closed-loop adjustment, specifically:

[0093] The system first calculates the sum of squared spectral errors of the inversion results and verifies whether the inversion results exceed the preset industrial emission range and historical average range. If an anomaly is detected, the system triggers closed-loop adjustment based on the anomaly type:

[0094] In the first scenario, in response to abnormal data quality, the sensor is automatically purged and recalculated for drift compensation.

[0095] In the second scenario, in response to an anomaly in the prior constraints, an adjustment is triggered to the constraint range of the elastic spectral prior model.

[0096] In the third scenario, in response to abnormal operating conditions, real-time coal quality and load data are acquired to correct model parameters.

[0097] By organically combining the above steps S1 to S7, this embodiment realizes a complete and adaptive industrial-grade multi-gas concentration inversion process.

[0098] Example 2:

[0099] This embodiment, building upon Embodiment 1, further details the core strategy for dynamic model switching control in the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints. The key issues addressed in this embodiment are: how to avoid repeated oscillations of the inversion model under critical conditions of frequent fluctuations in operating conditions, and how to maintain high-precision monitoring capabilities through algorithmic strategy adjustments in extreme cases where edge device computing resources are exhausted. The specific execution process of this embodiment includes three main stages: calculating and smoothing the operating condition complexity index, executing hysteresis model switching judgment, and executing resource conflict detection and prior compensation.

[0100] For example, calculating the complexity index of smooth operating conditions is the first step in achieving stable control:

[0101] In industrial settings, due to the instability of the combustion process or airflow disturbances, the real-time calculated operating condition complexity index often contains high-frequency random noise. Directly switching models based on instantaneous values ​​would cause drastic jumps between models of different complexities. Therefore, the system needs to calculate a smoothed operating condition complexity index. This smoothed operating condition complexity index is a weighted moving average of the operating condition complexity index within a preset time window. The data processing terminal maintains a length of [missing value] in memory. A time window queue used to store the most recent The instantaneous operating condition complexity index is calculated at each sampling time.

[0102] Specifically, the formula for calculating the complexity index of smooth operating conditions is expressed as follows:

[0103] ;

[0104] In the formula, Represents the current moment The calculated complexity index of the smooth operating condition; This represents the length of the preset time window, corresponding to the number of data sampling points; The index representing the time step of the backtracking; Representing a historical moment The calculated instantaneous operating condition complexity index; Represents a historical moment The weighting coefficients are determined. Through this weighted moving average algorithm, the system effectively filters out index spikes caused by instantaneous sensor fluctuations or minor environmental disturbances, providing a reliable data foundation for subsequent stable decision-making.

[0105] After obtaining a smoothed index, the system does not use a single threshold for simple logical judgment, but instead introduces the design concept of a hysteresis comparator. The system presets two key decision thresholds: an upgrade threshold and a downgrade threshold, and strictly sets the upgrade threshold to be greater than the downgrade threshold, thus creating a buffer dead zone between the two thresholds. Furthermore, to further enhance the system's anti-interference capability, a time-dimensional constraint is introduced: a first decision period and a second decision period.

[0106] Specifically, the system only performs a model upgrade operation, i.e., switches to a higher-complexity inversion model, when the smoothing condition complexity index exceeds the upgrade threshold and the duration exceeds the first determination period. This means that even if the condition complexity index spikes instantaneously and exceeds the upgrade threshold, if its duration is short and does not meet the requirements of the first determination period, the system will still consider it as an occasional pulse interference and maintain the current model unchanged.

[0107] Accordingly, the system only performs model degradation when the complexity index of the smoothed operating condition is less than the degradation threshold and the duration exceeds the second judgment period. This logic ensures that the system will only release the resources occupied by the high-precision model after the operating condition has indeed returned to stability and been maintained for a period of time. Through this dual hysteresis mechanism, the system can remain inert when facing critical operating conditions, greatly reducing the frequency of model switching and ensuring the continuity of monitoring data and the thermal stability of the system.

[0108] like Figure 5 The timing response characteristics of the dynamic switching process of the model based on the hysteresis comparator in this embodiment are demonstrated. Figure 5The horizontal axis represents the continuous operating time of the system, the vertical axis on the left corresponds to the value of the operating condition complexity index, and the vertical axis on the right corresponds to the complexity level of the inversion model.

[0109] Figure 5 The light blue thin waveform curve represents the instantaneous operating condition complexity index calculated by the system in real time. This curve exhibits significant high-frequency random fluctuations due to the influence of on-site noise. The dark blue solid curve superimposed on it represents the smoothed operating condition complexity index after weighted moving average processing, which filters out noise and reflects the true trend of operating condition changes.

[0110] Figure 5 The two horizontal dashed lines in the middle indicate the system's preset upgrade threshold (0.43) and downgrade threshold (0.37), respectively, and the two form a hysteresis buffer zone.

[0111] Figure 5 The black stepped solid lines in the diagram represent the model states of the system's final decision, where: Level 1 is the ultra-lightweight model and Level 2 is the standard model.

[0112] from Figure 5 The correspondence between the curves clearly shows that when the smoothing complexity index fluctuates within the buffer zone between the upgrade and downgrade thresholds, the black model state curve remains horizontal regardless of its changes, without any jumps. Only when the smoothing curve clearly breaks through the upgrade threshold does the model state transition upwards; conversely, only when the smoothing curve clearly falls below the downgrade threshold does the model state fall downwards. This curve trend intuitively demonstrates that the hysteresis comparator can effectively shield signal jitter under critical conditions, avoiding frequent and repeated switching of the model between different levels, thereby ensuring the stability of system operation.

[0113] Next, in edge computing scenarios of the Industrial Internet of Things (IIoT), data processing terminals often need to simultaneously undertake multiple tasks such as data acquisition, encrypted transmission, and alarm logic judgment. When the operating conditions are extremely complex, the lag judgment logic may instruct the system to switch to the full model. However, if the CPU or memory resources of the edge terminal are already occupied by other high-priority processes, forcibly loading the full model may lead to computation timeouts. Therefore, when determining that a switch to the full model is necessary, the system must first detect the real-time computing resource utilization of the data processing terminal.

[0114] Specifically, the system reads the kernel state of the operating system to obtain the current CPU load rate, memory usage rate, etc., and comprehensively evaluates to obtain the real-time computing resource utilization rate; then, the system compares this utilization rate with the preset resource utilization threshold.

[0115] Optionally, a priori compensation degradation mode can be activated to resolve the conflict between computing power and accuracy: if the real-time computing resource utilization rate exceeds a preset resource utilization threshold, it indicates that the device is currently under high load and cannot support the operation of the full model. In this case, the system automatically activates the priori compensation degradation mode. In this mode, the system performs two key operations to achieve a soft landing:

[0116] The first step is to force the inversion to use the standard model instead of the full model. The standard model only includes simplified multi-parameter coupling, and its computational cost is much lower than that of the full model, ensuring timely completion of the computational task even with limited resources.

[0117] The second step involves simultaneously applying a preset prior contraction coefficient to compress and correct the constraint interval of the elastic spectral prior model. This is the ingenious aspect of this embodiment. Because the standard model simplifies some nonlinear characteristics, its theoretical accuracy under complex conditions is lower than that of the full model. To compensate for this accuracy loss, the system utilizes the determinism of physical prior information to suppress the uncertainty of mathematical solutions. The system introduces a positive number less than 1 as a prior contraction coefficient to shrink the constraint interval originally determined by data quality grading.

[0118] Specifically, the compensation constraint interval after compression and correction can be defined by the following formula:

[0119] ;

[0120] In the formula, This represents the new constraint interval width used in the prior compensation degradation mode, such as the maximum allowable deviation of spectral line positions; This represents the preset prior contraction coefficient, whose value is strictly limited to between 0 and 1, with a typical value of 0.5. This represents the width of the elastic constraint interval originally determined based on the current data quality classification results.

[0121] like Figure 6 The transformation mechanism of the solution space of the inversion algorithm under the prior compensation degradation mode in this embodiment is revealed. Figure 6 The diagram shows a two-dimensional parameter solution plane, where the horizontal and vertical axes represent different parameter dimensions in the inversion model, such as spectral line position deviation and intensity change rate.

[0122] Figure 6 The light-colored circular area at the outermost layer of the graph represents the original elastic constraint solution space determined based on the current data quality classification results. Within this range, the inversion algorithm typically has a large degree of search freedom to adapt to the random fluctuations in the data. Figure 6The smaller, dark circular areas inside the graphic represent the compensated compressed solution space formed after the original space is forcibly shrunk by applying a preset a priori compression coefficient after the resource conflict protection mechanism is activated. Figure 6 The star-shaped marker at the geometric center represents the physical truth anchor point in the high-precision spectral database.

[0123] A comparison clearly shows that when edge devices lack sufficient computing power and are forced to downgrade the full model to the standard model, the system artificially compresses the allowed solution space significantly towards the physical truth anchor point. This forces the simplified standard model to search for the optimal solution within an extremely narrow, physically reasonable range. This geometric constraint enhancement strategy effectively utilizes the determinism of physical priors to compensate for the insufficient computing power of the model, preventing the simplified model from diverging due to fitting noise in a wide search space. It ensures that the inversion results still possess high physical reliability even under extremely resource-constrained conditions.

[0124] Through the aforementioned mathematical transformations, the system artificially narrows the solution space search range of the inversion algorithm. While this may theoretically sacrifice the ability to fit certain minute nonlinear details, in critical moments when computing power is insufficient, this strategy forces the inversion results to highly conform to the physical truth values ​​in the spectral database. This strategy of exchanging strong physical constraints for computational stability enables the standard model to output physically reliable solutions even under complex conditions, avoiding divergence or severe distortion of results caused by model simplification.

[0125] Example 3:

[0126] This embodiment, building upon Embodiments 1 and 2, further details the rapid response and authenticity discrimination logic under abnormal operating conditions in the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints. The key problem addressed in this embodiment is: in a multi-node collaborative inversion architecture, when the monitoring data of a certain edge node differs significantly from the global consensus, how to avoid blindly assuming it's a calculation error and triggering time-consuming model recalculation, thereby ensuring that in the event of a real, sudden high-concentration leak, the system can directly output a high-confidence alarm result without going through a cumbersome iterative process. The specific execution process of this embodiment includes key steps such as generating the theoretically reconstructed spectrum, calculating the spectral reconstruction residual, implementing the true / false anomaly diversion strategy, and activating the pass-through mechanism.

[0127] First, under the collaborative mechanism of federated learning, core nodes form global model parameters by aggregating data features from each edge terminal. When the deviation between the local coarse inversion result of a certain edge terminal and the global optimal solution determined based on the global model parameters exceeds a preset consistency threshold, it usually implies two possibilities:

[0128] First, the lightweight model of the edge terminal lacks accuracy; second, a localized, sudden high-concentration gas leak did indeed occur in the area where the edge terminal is located.

[0129] If it is the latter, blindly triggering recalculation will cause delays. Therefore, before triggering the data processing terminal to load the standard model and recalculate the standardized dataset, the system must first execute a physical-level introspection procedure, namely, a discrimination step based on the spectral reconstruction residuals.

[0130] Next, to verify the physical accuracy of the local coarse inversion results, the system uses the preliminary gas concentration result as input and performs a reverse inversion using a pre-set radiative transfer physical model. The radiative transfer physical model is based on the Lambert-Beer law and its nonlinear correction, describing the physical process of light being absorbed by gas molecules as it propagates through a medium. The system assumes that the currently inverted gas concentration is accurate and substitutes it into the physical equations to simulate the spectral pattern that the sensor should receive under ideal conditions.

[0131] Specifically, the process of generating the theoretically reconstructed spectrum can be defined by the following formula:

[0132] ;

[0133] In the formula, The intensity of the theoretically reconstructed spectrum generated by the system calculation is the wavelength. The function; The background baseline spectrum, representing the spectrum after subtracting gas absorption, is typically obtained from the preprocessing stage. Represents the effective absorption path length of the optical path system; This represents the total number of components in the gas being analyzed. Index number representing the gas component; Representing the The gas at the current temperature and air pressure The absorption cross section data is obtained from a high-precision spectral database. The first one is obtained by inversion from the ultra-lightweight model. Preliminary results on the gas concentration of the gas. Through this physical equation, a theoretical model parallel to the measured data is constructed.

[0134] Then, after obtaining the theoretical spectrum, the system compares it with the standardized data generated in Example 1. A point-by-point comparison is performed. The spectral reconstruction residual between the theoretically reconstructed spectrum and the spectral data in the standardized dataset is calculated, which is the intensity difference vector at each wavelength. Subsequently, to quantify the overall degree of this difference, the system calculates the root mean square error of the spectral reconstruction residual.

[0135] Specifically, the root mean square error of the spectral reconstruction residual is calculated using the following formula:

[0136] ;

[0137] In the formula, The root mean square error represents the residual of spectral reconstruction. The total number of sampling channels for spectral data; The index number representing the spectral channel; Representing the The center wavelength value corresponding to each channel; This represents the standardized dataset, i.e., the corrected spectrum. At wavelength The measured spectral intensity at a given location. This indicator objectively measures whether the calculated concentration truly matches the measured spectrum.

[0138] Next, the system makes a qualitative judgment on the current abnormal state based on the calculated root mean square error value. The core logic here is that if the calculation error is caused by a rudimentary model, the theoretical spectrum generated after substituting the erroneous result into the physical model usually cannot coincide with the measured spectrum, resulting in a large residual. Conversely, if it is a real sudden leak, although the concentration value is high and different from the surrounding nodes, this high concentration value is precisely the only solution that can perfectly explain the measured spectrum, so the residual will be very small.

[0139] The specific judgment logic is as follows: if the deviation exceeds the consistency threshold, it indicates that the data is inconsistent with the global data; however, if the root mean square error is less than the preset physical confidence threshold, it indicates that the data is physically highly consistent. At this point, the system determines the current state to be a high-confidence sudden event. This signifies that the system confirms that the currently detected high concentration is not a false alarm from the algorithm, but a real sudden event.

[0140] like Figure 7 The stark contrast between the two sub-figures on the left and right illustrates in detail the core mechanism of true and false anomaly discrimination based on spectral reconstruction residuals in this embodiment.

[0141] Figure 7The left subplot illustrates the spectral characteristics under scenario one, i.e., a high-confidence sudden operation. The solid blue line represents the measured spectrum acquired by the sensor, containing the absorption characteristics of the actual high-concentration gas. The dashed red line represents the theoretically reconstructed spectrum derived from the concentration value obtained through local coarse inversion combined with the physical model. It can be seen that in scenario one, despite the abnormally high concentration, the theoretical and measured curves exhibit a very high degree of overlap in peak shape, position, and intensity. The green shaded area between them, representing the spectral reconstruction residual, is extremely small and far below the preset physical confidence threshold. This fully demonstrates that the current high-concentration inversion result is physically self-consistent. Therefore, the system determines this as a genuine leak and immediately activates the pass-through mechanism.

[0142] Figure 7 The right-hand subplot illustrates the spectral characteristics of scenario two, where a calculation error leads to a false alarm. Although the inversion algorithm provides an incorrect high concentration value, the theoretical spectrum (red dashed line) generated after substituting it into the physical model differs significantly from the measured spectrum (blue). The measured spectrum may contain only non-gaseous interference signals or noise, failing to support the strong absorption peaks in the theoretical spectrum. This results in a large red shaded area—a significant spectral reconstruction residual. This physical inconsistency reveals that the high concentration result is a misjudgment by the algorithm. Based on this, the system intercepts the alarm and forcibly triggers a recalculation process of the standard model, effectively avoiding false alarms and ensuring the accuracy and seriousness of industrial monitoring.

[0143] Ultimately, the desired effect of this embodiment is to activate the pass-through mechanism to achieve millisecond-level response. Specifically:

[0144] Once a high-confidence emergency is identified, the system immediately activates the pass-through mechanism. This mechanism is an emergency bypass process designed to bypass all unnecessary computational steps. The pass-through mechanism is configured to perform the following three key operations:

[0145] First, the operation of loading the standard model is prohibited. In the normal process, excessive deviation will cause the system to attempt to switch to a more complex model to correct the error, but this will introduce a delay of seconds in model loading and initialization. The pass-through mechanism forcibly cuts off this recalculation path, maintaining the current running state of the ultra-lightweight model;

[0146] Second, the preliminary gas concentration result is directly used as the final output. Since the physical residual is extremely low, this indicates that the accuracy of the preliminary result already meets the requirements. The system immediately encapsulates this result into an alarm message and sends it to the control system via the industrial bus.

[0147] Third, a high-confidence weight adjustment instruction is sent to the core node. This is to prevent real high-concentration data from being diluted by low-concentration data from other nodes during the subsequent federated averaging process. The edge node informs the core node via the instruction: a real incident has occurred here, please trust this data. After receiving the instruction, the core node increases the weight of the edge terminal in the next round of global fusion, ensuring that the global model can quickly adapt to this sudden situation.

[0148] Through the implementation of the aforementioned direct-access mechanism, this embodiment constructs an intelligent monitoring system with self-identification capabilities. It not only utilizes federated learning to eliminate random errors but also escapes statistical pitfalls through physical verification at critical moments. This design perfectly resolves the contradictory demands of industrial sites for low false alarm rates, zero false alarm rates, and rapid response times.

[0149] Example 4:

[0150] See Figure 8 This embodiment provides a nonlinear deconvolution multi-gas concentration inversion system based on spectral prior constraints. This system typically runs as a software functional module on a data processing terminal with edge computing capabilities, such as an explosion-proof industrial control computer or an embedded intelligent gateway in an industrial setting. Addressing the technical challenges in existing technologies, such as sensor signal drift caused by monitoring equipment operating in extreme environments year-round, the inability of static constraints to adapt to changing operating conditions, and resource conflicts between edge computing power and high-precision models, this system achieves efficient data flow processing and intelligent decision-making through a modular design.

[0151] Specifically, the system mainly includes a data acquisition and preprocessing module, a data quality classification module, an elastic spectral prior construction module, a working condition adaptation and model selection module, and a dual-constraint inversion module.

[0152] First, the system is equipped with a data acquisition and preprocessing module. This module is the sensing front end of the entire system, responsible for establishing communication connections with the underlying hardware devices. Its primary task is to acquire raw multi-source data from the target area. This data includes spectral data collected by Fourier transform infrared spectrometers or laser spectrometers, tracer gas concentration data provided by industrial-grade injection devices and detectors, and environmental parameter data provided by thermometers, hygrometers, and dust meters. Given the interference of high temperature, high humidity, and high dust levels in industrial environments on signals, this module incorporates a dynamic drift compensation algorithm. Based on a preset environmental and baseline drift correlation model and real-time dust concentration values, it denoises and corrects the raw spectral signal, thereby eliminating baseline drift and dust scattering interference, ultimately obtaining a high-quality standardized dataset. This processing step effectively solves the problem of unavoidable nonlinear noise interference in sensor signals mentioned in the background technology, laying a solid foundation for subsequent processing.

[0153] Secondly, the system is connected to a data quality grading module. To overcome the drawbacks of the one-size-fits-all approach in traditional methods, this module performs refined quality assessments on the aforementioned standardized dataset. It divides the data into different quality levels, such as first quality level and second quality level, by calculating multi-dimensional indicators such as signal-to-noise ratio, spectral line overlap, and the proportion of interference peaks. The data quality grading result output by this module is not merely a label, but also a guide for adjusting the strategies of subsequent modules. Through this grading mechanism, the system can intelligently identify the reliability of the current data, thereby avoiding resource waste or result divergence caused by forcibly performing high-precision calculations when the data quality is extremely poor.

[0154] Next, the core of the system is configured with an elastic spectral prior construction module. This module aims to address the limitations of static constraints in existing inversion methods. It stores and calls upon a high-precision spectral database, which contains physical truth values ​​such as the positions and intensities of standard spectral lines for gas molecules. The innovation of this module lies in its elastic mechanism, namely, constructing an elastic spectral prior model based on the output of the data quality grading module. When the data quality is good, the module sets a strict constraint range, using tight constraints to filter out noise; while when the data quality is poor, the module automatically relaxes the constraint range, achieving flexible loosening of the physical prior information. This design of dynamically adjusting the constraint strength of physical prior information on the inversion process significantly improves the robustness of the algorithm under complex and variable industrial conditions, preventing the solution process from getting trapped in local optima or diverging.

[0155] In addition, the system includes a condition adaptation and model selection module. This module is designed to resolve the conflict between inversion accuracy and real-time response capability. It first calculates a condition complexity index, which comprehensively reflects the target gas type, the proportion of low-concentration components, and environmental stability. Then, based on this, the module selects a corresponding nonlinear deconvolution inversion model from a pre-set model library. The pre-set model library includes ultra-lightweight models, standard models, and full models. Furthermore, this module integrates hysteresis buffering and resource conflict compensation strategies, which can avoid frequent model switching at critical points of condition fluctuations and ensure stable system operation through forced degradation and prior compensation when edge devices are under computing power constraints. This design ensures that the system can handle extreme conditions even at edge devices with limited computing power, achieving optimized allocation of computing resources.

[0156] Finally, the system runs the dual-constraint inversion module. This module is the execution unit that generates the final monitoring data. It receives physical constraints from the elastic spectral prior construction module and source resolution constraints from the tracer gas data, forming a robust dual-constraint system. Using the selected nonlinear deconvolution inversion model, this module iteratively solves the standardized dataset. During the iteration process, this module also has anomaly detection capabilities, i.e., it verifies the authenticity of data by combining the spectral reconstruction residuals. For normal operating conditions, it obtains multi-gas concentration inversion results through rigorous mathematical iteration; while for sudden high-concentration leaks, it can activate a pass-through mechanism, skipping tedious calculations and directly outputting results, thereby achieving millisecond-level emergency response.

[0157] In summary, the various modules in this embodiment work together to not only achieve full-process automation, but also perfectly solve the problem of balancing accuracy, speed and stability in industrial gas monitoring through flexible strategy adjustments and intelligent resource management, thus significantly improving the reliability of industrial safety monitoring.

[0158] Example 5:

[0159] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0160] 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.

[0161] 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.

[0162] Bus 102 may include a pathway 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 an address bus, a data bus, a 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.

[0163] The memory 103 is used to store a computer program corresponding to the nonlinear deconvolution multi-gas concentration inversion method based on spectral prior constraints in the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 is used to execute the computer program stored in the memory 103 to implement the content shown in the foregoing method embodiments.

[0164] 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.

[0165] 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 nonlinear deconvolution method for multi-gas concentration inversion based on spectral prior constraints, characterized in that, Applications in data processing terminals include: The raw multi-source data of the target area is acquired and preprocessed to obtain a standardized dataset. The raw multi-source data includes spectral data, tracer gas concentration data and environmental parameter data. The standardized dataset is graded for data quality to obtain the data quality grading results. An elastic spectral prior model is constructed based on the data quality grading results. The elastic spectral prior model is constructed by combining physical prior information from a high-precision spectral database and dynamically adjusting the constraint strength of the physical prior information on the inversion process. The calculation of the operating condition complexity index is used to select the corresponding nonlinear deconvolution inversion model from the preset model library. The calculation of the operating condition complexity index includes weighted summation of the number of industrial target gas types, spectral overlap, proportion of low-concentration components, environmental interference level and operating condition stability level using preset weight coefficients. The step of selecting a corresponding nonlinear deconvolution inversion model from a preset model library includes: if the operating condition complexity index is less than a first complexity threshold, calling an ultra-lightweight model, which is configured to retain only the core spectral line position and intensity priors; if the operating condition complexity index is between the first and second complexity thresholds, calling a standard model, which is configured to include complete priors and simplified multi-parameter coupling; if the operating condition complexity index is greater than the second complexity threshold, calling a full model, which is configured to include complete priors and full-parameter joint inversion. Based on the dual constraints of the elastic spectral prior model and the tracer gas concentration data, the standardized dataset is iteratively solved using the nonlinear deconvolution inversion model to obtain multi-gas concentration inversion results.

2. The method according to claim 1, characterized in that, The process of constructing an elastic spectral prior model based on the data quality grading results includes: In response to the data quality grading result being a first quality level, a first constraint interval is set, wherein the first constraint interval limits the spectral line position deviation and intensity deviation to a range that is less than a preset first deviation threshold. In response to the data quality grading result being a second quality level, a second constraint interval is set. The second constraint interval restricts the spectral line position deviation and intensity deviation by a greater range than the first deviation threshold. The data quality of the second quality level is lower than that of the first quality level, so as to achieve flexible loosening of physical prior information.

3. The method according to claim 1, characterized in that, The acquisition and preprocessing of raw multi-source data of the target region includes dynamic drift compensation, the process of which includes: The baseline drift correction amount is determined based on a pre-defined environment and baseline drift correlation model. Obtain real-time dust concentration values ​​and calculate dust interference terms using preset dust correction coefficients; The corrected spectral signal is obtained by subtracting the baseline drift correction and the dust interference term from the original spectral signal.

4. The method according to claim 2, characterized in that, The process of classifying the data quality of the standardized dataset includes: Calculate the signal-to-noise ratio, spectral overlap, proportion of interference peaks, environmental interference level, and operating condition stability level of the standardized dataset; Each calculated indicator is compared with a preset first-level threshold set; When the signal-to-noise ratio is greater than the signal-to-noise ratio threshold, and the spectral overlap, interference peak ratio, environmental interference level, and operating condition stability level are all less than their respective evaluation thresholds, the data quality classification result is determined to be the first quality level. When any one of the signal-to-noise ratio, spectral line overlap, interference peak ratio, environmental interference level, or operating condition stability level meets the preset poor quality conditions, the data quality grading result is determined to be the second quality level.

5. The method according to claim 1, characterized in that, It also includes hysteresis buffering and resource conflict compensation strategies for model switching, specifically including: Calculate the smoothing operating condition complexity index, wherein the smoothing operating condition complexity index is a weighted moving average of the operating condition complexity index within a preset time window; The hysteresis model switching judgment is as follows: the model is switched to a higher complexity inversion model only when the complexity index of the smoothing condition is greater than the upgrade threshold and the duration exceeds the first judgment period; the model is switched to a lower complexity inversion model only when the complexity index of the smoothing condition is less than the downgrade threshold and the duration exceeds the second judgment period. Perform resource conflict detection and prior compensation: When it is determined that it is necessary to switch to the full model, the real-time computing resource occupancy rate of the data processing terminal is detected; if the real-time computing resource occupancy rate exceeds the preset resource occupancy threshold, the prior compensation degradation mode is activated, the standard model is forcibly called to replace the full model for inversion, and at the same time, the constraint range of the elastic spectral prior model is compressed and corrected by a preset prior compression coefficient.

6. The method according to claim 1, characterized in that, The dual constraint based on the elastic spectral prior model and the tracer gas concentration data includes: A tracer gas emission contribution rate model is constructed based on the tracer gas injection concentration of each unit and the emission contribution rate to be solved, where the total tracer concentration at the total outlet is represented as the sum of the products of the injection concentration of each unit and the corresponding emission contribution rate; The emission contribution rate of each unit is solved by constrained least squares method, and the sum of the emission contribution rates is constrained to be one, and the individual emission contribution rate is between zero and one. The target gas concentration of each unit is corrected using the emission contribution rate obtained from the solution, and the corrected concentration is used as the input constraint condition of the nonlinear deconvolution inversion model.

7. The method according to claim 1, characterized in that, The iterative solution of the standardized dataset using the nonlinear deconvolution inversion model includes performing federated learning collaborative steps: Based on the ultra-lightweight model, a local coarse inversion was performed to obtain preliminary results of gas concentration. Receive global model parameters from the core node. The global model parameters are obtained by the core node by assigning weights based on the data quality classification results of each data processing terminal and then fusing them. Perform a collaborative consistency check and calculate the deviation between the local coarse inversion result and the global optimal solution determined based on the global model parameters; When the deviation exceeds a preset consistency threshold, the data processing terminal is triggered to load the standard model and recalculate the standardized dataset.

8. The method according to claim 7, characterized in that, Before triggering the data processing terminal to load the standard model and recalculate the standardized dataset, the process also includes performing a true / false anomaly detection and pass-through step based on the spectral reconstruction residual: Using the preliminary results of the gas concentration, the spectrum was reconstructed based on the reverse generation theory of the radiative transfer physics model; Calculate the spectral reconstruction residual between the theoretically reconstructed spectrum and the spectral data in the standardized dataset, and calculate the root mean square error of the spectral reconstruction residual; If the deviation exceeds the consistency threshold, but the root mean square error is less than the preset physical confidence threshold, then the current state is determined to be a high-confidence sudden condition, and the pass-through mechanism is activated. The pass-through mechanism is configured to: prohibit the operation of loading the standard model, directly lock the preliminary gas concentration result as the final output result, and send a high confidence weight adjustment instruction to the core node.

9. The method according to claim 1, characterized in that, After obtaining the multi-gas concentration inversion results, the process also includes multi-dimensional anomaly verification and closed-loop adjustment steps: Calculate the sum of squared spectral errors of the inversion results and verify whether the inversion results exceed the preset industrial emission range and historical average range; If an anomaly is detected, a closed-loop adjustment is triggered based on the anomaly type: In response to data quality anomalies, the sensor is automatically purged and recalculated for drift compensation. In response to anomalies in prior constraints, an adjustment is triggered to the constraint range of the elastic spectral prior model; In response to abnormal operating conditions, real-time coal quality and load data are acquired to correct model parameters.

Images