A method for accurately inverting carbon dioxide, methane and nitrous oxide concentrations based on atmospheric composition background observation

Abstract

The present application relates to the technical field of atmospheric environment monitoring and data processing, and discloses a carbon dioxide, methane and nitrous oxide concentration accurate inversion method based on atmospheric composition background observation. The present application aims to solve the problem that the existing inversion method is difficult to decouple meteorological interference under the condition of multivariate coupling, and is easily affected by local emission pollution, resulting in unreliable results. The present application constructs a multivariate correlation graph by acquiring multi-source observation data, and uses an inversion reasoning model to perform spatial dynamic feature aggregation and time series dependent feature extraction. During parameter training, the network parameters are updated by fusing the gas optical physical consistency constraint and the contrast discrimination mechanism. During actual inversion, the reliability is evaluated based on independent repeated prediction, and the output parameters are periodically fine-tuned. The present application combines physical laws and data driving, decouples the cross interference of environmental properties on spectral characteristics, filters out local emission pollution characteristics, and improves the accuracy and stability of the inversion of the background concentration of multi-component greenhouse gases.

Description

Technical Field

[0001] This invention relates to the field of atmospheric environmental monitoring data processing technology, specifically, to a method for accurately retrieving the concentrations of carbon dioxide, methane, and nitrous oxide based on background atmospheric composition observations. Background Technology

[0002] Currently, field observation stations in this field typically deploy high-precision gas optical detection equipment in conjunction with environmental sensing equipment for joint observation. After acquiring multi-source observation data, existing concentration inversion techniques mainly rely on traditional single-variable physical empirical formulas or employ purely data-driven conventional machine learning models to output inversion results by establishing a mapping relationship between optical absorption signals and target gas concentrations.

[0003] However, in actual long-term field station observation operations, the atmospheric environment is in a highly complex and dynamically evolving state. The optical absorption characteristics of target gases are often deeply coupled with dynamically fluctuating meteorological factors. At the same time, long-term continuous observation sequences are inevitably affected by pollution interference from local short-term anthropogenic emissions and the systematic baseline drift caused by long-term instrument operation. Existing inversion methods, when faced with the above-mentioned complex conditions of multivariate coupling, are difficult to effectively decouple environmental interference from actual concentration changes, and lack the ability to accurately identify abnormal pollution signals. They are prone to misinterpreting transient meteorological fluctuations or local emissions as the actual evolution of background concentrations, resulting in large errors in the final inversion results and poor long-term reliability. This makes it difficult to meet the stringent requirements of climatological benchmarks for high-confidence, high-precision true atmospheric background concentration data. Summary of the Invention

[0004] Given that existing inversion methods are unable to effectively separate complex meteorological cross-interferences and localized sudden pollution fluctuations, and lack mechanisms to address long-term equipment aging and drift, this invention provides the following technical solution to address these issues: This invention provides a method for accurately inverting the concentrations of carbon dioxide, methane, and nitrous oxide based on background atmospheric composition observations, which includes the following steps: S1. Acquire multi-source observation data collected by gas optical detection equipment and environmental sensing equipment, perform time alignment and baseline calibration, extract environmental state attributes from the multi-source observation data as independent graph nodes and assign initial numerical weights to the connections, and construct a multivariate correlation graph. S2. Input the multivariate correlation map into the inversion inference model, and the inversion inference model performs dynamic feature aggregation in the spatial dimension and extraction of time-dependent features in the temporal dimension. The extracted features are then input into its internal output calculation layer to output the preliminary predicted concentration of the target greenhouse gas. S3. When training the parameters of the inversion inference model, a physical consistency constraint is constructed based on the gas optical absorption law to adjust the preliminary predicted concentration, and a comparison and discrimination mechanism is constructed to suppress the influence of local pollution characteristics on the concentration inversion result, so as to obtain the true background concentration. S4. When performing actual inversion calculations, multiple independent and repeated predictions are performed for the same set of input data. The dispersion of the multiple concentration prediction results is used as a reliability index, and the calculation parameters of the output calculation layer are periodically fine-tuned using calibration data.

[0005] As a preferred embodiment of the method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on atmospheric background observations according to the present invention, wherein: in S1, the process of acquiring multi-source observation data collected by gas optical detection equipment and environmental sensing equipment for time alignment and baseline calibration includes: calculating the numerical difference of the multi-source observation data at different sampling frequencies at each time node, finding a corresponding path with the smallest cumulative difference, synchronizing all the multi-source observation data to a unified time resolution according to the corresponding path; processing the synchronized multi-source observation data using time-series signal separation methods, separating the synchronized multi-source observation data into a low-frequency trend component characterizing systematic drift and a high-frequency signal component characterizing target concentration changes, and filtering out the low-frequency trend component to complete the baseline calibration.

[0006] As a preferred embodiment of the method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on atmospheric background observations according to the present invention, wherein: in S1, the process of extracting environmental state attributes from the multi-source observation data as independent graph nodes and assigning initial numerical weights to the connections includes: taking the spectral characteristics of the target greenhouse gas and various environmental state attributes that physically interfere with the spectral characteristics as independent graph nodes; extracting the pre-stored absorption spectral physical parameters of the target greenhouse gas; determining the degree of physical interference between each graph node based on the gas absorption spectral physical parameters; and determining the initial numerical weights of the connections between each graph node based on the degree of physical interference.

[0007] As a preferred embodiment of the method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on atmospheric background observations as described in this invention, wherein: in S2, the process of dynamic feature aggregation in the spatial dimension includes: calculating the importance score of the mutual influence between each graph node, dynamically allocating processing weights according to the magnitude of the importance score; and aggregating the information of surrounding graph nodes to the current graph node according to the processing weights to complete the feature update of the graph node in the spatial dimension.

[0008] As a preferred embodiment of the method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on atmospheric background observations as described in this invention, wherein: in S2, the process of extracting time-dependent features in the time dimension and outputting preliminary predicted concentrations includes: constructing a time-series processing structure with a historical information filtering and context state transfer mechanism; inputting the updated graph node features in the spatial dimension into the time-series processing structure in chronological order; dynamically evaluating and extracting historical cumulative features using the historical information filtering and context state transfer mechanism; inputting the extracted historical cumulative features into the output calculation layer; and having the output calculation layer calculate the preliminary predicted concentration of the target greenhouse gas at the current moment.

[0009] As a preferred embodiment of the method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on atmospheric background observations according to the present invention, wherein: in step S3, the process of constructing physical consistency constraints based on gas optical absorption laws to adjust the preliminary predicted concentration includes: defining the preliminary predicted concentration as a three-dimensional concentration vector containing the carbon dioxide, methane, and nitrous oxide; obtaining the intrinsic absorption line parameters of the carbon dioxide, methane, and nitrous oxide within independent detection windows based on a high-resolution transmission molecular absorption spectral database; calculating the dynamic absorption coefficient matrix after Doppler and collision line broadening and cross-absorption correction by combining synchronously acquired temperature, pressure, and humidity data; performing multi-component forward absorption calculations using the three-dimensional concentration vector and the dynamic absorption coefficient matrix to back-calculate the theoretical light attenuation spectrum corresponding to a specific wavenumber; calculating the difference between the theoretical light attenuation spectrum and the actual light attenuation spectrum measured by the instrument, and feeding the difference as an error constraint term back into the parameter training process to adjust the internal calculation parameters.

[0010] As a preferred embodiment of the accurate inversion method for carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations described in this invention, wherein: in step S3, the process of constructing a contrastive discrimination mechanism to suppress the influence of local pollution characteristics on the concentration inversion results and obtain the true background concentration includes: constructing a feature comparison space, acquiring historical observation data sequences, using stable data in the historical observation data sequences as positive samples, and using data with local emission characteristics as negative samples; during feature extraction, calculating the distance between the current input data features and the positive and negative samples in the feature comparison space, and suppressing the influence of local pollution characteristics on the concentration inversion results by narrowing the distance to the positive samples and widening the distance to the negative samples, and outputting the true background concentration.

[0011] As a preferred embodiment of the method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on atmospheric background observations as described in this invention, wherein: in step S3, the process of using data with local emission characteristics as negative samples includes: extracting the current concentration of each target greenhouse gas in the historical observation data sequence, and calculating the absolute value of the rate of change including both positive rapid increase and negative significant decrease states; setting corresponding fluctuation judgment thresholds based on the natural background magnitude of each target greenhouse gas and the baseline variance of the historical observation data; generating an initial anomaly candidate set when the absolute value of the rate of change of any of the target greenhouse gases is greater than its corresponding fluctuation judgment threshold, and the duration of the abnormal fluctuation meets the short-time pulse characteristic; and introducing... The historical observation data sequence is synchronized with external meteorological wind field data, pre-configured local emission sector data, or independent reference station observation sequences. A local source cross-confirmation rule is applied: if, within the time period corresponding to the initial anomaly candidate set, the synchronized environmental wind direction falls within the spatial orientation defined by the local emission sector data, the wind speed is lower than the preset retention and diffusion threshold, or the data trend exhibits spatial inconsistency with the independent reference station observation sequence, then the initial anomaly candidate set is confirmed as a local interference event. Data sequences confirmed as local interference events are extracted as negative samples, and continuous stable data sequences that do not trigger the fluctuation judgment threshold and whose wind direction is within a clean wind direction sector are extracted as positive samples, thus constructing a comparative sample set for model training and validation.

[0012] As a preferred embodiment of the method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on atmospheric background observations according to the present invention, in step S4, the process of performing multiple independent repeated predictions for the same set of input data and calculating the dispersion of the multiple concentration prediction results as a reliability index includes: for the same set of input data, dynamically and randomly selecting some calculation nodes and directly assigning their output results to zero, and performing a preset number of independent repeated prediction calculations; calculating the average value of the preset number of repeated prediction calculation results as the final output concentration, and calculating the standard deviation of the preset number of repeated prediction calculation results as the reliability index; extracting a preset baseline percentage ratio, multiplying the average value by the baseline percentage ratio, and generating a standard deviation threshold corresponding to the current input data in real time; if the standard deviation is greater than the standard deviation threshold, it is determined that feature analysis dispersion has occurred, and anomaly interception is performed on the current output concentration.

[0013] As a preferred embodiment of the accurate inversion method for carbon dioxide, methane, and nitrous oxide concentrations based on atmospheric background observations as described in this invention, wherein: in step S4, the process of periodically fine-tuning the calculation parameters of the output calculation layer using calibration data includes: stopping the backpropagation of the network weight gradient of the front feature extraction structure in the inversion inference model, keeping its parameters frozen; introducing a standard reference gas according to a preset calibration cycle, inputting the acquired calibration observation data into the inversion inference model, performing forward propagation calculation through the front feature extraction structure, extracting the corresponding hidden state feature vector and inputting it into the output calculation layer to obtain the predicted calibration concentration; obtaining the true concentration label of the standard reference gas, and based on the predicted calibration concentration and the true concentration label, calling the Huber loss function to calculate the fine-tuning loss value and solving the bottom error gradient; based on the bottom error gradient, according to a preset fine-tuning learning rate and maximum fine-tuning steps, incrementally updating only the weight matrix and bias term parameters within the output calculation layer.

[0014] The beneficial effects of this invention are as follows: By constructing multivariate correlation maps from multi-source observation data and combining dynamic feature aggregation in the spatial dimension with temporal-dependent feature extraction in the temporal dimension, this invention can characterize the correlation between different environmental state attributes and the spectral characteristics of target greenhouse gases; by introducing physical consistency constraints and contrast discrimination mechanisms based on the optical absorption law of gases, the impact of environmental state changes and local emission interference on the concentration inversion results is reduced; at the same time, the reliability of the inversion results is assessed by evaluating the dispersion of multiple independent repeated prediction results, and the calculation parameters of the output calculation layer are periodically fine-tuned using calibration data, thereby reducing the impact of instrument state changes on the inversion results during long-term operation, thus improving the accuracy and long-term operational stability of the background concentration inversion results of carbon dioxide, methane, and nitrous oxide. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a method for accurately inverting the concentrations of carbon dioxide, methane, and nitrous oxide based on background atmospheric composition observations.

[0017] Figure 2 A flowchart for preprocessing multi-source data and constructing multivariate association maps.

[0018] Figure 3 This is a flowchart of spatiotemporal dual-dimensional feature extraction and preliminary concentration prediction.

[0019] Figure 4 A visual diagram illustrating data sampling and baseline calibration.

[0020] Figure 5 This is a schematic diagram of local interference identification and background output. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0024] Example 1 Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for accurately inverting the concentrations of carbon dioxide, methane, and nitrous oxide based on background atmospheric composition observations, comprising the following steps: S1. Acquire multi-source observation data collected by gas optical detection equipment and environmental sensing equipment, perform time alignment and baseline calibration, extract environmental state attributes from the multi-source observation data as independent graph nodes and assign initial numerical weights to the connections, and construct a multivariate correlation graph. The numerical differences of the multi-source observation data at different sampling frequencies at each time node are calculated, and a corresponding path with the smallest cumulative difference is found. Based on the corresponding path, all the multi-source observation data are synchronized to a unified time resolution. The synchronized multi-source observation data are processed using time-series signal separation methods to separate the synchronized multi-source observation data into a low-frequency trend component characterizing systematic drift and a high-frequency signal component characterizing target concentration changes. The low-frequency trend component is filtered out to complete baseline calibration.

[0025] The spectral characteristics of the target greenhouse gas and the environmental state attributes that physically interfere with the spectral characteristics are each used as an independent graph node; the physical parameters of the absorption spectral lines of the target greenhouse gas are extracted from the pre-stored data, the degree of physical interference between each graph node is determined based on the physical parameters of the gas absorption spectral lines, and the initial numerical weight of the connection between each graph node is determined based on the degree of physical interference.

[0026] In actual field background observations, the sampling frequency of meteorological sensors is typically lower than that of high-precision optical analysis instruments. This inconsistency in the pace of multi-source heterogeneous data acquisition hinders subsequent joint computation. Faced with multi-source heterogeneous data streams, the system first aggregates the data sequences generated by various sensors and constructs a two-dimensional comparison grid within the program. The system distributes the observation sequences with faster sampling on the horizontal axis and the observation sequences with slower sampling on the vertical axis. Next, it calculates the numerical difference between each time data point on the horizontal axis and each time data point on the vertical axis, filling these values ​​reflecting local differences into the grid to form an error matrix. Subsequently, the system starts from the beginning of the grid and progresses towards the end. During this traversal, the system restricts the matching direction to forward, downward, or diagonally, representing one-to-many, many-to-one, or one-to-one mapping relationships between data points, respectively.

[0027] While traversing and calculating all potential connecting routes, the system continuously accumulates the local difference values ​​along the way. By comparing the total accumulated values ​​of all routes, the system selects the route with the smallest total accumulated time error as the optimal mapping path. Along this path, the system smoothly resamples all data and synchronously aligns it to a standard time scale once per second.

[0028] After time alignment of multi-source data, processing is performed to address changes in the physical characteristics of the underlying hardware. When precision optical equipment operates continuously for extended periods, aging of the internal light source or minute deformation of the cavity can cause a long-term, low-frequency trend shift in the measurement baseline. To address this systematic error, the system employs time-series signal separation techniques such as multi-scale decomposition to process the synchronized data waveform. In practice, the system first internally sets multiple preset time observation windows, ranging from wide to narrow.

[0029] The large-scale time-domain observation window is limited to between ten and thirty minutes. This range is chosen because the thermal balance of the optical instrument cavity and the aging drift of the light source typically exhibit a low-frequency, gradual evolution cycle on the order of minutes. Meanwhile, the small-scale time-domain observation window is limited to between one and five seconds. This range is chosen because fluctuations in the concentration of real gases in nature, driven by atmospheric turbulence, typically undergo rapid, short-term changes within this short period. The system performs a segmented sliding scan of these observation windows at different scales along the data sequence synchronized by multi-level decoupling and signal separation transformation.

[0030] The large-scale time-domain observation window automatically spans and smoothly filters out localized short-term high-frequency fluctuations, extracting low-frequency trend features that evolve gradually over long periods in the data sequence—that is, low-frequency trend components representing long-term drift in the equipment hardware. The aforementioned small-scale time-domain observation window is specifically used to separate and extract high-frequency pulse features occurring within a short period, extracting high-frequency signal components representing rapid fluctuations in the true atmospheric concentration in the natural environment. The system then directly filters out the separated low-frequency trend signals, completing baseline calibration without compromising the true concentration fluctuation characteristics.

[0031] Furthermore, multiple variables in the external meteorological environment can interfere with the absorption spectrum of the target gas. In actual measurements, instruments calculate concentration based on the characteristics of gas absorption light. However, fluctuations in temperature or changes in air pressure can cause the characteristic spectral lines of the gas absorption spectrum to broaden or deform. Since these environmental state attributes are highly coupled with the spectral characteristics of the gas in the original observation data sequence, a data dimension decoupling operation is performed to address this coupling feature.

[0032] In practical implementation, the system independently extracts the spectral characteristic data, temperature data, air pressure data, and humidity data of the target greenhouse gas, maps them to different computing units, and defines each independent computing unit as a graph node. After completing the node definition, to quantify the physical coupling relationship between various environmental state attributes and gas spectral characteristics, connecting edges are constructed between the above graph nodes.

[0033] The specific construction process is as follows: The system constructs an association mapping matrix internally, mapping all defined graph nodes to the row and column indices of this matrix. Subsequently, the system traverses each row and column intersection of the matrix, performing matching operations on each intersection according to a preset physical association rule table. If two graph nodes corresponding to a certain intersection meet the preset association conditions, the system assigns a connectivity attribute to that intersection to generate a valid data transmission channel. The above-mentioned process of assigning connectivity attributes completes the construction of connecting edges at the data structure level.

[0034] After the connection edges are constructed, the system calls a publicly available high-resolution transmission molecular absorption spectrum database, which records the physical parameters of gas absorption lines for various gases. The system extracts the pre-recorded physical parameters of the gas absorption lines and performs correlation calculations on the corresponding physical parameters of the gas absorption lines based on the influence of the physical quantities corresponding to different graph nodes on the absorption spectra of the target greenhouse gases. This yields correlation values ​​characterizing the degree of physical interference between the graph nodes. Subsequently, these correlation values ​​are dimensionless, and the processed correlation values ​​are used as the initial numerical weights for the connections between the graph nodes.

[0035] The magnitude of the initial numerical weights is positively correlated with the degree of physical interference between corresponding graph nodes, and is used to characterize the relative strength of the physical interference between corresponding nodes. By decoupling multidimensional attributes into independent nodes and determining the degree of physical interference between nodes based on the physical parameters of gas absorption spectra at the corresponding intersections of the correlation matrix, the initial numerical weights of the connection edges are assigned. The system transforms the multivariate coupled atmospheric environmental physical relationships into a network topology structure with initial numerical weights, i.e., a multivariate correlation graph, thereby providing a standardized input data structure for the subsequent inversion inference model calculations.

[0036] S2, the multivariate correlation map is input into the inversion inference model, the inversion inference model performs dynamic feature aggregation in the spatial dimension and extraction of time-dependent features in the temporal dimension, and the extracted features are input into its internal output calculation layer to output the preliminary predicted concentration of the target greenhouse gas; Calculate the importance score of the mutual influence between each graph node, and dynamically allocate processing weights according to the magnitude of the importance score; aggregate the information of surrounding graph nodes to the current graph node according to the processing weight, and complete the graph node feature update in the spatial dimension.

[0037] A time-series processing structure with historical information filtering and context state passing mechanisms is constructed. The updated graph node features in the spatial dimension are input into the time-series processing structure in chronological order. The historical information filtering and context state passing mechanisms are used to dynamically evaluate and extract historical cumulative features. The extracted historical cumulative features are input into the output calculation layer, and the output calculation layer calculates the preliminary predicted concentration of the target greenhouse gas at the current moment.

[0038] Before performing dynamic feature aggregation and temporal feature extraction, the system first internally constructs and initializes the inversion inference model. Specifically, the construction process involves the system sequentially building, at the code level, a spatial mapping network layer for processing multivariate graph features, a temporal processing structure for processing time-series features, and an output computation layer at the end of the network, according to the data processing flow. Initialized random weight parameters are then assigned to the node matrices in each of these computation layers. After the inversion inference model is constructed, the system inputs the data stream containing the multivariate correlation graph into the inversion inference model.

[0039] The atmospheric environment is in a constant state of dynamic evolution, and the environmental factors that dominate the spectral characteristics of the target gas are constantly changing at different times. In order to dynamically capture this characteristic, the spatial mapping network layer configured inside the system processes multivariate correlation graphs in real time and calculates the importance scores of the mutual influence of each node at the current specific moment.

[0040] The specific process for calculating importance scores is as follows: First, the system extracts the current feature data of the target greenhouse gas map node and the surrounding environmental state map nodes. Next, it combines the feature data of the target map node with the feature data of a certain surrounding map node in pairs, and uses the system's built-in parameter matrix to project the combined data, obtaining a raw value reflecting the initial correlation strength between the two. Then, the system performs nonlinear activation processing on this raw value to obtain the corresponding feature response value. Subsequently, the system aggregates the raw values ​​calculated for the target map node and all surrounding map nodes, converts all raw values ​​to positive scales through exponential mapping, and calculates the sum of these positive scales. Finally, the system divides each positive scale by this sum, thereby converting and allocating these raw values ​​into a set of relative weight proportions between zero and one, summing to 100%.

[0041] The relative weight ratio represents the importance score of each environmental node in influencing the target gas at a specific time. When the characteristic value of a certain meteorological attribute changes due to external environmental influences, the importance score corresponding to that environmental state map node is dynamically adjusted based on its current characteristic data after the above splicing and conversion. The system dynamically allocates processing weights according to the importance scores of each map node, and performs weighted aggregation of the feature information of the surrounding environmental state map nodes according to the processing weights to complete the feature update of the target greenhouse gas map node.

[0042] The underlying mathematical logic for the dynamic feature aggregation process that calculates the importance scores of the mutual influence between graph nodes is as follows: First, the system uses the feature projection matrix to perform combination and nonlinear activation calculations on the target graph node and its surrounding nodes to obtain the initial association strength. : In the formula, The target graph node represents the current feature vector of the target greenhouse gas. A feature vector representing a node in a state diagram of a certain surrounding environment; This is the system's built-in feature mapping weight matrix; This represents the concatenation operation of feature vectors; This is the projection parameter matrix; The modified linear unit activation function with leakage is used to amplify significant feature signals while preserving weak negative interference features. Subsequently, the system performs exponential mapping and sum-normalization operations on the initial correlation strengths to obtain the relative weight ratios. : In the formula, Characterizes the initial association strength between target graph nodes and surrounding specific environmental state graph nodes; Representative set Each node in the environment graph that is traversed; Representation of target graph nodes and sets The initial association strength between nodes in the environment graph; This represents the set of all environment graph nodes connected to the target graph node. This normalization operation ensures that the sum of the interference weights of all input environment nodes on the target gas is 100%. Finally, the updated target features in the spatial dimension are the features of the surrounding nodes based on this weight. The weighted summation.

[0043] Furthermore, atmospheric flow and mass diffusion are physical processes with time lags; the air mass transport state in the previous stage affects the subsequent concentration trend. Based on this, the system operates a second-level time-series processing structure. Specifically, the system constructs a long-range state memory channel spanning all time points within this structure to carry historically accumulated contextual information along the timeline. Simultaneously, at each specific time point, a gated operation matrix constructed earlier is configured to perform numerical control on the historical information filtering and contextual state transmission mechanism. The data sequence, updated after spatial feature aggregation, is sequentially input into the time-series processing structure. When the data features of the current moment enter, the system concatenates them with the short-term state features transmitted from the previous moment. Next, the system calculates a forgetting ratio between zero and one based on the forgetting gate matrix. This forgetting ratio is then used to perform a multiplicative decay operation on the old information in the long-range state memory channel, thereby terminating the transmission of redundant environmental state data to subsequent nodes.

[0044] Subsequently, the system again uses the updated gating matrix to calculate an update ratio, extracts the latest feature vectors representing gas accumulation and diffusion from the input features at the current moment, and fuses these new features into the long-range state memory channel through an addition operation. Finally, based on the fused latest long-range state memory, the system outputs the current accumulated state to the next time node. Through this time-node-by-time state filtering, truncation, and new data fusion operation, the system sends these time-dependent accumulated features to the third-level output calculation layer. The output calculation layer performs matrix mapping operations to calculate the preliminary predicted concentration of the target greenhouse gas at the current moment.

[0045] Regarding the historical information filtering and context state fusion process in the aforementioned time-series processing structure, the underlying computational logic based on the gating operation matrix is ​​as follows: the forgetting ratio, i.e., the forgetting gating... The calculation process is as follows: Update ratio, i.e., input gating With candidate new feature vectors The calculation process is as follows: Accumulated state in long-range state memory channel The fusion update formula is: In the formula, The data characteristics input to the time-series processing structure at the current moment; The short-term state characteristics passed from the previous moment; , , as well as , , Both are mapping matrices and bias term parameters that can be used for gradient updates; Characterizing the hyperbolic tangent activation function; It is an S-type logic activation function that can restrict the output value to between zero and one to achieve filtering and truncation. This represents the element-wise multiplication operation of a matrix. Represents the cumulative state at the previous moment. Through The multiplicative decay and discarding of old information have been completed, through The additive fusion of new features has been completed.

[0046] S3, when training the parameters of the inversion inference model, a physical consistency constraint is constructed based on the gas optical absorption law to adjust the preliminary predicted concentration, and a comparison and discrimination mechanism is constructed to suppress the influence of local pollution characteristics on the concentration inversion result, so as to obtain the true background concentration. The preliminary predicted concentration is defined as a three-dimensional concentration vector containing the carbon dioxide, methane, and nitrous oxide. Based on a high-resolution transmission molecular absorption spectroscopy database, the intrinsic absorption line parameters of the carbon dioxide, methane, and nitrous oxide within independent detection windows are obtained. Combining synchronously acquired temperature, pressure, and humidity data, a dynamic absorption coefficient matrix is ​​calculated after Doppler and collision line broadening and cross-absorption correction. Multi-component forward absorption calculations are performed using the three-dimensional concentration vector and the dynamic absorption coefficient matrix to back-calculate the theoretical light attenuation spectrum corresponding to a specific wavenumber. The difference between the theoretical light attenuation spectrum and the actual light attenuation spectrum measured by the instrument is calculated, and this difference is fed back as an error constraint to the parameter training process to adjust the internal calculation parameters.

[0047] A feature comparison space is constructed, and historical observation data sequences are obtained. Stable data in the historical observation data sequences are used as positive samples, and data with local emission characteristics are used as negative samples. When performing feature extraction, the distance between the current input data features and the positive and negative samples in the feature comparison space is calculated. By narrowing the distance with the positive samples and widening the distance with the negative samples, the influence of local pollution features on the concentration inversion results is suppressed, and the true background concentration is output.

[0048] The current concentration of each target greenhouse gas is extracted from the historical observation data sequence, and the absolute value of the rate of change, including both positive rapid increase and negative significant decrease states, is calculated. Based on the natural background level of each target greenhouse gas and the baseline variance of the historical observation data, corresponding fluctuation judgment thresholds are set. When the absolute value of the rate of change of any of the target greenhouse gases exceeds its corresponding fluctuation judgment threshold, and the duration of the abnormal fluctuation meets the short-time pulse characteristic, an initial anomaly candidate set is generated. External meteorological wind field data synchronized with the historical observation data sequence, pre-configured local emission sector data, or independent data are introduced. For the reference station observation sequence, the local source cross-confirmation rule is executed: if, within the time period corresponding to the initial anomaly candidate set, the synchronous environmental wind direction falls within the spatial orientation defined by the local emission sector data, the wind speed is lower than the preset retention and diffusion threshold, or the data trend of the independent reference station observation sequence shows spatial inconsistency characteristics, then the initial anomaly candidate set is confirmed as a local interference event; the data sequence confirmed as the local interference event is extracted as a negative sample, and the continuous stable data sequence that has not triggered the fluctuation judgment threshold and whose wind direction is in the clean wind direction sector is extracted as a positive sample, thus constructing a comparison sample set for model training and validation.

[0049] If only a data-driven approach is used for parameter training, the model output may not conform to the laws of gas optical absorption. Since there are multiple greenhouse gases in the atmosphere, and their absorption lines overlap in certain bands, a simple single-variable Beer-Lambert law cannot achieve physical closure. Therefore, this embodiment constructs a multi-component coupled forward physical absorption calculation model.

[0050] The specific reverse alignment calculation process is as follows: First, the output of the inversion inference model is configured as a three-dimensional concentration vector. Secondly, the system calls a standard molecular spectral database to obtain the intrinsic absorption line parameters of the aforementioned gases within the corresponding independent detection windows of the optical instrument. These intrinsic absorption line parameters are the aforementioned physical parameters of the gas absorption lines, specifically including the center wavelength, reference line intensity, air broadening half-width at half-maximum (FWHM), and low-energy state energies of each line.

[0051] Next, the system combines the currently synchronously collected environmental status node data, specifically the absolute temperature. Ambient air pressure and the water vapor mixing ratio, which characterizes humidity Temperature correction based on the partition function is applied to the standard reference line intensity, and the spectral deformation, including Doppler broadening and collisional broadening, is calculated using the Vogt line shape function. By linearly superimposing the absorption contributions of each gas component at the same wavelength, a dynamic absorption coefficient matrix including the effects of cross-absorption is calculated. .

[0052] Subsequently, the system calculates a specific wavenumber based on the multi-component forward absorption equation. Theoretical light attenuation value : In the formula, The effective physical folding optical path for light to penetrate the gas inside an optical instrument; For the first The dynamic absorption coefficient of the gas after temperature, pressure and humidity correction; This is the preliminary predicted concentration of the corresponding gas output by the model.

[0053] Simultaneously, the system retrieves the actual spectral scan response of the optical instrument and calculates the actual light attenuation spectrum within the measured wavelength band. In the formula, This is the initial light source intensity data; This represents the final captured light source intensity data. The final generated physical error constraint term. The calculation formula has been updated to calculate the root mean square error for discrete scan wavenumber points: In the formula, These are the physical constraint weighting coefficients; This represents the squared Euclidean distance used to calculate the difference between the two. The introduction of this multi-component physical forward model ensures that the inversion inference model, when performing gradient backpropagation, is constrained by physical laws to simultaneously satisfy the physical closure condition for cross-absorption of the three gases under temperature, pressure, and humidity dynamic broadening, thereby effectively suppressing spurious concentration vectors in the model output that do not conform to the spectral overlap law.

[0054] On the other hand, in actual field station observations, relying solely on the rate of concentration increase and duration to automatically identify local emissions lacks reliable true data and is prone to misjudging real long-distance air mass transport or short-term pulse fluctuations caused by instrument valve switching as pollution. To construct a sample set for comparative learning, the system established an automatic labeling mechanism based on cross-verification of multi-component differential thresholds and external physical evidence.

[0055] The specific sample selection and cross-validation rules are as follows: First, given that the natural background concentration of carbon dioxide is in the ppm range, while that of methane and nitrous oxide is in the ppb range, the system uses absolute change rate thresholds set based on the baseline variance of the historical observation data for each of the three gases. Specifically, the system sets corresponding fluctuation thresholds for carbon dioxide in the range of 0.5~2.0 ppm / s, for methane in the range of 5~20 ppb / s, and for nitrous oxide in the range of 1~5 ppb / s.

[0056] Secondly, the system calculates the absolute value of the rate of change of gas concentration in the historical sequence. This calculation simultaneously identifies positive rapid rise states and negative significant decline states; whereby the positive rapid rise state represents local emissions of gas from fuel vehicles or human activities passing through the vicinity of the station, and the negative significant decline state represents hardware system interference signals caused by instrument calibration gas switching leaks or short-term purging of the intake pipeline. When the absolute value of the rate of change of any of carbon dioxide, methane, or nitrous oxide is greater than the corresponding fluctuation judgment threshold, and the duration of the abnormal fluctuation is within a short pulse time window of 3 to 15 seconds, the system extracts the data sequence of the corresponding time period to construct an initial abnormal candidate set.

[0057] Subsequently, the system incorporates time-synchronized external meteorological wind field data and pre-configured local emission sector data for cross-verification. This local emission sector data is used to characterize the spatial location of anthropogenic emission sources around the observation station. The system retrieves real-time wind direction and speed data from the anemometer tower. If the time period corresponding to the initial anomaly candidate set meets one of the following criteria, the initial anomaly candidate set is confirmed as a local disturbance event: First, the wind speed is lower than a preset retention and diffusion threshold, which is set within the range of 0.5~1.5 m / s, indicating local calm wind accumulation of the target gas; Second, the ambient wind direction falls within the spatial orientation defined by the local emission sector data; Third, the system calls the synchronous observation sequence of an independent reference station within a preset spatial distance range for comparison. The preset spatial distance range is set to 10 to 50 kilometers. If the target station shows an independent high-value fluctuation signal while the independent reference station does not show a corresponding signal, it indicates that the data exhibits inconsistent characteristics in spatial distribution.

[0058] After cross-verification, the system marks data sequences verified as local interference events as negative samples. Simultaneously, the system extracts continuous stationary data sequences with wind direction within clean wind sectors and absolute rate of change below the aforementioned fluctuation threshold as positive samples. The system allocates the positive and negative samples to a training sample set and a validation sample set according to a preset ratio. This preset ratio represents the ratio of the data volume in the training sample set to the validation sample set, and its specific value is set within the range of 7:3 to 9:1. The system synchronously inputs the allocated samples into the high-dimensional feature comparison space.

[0059] In the feature extraction computation stage, the system quantifies the vector distance between the current input feature and the two classes of samples in a high-dimensional feature comparison space. This high-dimensional feature comparison space is essentially a set of multi-dimensional mathematical matrices constructed by the hidden layer nodes of the inversion inference model. Within this spatial structure, different types of features in the original observation sequence are transformed into multiple preset number of feature dimensions through mapping matrices, so that data with different attributes form corresponding spatial distance differences in the feature comparison space.

[0060] The specific implementation and calculation process is as follows: First, the system uses the mapping calculation matrix inside the inversion inference model to synchronously map the current input data, positive samples, and negative samples to the high-dimensional feature comparison space and extract their corresponding feature vectors. Then, the system performs Euclidean vector distance calculation to calculate the first spatial distance between the current input feature vector and the feature vector of the positive sample, and the second spatial distance between the current input feature vector and the feature vector of the negative sample. Next, the system constructs a relative distance evaluation function internally, which sets an optimization constraint objective that the first spatial distance should be less than the second spatial distance, and the two should maintain a pre-set safety boundary difference.

[0061] The safety boundary difference here is set as a Euclidean distance scalar between 0.1 and 0.5. This scalar range is chosen as the basis for boundary constraints because: if the preset difference is less than 0.1, the distance constraint between pollution features and background features decreases, which may reduce the distinguishability between the two types of features; if the preset difference is greater than 0.5, the inversion inference model will fail to converge.

[0062] If the features extracted by the current inversion inference model do not meet the optimization constraint objective, the system will generate a feature rejection error term based on the difference between the two distances mentioned above. The system feeds this feature rejection error term back to the inversion inference model, continuously adjusting its internal weights. As the weights are updated, the feature mapping method of the inversion inference model for the current input data changes, causing a shift in its mapping coordinates in the comparison space. This achieves feature isolation operations by reducing the spatial distance threshold between the model and positive samples while simultaneously expanding the spatial distance threshold between the model and negative samples in the geometric dimension of the data. Through this distance reconstruction mechanism in the feature space, the inversion inference model effectively suppresses the influence of short-term pollution characteristics caused by local emission sources on the concentration inversion results, outputting the true background concentration representing the regional macro-climate evolution trend.

[0063] Regarding the calculation process of feature space alignment and generation of feature exclusion error terms, the specific calculation logic executed by the distance evaluation function is as follows: in, This is the feature rejection error term generated; This represents the first spatial distance between the current input feature vector and the feature vector of the positive sample. ; The second spatial distance represents the distance between the current input feature vector and the feature vector of the negative sample. ; Represents the current input feature vector. This represents the feature vector of a positive sample. Represents the feature vector of the negative sample; This represents a scalar value indicating the predefined safety boundary difference. The function's operation mechanism is as follows: the second spatial distance... Failed to exceed the first spatial distance Difference from safety boundary When the sum is equal to the sum, the function will calculate an error term greater than zero, thereby triggering the weight iteration of the inversion inference model and performing feature push-pull and isolation actions.

[0064] S4. During the actual inversion calculation, multiple independent and repeated predictions are performed for the same set of input data. The dispersion of the multiple concentration prediction results is used as a reliability index, and the calculation parameters of the output calculation layer are periodically fine-tuned using calibration data.

[0065] For the same set of input data, some computing nodes are dynamically and randomly selected, and their output results are directly assigned to zero. A preset number of independent repeated prediction calculations are then performed. The average value of the preset number of repeated prediction calculation results is calculated as the final output concentration, and the standard deviation of the preset number of repeated prediction calculation results is calculated as the reliability index. A preset baseline percentage is extracted, and the average value is multiplied by the baseline percentage to generate a standard deviation threshold corresponding to the current input data in real time. If the standard deviation is greater than the standard deviation threshold, feature parsing discrepancies are determined, and anomaly interception is performed on the current output concentration.

[0066] Stop the backpropagation of the network weight gradient of the front feature extraction structure in the inversion inference model, and keep its parameters frozen; introduce standard reference gas according to a preset calibration period, input the acquired calibration observation data into the inversion inference model, perform forward propagation calculation through the front feature extraction structure, extract the corresponding hidden state feature vector and input it into the output calculation layer to obtain the predicted calibration concentration; obtain the true concentration label of the standard reference gas, and calculate the fine-tuning loss value based on the predicted calibration concentration and the true concentration label, and solve the bottom error gradient; based on the bottom error gradient, according to the preset fine-tuning learning rate and the maximum fine-tuning steps, only update the weight matrix and bias term parameters inside the output calculation layer incrementally.

[0067] When the inversion system is running online for an extended period and is fed meteorological data whose feature distribution exceeds the boundaries of the original weighted optimized sample, the inversion inference model risks outputting low-confidence error values. To quantitatively assess the reliability of operational inversion results, an uncertainty calculation mechanism is introduced into the inference operation phase of the system.

[0068] For the same batch of multimodal input data, the system dynamically and randomly selects some computational nodes within the inversion inference model according to a preset probability and directly assigns their output to zero, simulating the non-fully connected operation state of the inversion inference model's internal structure. The preset probability is limited to between 10% and 30%, based on the following: if the probability of setting a node to zero is less than 10%, the output of the inversion inference model is highly homogeneous and consistent, making it impossible to effectively assess its uncertainty under the current environmental characteristics; if the probability of setting a node to zero is greater than 30%, it will disrupt the integrity of the front-end feature extraction structure, causing feature mapping degradation even under normal weather conditions and outputting predicted values ​​that deviate from the reasonable physical range, thereby reducing the effectiveness of the reliability assessment mechanism.

[0069] In the non-fully connected mode where some nodes are set to zero, the system performs a preset number of independent repeated prediction calculations on the current data. The preset number of repeated predictions is limited to between fifty and one hundred. Subsequently, the system aggregates the multiple concentration predictions generated by these independent repeated predictions into a numerical sequence, calculates the average of this sequence as the final output inversion concentration, and calculates the standard deviation of these repeated prediction results. Because this standard deviation maintains the same physical dimensions as the mean, it can intuitively reflect the analytical dispersion of each computational layer within the inversion inference model for the current input data.

[0070] The underlying calculation logic for the statistical evaluation process of the repeated prediction calculation and its standard deviation reliability index for the above-mentioned preset number of times is as follows: Final output concentration The calculation formula is: The standard deviation of prediction reflects the reliability index. The calculation formula is: In the formula, This represents the number of repeated predictions performed in a non-fully connected operation state, and its value is... ; Characterizing the first The concentration prediction value is calculated independently and repeatedly.

[0071] The aforementioned baseline percentage for anomaly detection was determined during the early testing and validation phase of the inversion inference model by statistically analyzing the prediction dispersion of historical normal observation data in the non-fully connected mode. In practical deployments, this baseline percentage is specifically quantified to be between 0.5% and 2.0%. During each actual inversion calculation, the system dynamically extracts the aforementioned average predicted concentration. This value is then multiplied by the baseline percentage to generate the standard deviation threshold for the current input data in real time. Since the standard deviation and the concentration mean have the same physical dimensions, the threshold generation and condition determination process is logically executable at the computer level.

[0072] The core physical basis for selecting this specific proportion range as the anomaly detection threshold lies in the high standards and regulations imposed by the World Meteorological Organization on the long-term accuracy of background greenhouse gas observations. If the standard deviation of the output of the aforementioned preset number of predictions exceeds the standard deviation threshold generated in real time, it is determined that the various computational layers within the inversion inference model have produced significant feature analysis discretization of the current anomalous meteorological input data, and its potential output error significantly exceeds the reasonable error range specified by the climatological benchmark. In this case, the system triggers a risk interception command for this inversion prediction, directly terminating the output and storage process of this batch of anomalous data to prevent low-quality data from being mixed into the final atmospheric background core database.

[0073] Finally, with the wear and tear of optical analysis equipment over long-term use, and the gradual shift in data feature distribution caused by the annual increase in global atmospheric background concentration, the inversion inference model is prone to gradually accumulating a systematic trend bias with a single direction over time. If global parameter updates are frequently performed on the inversion inference model, recent calibration data may change the front feature extraction parameters formed based on historical training data, thereby affecting the original feature extraction results. Therefore, this embodiment adopts a local parameter incremental fine-tuning strategy based on gradient freezing.

[0074] In practice, the system has a pre-defined calibration cycle for periodically performing parameter calibration, with the cycle duration ranging from thirty to ninety days. When the system reaches the set point of the calibration cycle, the optical instrument's measurement channel switches to purging a standard reference gas. This standard reference gas is a pre-prepared mixed standard gas containing a constant, known concentration of the target greenhouse gas, having undergone a metrological verification procedure. The system acquires the observation data generated by the optical instrument during the purging of the standard reference gas and uses it as calibration observation data. Simultaneously, the system extracts the inherent baseline concentration value of the standard reference gas and uses it as the true concentration label for fine-tuning calculations.

[0075] During parameter fine-tuning calculations, the system first defines the hierarchical structure of the inversion inference model, collectively referring to the spatial mapping network layer responsible for processing multivariate correlation graphs and the temporal processing structure as the front-end feature extraction structure. Subsequently, the system modifies the update permissions of the underlying computation graph, truncating the backpropagation path of the network weight gradients for the front-end feature extraction structure. This truncation operation prevents the weight matrix within the front-end feature extraction structure from participating in gradient updates during fine-tuning and keeps it frozen. The system then inputs the aforementioned calibration observation data into the inversion inference model. The frozen front-end feature extraction structure, utilizing its internally locked weight matrix, sequentially performs feature projection, weighted aggregation, and state fusion calculations based on the gating matrix on the calibration observation data. This process of mapping data according to a unidirectional data flow is the forward propagation calculation. After this calculation, the front-end feature extraction structure converts the underlying calibration observation data into a high-dimensional hidden-state feature vector and transmits this hidden-state feature vector to the output computation layer, which calculates and outputs the corresponding predicted calibration concentration.

[0076] To address the small reference offset caused by long-period drift, the system employs the Huber loss function as the fine-tuning loss function. Its calculation logic is as follows: In the formula, The output calculation layer outputs the predicted calibration concentration based on the calibration observation data; The actual concentration label for the standard reference gas; This is a constant threshold for smooth conversion, used to prevent gradient value divergence caused by abnormal calibration data fluctuations.

[0077] The system calculates the underlying error gradient based on this fine-tuning loss function. During parameter updates, the system uses this error gradient to incrementally update only the weight matrix and bias parameters within the output computation layer. To avoid excessively large single weight updates that could cause the model to fail to converge and disrupt existing underlying feature mapping logic, the system sets the fine-tuning learning rate to a range of 1×10⁻⁶. -5 Up to 5×10 -5 It also limits the maximum number of fine-tuning steps to between one hundred and five hundred steps.

[0078] By combining the aforementioned forward feature extraction and gradient truncation computational mechanism, while keeping the structural parameters of the forward feature extraction frozen, only the output computation layer parameters are incrementally updated. This reduces the impact of systematic optical drift caused by long-term continuous operation of the instrument on the concentration inversion results, and is used for long-term continuous inversion of background concentrations of carbon dioxide, methane, and nitrous oxide.

[0079] Example 2 Reference Figures 1-4 This is the second embodiment of the present invention. To further illustrate the invention, the technical solution is described using an example of a high-altitude national atmospheric background field observation station operating under meteorological conditions where the characteristic distribution exceeds the sample boundary and under local emission interference conditions. The station is equipped with a multi-channel optical cavity ring-down spectrometer to collect high-frequency greenhouse gas spectral attenuation signals, and simultaneously equipped with an automatic weather station to collect relatively low-frequency environmental state signals. The currently set target gases for inversion are carbon dioxide, methane, and nitrous oxide, each with corresponding typical atmospheric background concentration levels. Specifically, in the macroscopic background of nature, carbon dioxide is at approximately 420 ppm, methane at approximately 1900 ppb, and nitrous oxide at approximately 330 ppb.

[0080] S1. Acquire multi-source observation data collected by gas optical detection equipment and environmental sensing equipment, perform time alignment and baseline calibration, extract environmental state attributes from multi-source observation data as independent graph nodes and assign initial numerical weights to the connections, and construct a multivariate correlation graph.

[0081] In daily operations, the system continuously acquires the aforementioned multi-source heterogeneous observation data and performs local difference matrix accumulation operations in a two-dimensional comparison grid to select the optimal mapping path. It then uniformly and smoothly resamples both types of data to a standard time scale once per second. Addressing the optical lens aging phenomenon that occurs after several months of continuous operation of the cavity ring-down spectrometer, the system sets a large-scale time-domain observation window with a width ranging from ten to thirty minutes. This window slides along the data sequence to extract and filter out low-frequency trend components, while retaining high-frequency signal components to complete baseline calibration. Subsequently, the system decouples the spectral data of the target gas from data such as temperature, pressure, and humidity into independent graph nodes. It then extracts pre-stored absorption spectral physical parameters of the target greenhouse gas from the built-in database. Based on these parameters, it determines the degree of physical interference between graph nodes, performs dimensionless processing on the correlation values ​​characterizing this degree of physical interference, and uses the processed correlation values ​​as the initial numerical weights for the corresponding connection edges, thereby generating a multivariate correlation graph.

[0082] S2 performs spatial dynamic feature aggregation and temporal dependency feature extraction on the multivariate correlation map, and inputs the extracted features into the output computation layer to output the preliminary predicted concentration of the target greenhouse gas.

[0083] When the station is under localized strong convection conditions and the characteristic values ​​of the air pressure parameters change, the spatial mapping network layer inside the inversion inference model performs feature projection and importance score calculation. The tensor values ​​at the bottom layer of the system at the current moment are extracted: the original values ​​representing the initial correlation strength of the air pressure node, temperature node, and humidity node are denoted as follows: , and Because the characteristic values ​​of the current air pressure parameters change, the corresponding input feature vectors undergo numerical changes after matrix concatenation. These changes are then processed through feature projection matrix operations and nonlinear activation functions within the system to calculate the final value. The value is greater than and The system then substitutes the above values ​​and performs a uniform normalization mapping operation: Based on the monotonically increasing mathematical property of exponential functions, since The pressure node has the highest numerical value, resulting in the highest importance score among the three environmental state map nodes, while the weighting of temperature and humidity decreases accordingly. Based on this dynamic weight distribution, the system assigns a higher processing weight to the pressure node than to the temperature and humidity nodes, and aggregates the feature information corresponding to the pressure node to the target greenhouse gas map node based on this weight. Subsequently, the data stream enters the second-level time-series processing structure. Since atmospheric transport and diffusion processes are time-dependent, the system performs historical state fusion calculations based on its internal matrix configuration. The operating parameters of the time-series processing structure at this moment are extracted: the accumulated state transmitted through the long-range state memory channel from the previous moment. The system calculates the forgetting ratio based on the forgetting gating. Based on the input ratio calculated by the updated gating And candidate new feature vectors extracted from the latest meteorological data by the current inversion inference model. The system performs cumulative state fusion calculations in the long-range state memory channel based on the fusion update formula: The historical cumulative features obtained through the above cumulative state fusion calculation are input into the output calculation layer. The output calculation layer calculates the preliminary predicted concentration of the target gas at the current moment to preserve the time-dependent characteristics of the continuous change of gas concentration over time.

[0084] S3. When iteratively optimizing network parameters, a physical consistency constraint is constructed based on the optical absorption law of gas to adjust the preliminary predicted concentration, and a comparison and discrimination mechanism is constructed to suppress the influence of local pollution characteristics on the concentration inversion results, so as to obtain the true background concentration.

[0085] During the network parameter training and iteration phase, the system synchronously incorporates the aforementioned preliminary predicted concentrations into the forward physical absorption calculation model. The system configures the preliminary predicted concentrations as a three-dimensional concentration vector containing the three gases and calls upon a high-resolution transmission molecular absorption spectroscopy database to obtain the intrinsic absorption line parameters within independent detection windows near 1.6 μm for carbon dioxide, 1.65 μm for methane, and 4.5 μm for nitrous oxide. Subsequently, the system combines temperature, air pressure, and humidity data synchronously collected from an automatic weather station and uses the Vogt line shape function to calculate a broadened and corrected dynamic absorption coefficient matrix. The system uses this dynamic absorption coefficient matrix and the three-dimensional concentration vector to perform multi-component forward absorption calculations, inversely deducing the theoretical light attenuation spectrum corresponding to a specific wavenumber. Simultaneously, the system retrieves the light attenuation spectrum measured by the instrument, calculates the difference between the two to generate physical error constraint terms, and constrains the output of the inversion inference model to satisfy the physical laws of multi-component spectral overlap.

[0086] During the observation period, a diesel-powered transport vehicle passed near the air intake of the observation station, causing an increase in the concentrations of carbon dioxide and nitrous oxide in the observation sequence. The system scanned the observation data sequence and calculated the absolute values ​​of the rate of change of the concentration of each target greenhouse gas. The calculated absolute value of the rate of change of carbon dioxide was 1.5 ppm / s, and the simultaneous absolute value of the rate of change of nitrous oxide was 3.0 ppb / s, exceeding the corresponding fluctuation judgment thresholds of 1.0 ppm / s and 2.0 ppb / s set in this verification scenario, respectively. Furthermore, the duration of the abnormal fluctuations fell within a short pulse time window of three to fifteen seconds, triggering the initial anomaly candidate set truncation condition. Subsequently, the system retrieved the wind tower data and detected that the ambient wind direction at this time fell within the spatial orientation defined by the pre-configured local emission sector data, triggering the cross-confirmation rule, confirming it as a local interference event, and marking it as a negative sample. In the distance evaluation of the feature comparison space, the safety boundary difference scalar was set to be within the range of 0.1 to 0.5. The first spatial distance between the input feature corresponding to the current local interference event and the positive sample was extracted. and the second spatial distance from the reverse sample The system substitutes the relative distance evaluation function to perform the calculation: If the feature rejection error term is greater than zero, the system feeds it back to the inversion inference model. Based on this feature rejection error term, the system adjusts the weights of the internal feature mapping matrix, expanding the second spatial distance between the current feature and the negative sample in the data geometric dimension. This suppresses the influence of local emission features on the concentration inversion results and outputs the true background concentration representing the climate characteristics of high-altitude areas.

[0087] S4, during actual inversion calculations, performs multiple independent and repeated predictions for the same set of input data. The dispersion of the calculated concentration predictions is used as a reliability index, and the calculation parameters of the output calculation layer are periodically fine-tuned using calibration data. When inputting meteorological conditions with low coverage, such as snowfall, in the training samples, the input data features may exceed the original weight optimization sample boundaries.

[0088] In non-fully connected mode, the system dynamically and randomly selects a portion of computational nodes with a preset node-to-zero probability of 10% to 30%, assigns their output values ​​to zero, and performs a preset number of independent repeated prediction calculations. The system aggregates the concentration prediction values ​​obtained from each independent repeated prediction, calculates their average value, and generates the corresponding standard deviation threshold in real time. The system calculates the standard deviation of the independent repeated prediction results and compares the standard deviation with the real-time generated standard deviation threshold; when the standard deviation is greater than the standard deviation threshold, it determines that the repeated prediction result corresponding to the current input data has feature parsing discrepancies, and the system performs an anomaly interception operation, terminating the output process of this batch of data.

[0089] During routine station maintenance, when the system runtime reaches a set node in the calibration cycle, the measurement channel of the optical instrument automatically switches to receive a mixed standard gas to acquire calibration observation data and corresponding baseline concentration labels. The system truncates the backpropagation path of the network weight gradient of the front-end feature extraction structure, freezing it. The calibration observation data is extracted into hidden-state feature vectors via forward propagation and input into the output computation layer. The system calls the Huber loss function to calculate the fine-tuning loss value and solve for the underlying error gradient, and incrementally updates only the parameters within the output computation layer based on the set fine-tuning learning rate. By freezing the front-end feature extraction structure and only incrementally updating the parameters of the output computation layer, the calibration data is used to adjust the output mapping relationship, thereby reducing the impact of systematic optical drift caused by long-term operation on the concentration inversion results.

[0090] To verify the multi-component greenhouse gas inversion results of this embodiment, the system extracted observation data from the station for twelve consecutive months and compared the background true concentration output by this method with the measurement data of an independently set and periodically calibrated reference instrument.

[0091] In the specific parameter configuration of this verification experiment, the fluctuation judgment thresholds for each gas were fixed as follows: carbon dioxide 1.0 ppm / s, methane 10 ppb / s, and nitrous oxide 2 ppb / s, respectively; the short-time pulse time window was set to 5 seconds; the node zeroing probability in the reliability assessment was set to 20%, the number of independent repeated predictions was set to 100, and the baseline percentage corresponding to the standard deviation threshold was set to 1.0%; the calibration period for fine-tuning the output computation layer was set to 30 days, and the fine-tuning learning rate was set to 2 × 10⁻⁶. -5The maximum number of fine-tuning steps is set to two hundred.

[0092] Meanwhile, during the verification calculation, the system shielded the aforementioned physical consistency constraints and contrast discrimination mechanisms, using the preliminary predicted concentration sequence directly output by the inversion inference model without incorporating these two mechanisms for training and optimization as the benchmark data for error comparison. Verification results show that the complete inversion mechanism of this invention reduced the root mean square error (RMSE) of carbon dioxide inversion from 1.2 ppm to 0.15 ppm, with the mean absolute deviation controlled within ±0.05 ppm; the RMS error of methane decreased from 15.0 ppb to 1.2 ppb; and the RMS error of nitrous oxide decreased from 3.5 ppb to 0.3 ppb. The error comparison results covering a continuous twelve-month observation period demonstrate that after introducing physical consistency constraints and contrast discrimination mechanisms, the RMS errors of carbon dioxide, methane, and nitrous oxide inversion are all lower than the corresponding benchmark data, indicating that the inversion method can be used for concentration inversion at the ppm level for carbon dioxide and at the ppb level for methane and nitrous oxide.

[0093] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for accurately inverting the concentrations of carbon dioxide, methane, and nitrous oxide based on background atmospheric composition observations, characterized in that, include: S1. Acquire multi-source observation data collected by gas optical detection equipment and environmental sensing equipment, perform time alignment and baseline calibration, extract environmental state attributes from the multi-source observation data as independent graph nodes and assign initial numerical weights to the connections, and construct a multivariate correlation graph. S2, the multivariate correlation map is input into the inversion inference model, the inversion inference model performs dynamic feature aggregation in the spatial dimension and extraction of time-dependent features in the temporal dimension, and the extracted features are input into its internal output calculation layer to output the preliminary predicted concentration of the target greenhouse gas; S3, when training the parameters of the inversion inference model, a physical consistency constraint is constructed based on the gas optical absorption law to adjust the preliminary predicted concentration, and a comparison and discrimination mechanism is constructed to suppress the influence of local pollution characteristics on the concentration inversion result, so as to obtain the true background concentration. S4. During the actual inversion calculation, multiple independent and repeated predictions are performed for the same set of input data. The dispersion of the multiple concentration prediction results is used as a reliability index, and the calculation parameters of the output calculation layer are periodically fine-tuned using calibration data.

2. The method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations according to claim 1, characterized in that, In step S1, the process of acquiring multi-source observation data collected by gas optical detection equipment and environmental sensing equipment for time alignment and baseline calibration includes: The numerical differences of the multi-source observation data at different sampling frequencies at each time node are calculated, and a corresponding path with the smallest cumulative difference is found. Based on the corresponding path, all the multi-source observation data are synchronized to a unified time resolution. The synchronized multi-source observation data are processed using time-series signal separation methods to separate the synchronized multi-source observation data into a low-frequency trend component characterizing systematic drift and a high-frequency signal component characterizing target concentration changes. The low-frequency trend component is filtered out to complete baseline calibration.

3. The method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations according to claim 1, characterized in that, In step S1, the process of extracting environmental state attributes from the multi-source observation data as independent graph nodes and assigning initial numerical weights to the connections includes: The spectral characteristics of the target greenhouse gas and the environmental state attributes that physically interfere with the spectral characteristics are each used as an independent graph node; the physical parameters of the absorption spectral lines of the target greenhouse gas are extracted from the pre-stored data, the degree of physical interference between each graph node is determined based on the physical parameters of the gas absorption spectral lines, and the initial numerical weight of the connection between each graph node is determined based on the degree of physical interference.

4. The method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations according to claim 1, characterized in that, In S2, the process of dynamically aggregating spatial features includes: Calculate the importance score of the mutual influence between each graph node, and dynamically allocate processing weights according to the magnitude of the importance score; aggregate the information of surrounding graph nodes to the current graph node according to the processing weight, and complete the graph node feature update in the spatial dimension.

5. The method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations according to claim 4, characterized in that, In step S2, the process of extracting time-dependent features and outputting preliminary predicted concentrations includes: A time-series processing structure with historical information filtering and context state passing mechanisms is constructed. The updated graph node features in the spatial dimension are input into the time-series processing structure in chronological order. The historical information filtering and context state passing mechanisms are used to dynamically evaluate and extract historical cumulative features. The extracted historical cumulative features are input into the output calculation layer, and the output calculation layer calculates the preliminary predicted concentration of the target greenhouse gas at the current moment.

6. The method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations according to claim 1, characterized in that, In step S3, the process of constructing physical consistency constraints based on the gas optical absorption law to adjust the preliminary predicted concentration includes: The preliminary predicted concentration is defined as a three-dimensional concentration vector containing the carbon dioxide, methane, and nitrous oxide. Based on a high-resolution transmission molecular absorption spectroscopy database, the intrinsic absorption line parameters of the carbon dioxide, methane, and nitrous oxide within independent detection windows are obtained. Combining synchronously acquired temperature, pressure, and humidity data, a dynamic absorption coefficient matrix is ​​calculated after Doppler and collision line broadening and cross-absorption correction. Multi-component forward absorption calculations are performed using the three-dimensional concentration vector and the dynamic absorption coefficient matrix to back-calculate the theoretical light attenuation spectrum corresponding to a specific wavenumber. The difference between the theoretical light attenuation spectrum and the actual light attenuation spectrum measured by the instrument is calculated, and this difference is fed back as an error constraint to the parameter training process to adjust the internal calculation parameters.

7. The method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations according to claim 1, characterized in that, In step S3, the process of constructing a contrastive discrimination mechanism to suppress the influence of local pollution characteristics on the concentration inversion results and obtain the true background concentration includes: A feature comparison space is constructed, and historical observation data sequences are obtained. Stable data in the historical observation data sequences are used as positive samples, and data with local emission characteristics are used as negative samples. When performing feature extraction, the distance between the current input data features and the positive and negative samples in the feature comparison space is calculated. By narrowing the distance with the positive samples and widening the distance with the negative samples, the influence of local pollution features on the concentration inversion results is suppressed, and the true background concentration is output.

8. The method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations according to claim 7, characterized in that, In step S3, the process of using data with local emission characteristics as negative samples includes: The current concentration of each target greenhouse gas is extracted from the historical observation data sequence, and the absolute value of the rate of change, including both positive rapid increase and negative significant decrease states, is calculated. Based on the natural background level of each target greenhouse gas and the baseline variance of the historical observation data, corresponding fluctuation judgment thresholds are set. When the absolute value of the rate of change of any of the target greenhouse gases exceeds its corresponding fluctuation judgment threshold, and the duration of the abnormal fluctuation meets the short-time pulse characteristic, an initial anomaly candidate set is generated. External meteorological wind field data synchronized with the historical observation data sequence, pre-configured local emission sector data, or independent data are introduced. For the reference station observation sequence, the local source cross-confirmation rule is executed: if, within the time period corresponding to the initial anomaly candidate set, the synchronous environmental wind direction falls within the spatial orientation defined by the local emission sector data, the wind speed is lower than the preset retention and diffusion threshold, or the data trend of the independent reference station observation sequence shows spatial inconsistency characteristics, then the initial anomaly candidate set is confirmed as a local interference event; the data sequence confirmed as the local interference event is extracted as a negative sample, and the continuous stable data sequence that has not triggered the fluctuation judgment threshold and whose wind direction is in the clean wind direction sector is extracted as a positive sample, thus constructing a comparison sample set for model training and validation.

9. The method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations according to claim 1, characterized in that, In step S4, the process of performing multiple independent and repeated predictions on the same set of input data and calculating the dispersion of the resulting concentration predictions as a reliability index includes: For the same set of input data, some computing nodes are dynamically and randomly selected, and their output results are directly assigned to zero. A preset number of independent repeated prediction calculations are then performed. The average value of the preset number of repeated prediction calculation results is calculated as the final output concentration, and the standard deviation of the preset number of repeated prediction calculation results is calculated as the reliability index. A preset baseline percentage is extracted, and the average value is multiplied by the baseline percentage to generate a standard deviation threshold corresponding to the current input data in real time. If the standard deviation is greater than the standard deviation threshold, feature parsing discrepancies are determined, and anomaly interception is performed on the current output concentration.

10. The method for accurate inversion of carbon dioxide, methane, and nitrous oxide concentrations based on background atmospheric composition observations according to claim 1, characterized in that, In step S4, the process of periodically fine-tuning the calculation parameters of the output calculation layer using calibration data includes: Stop the backpropagation of the network weight gradient of the front feature extraction structure in the inversion inference model, and keep its parameters frozen; introduce standard reference gas according to a preset calibration period, input the acquired calibration observation data into the inversion inference model, perform forward propagation calculation through the front feature extraction structure, extract the corresponding hidden state feature vector and input it into the output calculation layer to obtain the predicted calibration concentration; obtain the true concentration label of the standard reference gas, and calculate the fine-tuning loss value based on the predicted calibration concentration and the true concentration label, and solve the bottom error gradient; based on the bottom error gradient, according to the preset fine-tuning learning rate and the maximum fine-tuning steps, only update the weight matrix and bias term parameters inside the output calculation layer incrementally.

Images