A passive infrared gas analysis inversion method and system based on a convolution decoupling architecture and a background parameterization model

Abstract

The application relates to the technical field of spectral data processing, and discloses a passive infrared gas analysis inversion method and system based on a convolution decoupling architecture and a background parameterization model. The method comprises the following steps: obtaining a standard radiation brightness original observation tensor; extracting radiation distribution characteristics of an environmental background through a multi-scale space background estimation network; inputting the original observation tensor and the baseline background characteristic vector into a preset convolution decoupling architecture; performing bidirectional orthogonal constraint processing on a background decoupling branch and a gas feature extraction branch in the convolution decoupling architecture to separate a pure target gas spectrum feature map; and inputting the pure target gas spectrum feature map into a preset concentration inversion mapping network to obtain a column concentration distribution matrix of a target gas in a target monitoring area. Through the technical solution, the passive infrared gas signal and the complex environmental background are accurately separated, and the accuracy of target gas column concentration inversion is improved.

Description

Technical Field

[0001] This application relates to the field of spectral data processing, and in particular to a passive infrared gas analytical inversion method and system based on a convolutional decoupling architecture and a background parameterization model. Background Technology

[0002] Passive infrared gas remote sensing technology, with its core advantages of no active light source, long detection range, and wide coverage, has become a core technology for industrial leak monitoring, environmental pollution assessment, and safety control in chemical industrial parks. Its core principle is to detect the thermal radiation spectrum of the target area, identify the characteristic absorption and emission signals of gas molecules in specific infrared bands, and thus achieve gas type identification and quantitative concentration retrieval.

[0003] However, in complex industrial environments, passive infrared gas analysis and inversion faces a core technical bottleneck: deep coupling between dynamic background radiation and weak gas characteristics, leading to signal entanglement. In passively detected radiation energy, background thermal radiation accounts for a very high proportion, with only a small amount carrying gas characteristic information. When the temperature and emissivity of heat sources such as reactors, cooling towers, and chimneys in industrial settings undergo spatiotemporal dynamic non-uniform changes, the complex background radiation baseline will non-linearly couple with the weak gas absorption characteristics.

[0004] Existing technologies mostly adopt a paradigm of overall processing of mixed spectra, which lacks the ability to decouple features at the mechanistic level and cannot effectively separate high-frequency fluctuating background noise from low-frequency weak gas features. This results in a very high false alarm rate and serious distortion of concentration inversion results in scenarios with low thermal contrast and complex and fluctuating background environment. At the same time, traditional methods do not consider the influence of meteorological parameters and temporal background changes, which further reduces the accuracy and real-time performance of inversion.

[0005] Therefore, how to achieve efficient decoupling between dynamic background and gas characteristics, and improve the accuracy, robustness and real-time performance of passive infrared gas analysis and inversion, has become an urgent technical problem to be solved. Summary of the Invention

[0006] This application provides a passive infrared gas analytical inversion method based on a convolutional decoupling architecture and a background parameterization model, so as to at least partially solve the above-mentioned technical problems.

[0007] To achieve the above objectives, according to a first aspect of this application, a passive infrared gas analytical inversion method based on a convolutional decoupling architecture and a background parameterization model is provided, comprising: Step S101: Obtain the passive infrared hyperspectral data cube of the target monitoring area, and perform radiometric calibration and spectral alignment on the passive infrared hyperspectral data cube to obtain the original observation tensor of standard radiance. Step S102: Input the original observation tensor into the preset background parameterization model, extract the radiation distribution features of the environmental background through the multi-scale spatial background estimation network, and output the corresponding baseline background feature vector. Step S103: Input the original observation tensor and the baseline background feature vector into a preset convolutional decoupling architecture, and use the background decoupling branch and the gas feature extraction branch in the convolutional decoupling architecture to perform bidirectional orthogonal constraint processing to separate the pure target gas spectral feature map. Step S104: Input the spectral feature map of the pure target gas into a preset concentration inversion mapping network to obtain the column concentration distribution matrix of the target gas in the target monitoring area.

[0008] Preferably, the preset background parameterization model includes a time sliding window sampling module, a local contrast calculation unit, and a multi-scale spatial background estimation network connected in sequence; step S102 specifically includes: The reference frame sequence is extracted from the raw observation tensor of continuous time sequence using the time sliding window sampling module. The local contrast calculation unit is used to calculate the pixel-level radiometric difference between the current original observation tensor and the reference frame sequence to generate a background variation map; The current original observation tensor is merged with the background variation map and then input into the multi-scale spatial background estimation network to extract the radiation distribution features of the environmental background and generate the baseline background feature vector.

[0009] Preferably, the internal operating logic of the convolutional decoupling architecture specifically includes: The background decoupling branch receives the original observation tensor and the baseline background feature vector, and extracts the spectral deformation features of the background continuum through a three-dimensional convolutional layer; The gas feature extraction branch receives the original observation tensor and extracts the fine absorption peak features of gas molecules through a one-dimensional spectral hole convolution layer. The cosine similarity between the feature tensor output by the background decoupling branch and the feature tensor output by the gas feature extraction branch is calculated, and the network parameters are updated based on the objective of minimizing the cosine similarity to achieve the bidirectional orthogonal constraint processing.

[0010] Preferably, the gas feature extraction branch also includes a spectral spatial attention fusion mechanism, and the specific steps for separating the pure target gas spectral feature map include: The spectral attention module is used to redistribute the channel-dimensional weights of the features output by the one-dimensional spectral hole convolutional layer, thereby enhancing the response weights of specific gas absorption bands. The spatial attention module is used to perform spatial dimension masking on the features after weight redistribution to suppress background interference in non-gas distribution areas. The features processed by the spectral attention module and the spatial attention module are fused by element-wise multiplication to generate the pure target gas spectral feature map.

[0011] Preferably, the preset concentration inversion mapping network is a fully connected regression residual network, and the specific steps for obtaining the column concentration distribution matrix of the target gas within the target monitoring area include: Flatten the spectral feature map of the pure target gas into a one-dimensional feature vector; The one-dimensional feature vector is input into a multilayer fully connected regression residual network, and the corresponding column concentration value is output through nonlinear mapping layer by layer by a multilayer perceptron. The column concentration values ​​are reconstructed into a two-dimensional matrix based on the original spatial coordinate information to obtain the column concentration distribution matrix.

[0012] Preferably, the method further includes a training step for the preset convolutional decoupling architecture: Construct a training dataset containing mixed radiation spectrum samples, known background radiation spectrum samples, and known gas concentration labels; The mixed radiation spectrum samples are input into the convolutional decoupling architecture to be trained; The output of the background decoupling branch is constrained by known background radiation spectrum samples, the output of the concentration inversion mapping network is constrained by known gas concentration labels, and a joint loss function is constructed by combining a cosine similarity penalty term. The weight parameters of the convolutional decoupling architecture are updated using the backpropagation algorithm based on the joint loss function until the model converges.

[0013] Preferably, the time-sliding window sampling module employs a dynamic update strategy based on structural similarity when extracting the reference frame sequence: Calculate the structural similarity index between the historical frames and the current original observation tensor within the sliding window; Remove historical frames whose structural similarity index is below a preset threshold; The retained historical frames are weighted and averaged according to the time decay weight to generate a reference frame sequence for characterizing the transient background.

[0014] Preferably, before feature extraction, the convolutional decoupling architecture further includes an initial gas mask generation step based on a dynamic threshold: Calculate the spectral difference between the original observation tensor and the reference band in the main absorption band of the gas; An adaptive floating threshold is determined based on the distribution of spectral differences in the historical gas-free state of the target monitoring area. The calculated spectral difference values ​​are binarized using the adaptive floating threshold to generate an initial gas mask that marks the potential gas presence region. The initial gas mask is input as a prior attention map into the gas feature extraction branch.

[0015] Preferably, the background decoupling branch introduces a cross-scale spatial feature aggregation mechanism when extracting the spectral deformation features of the background continuum: The baseline background feature vector is processed in parallel using multiple two-dimensional dilated convolutions with different dilation rates to obtain spatial texture features under different receptive fields; Spatial texture features from different receptive fields are spliced ​​together along the channel dimension; By using point-to-point convolutional layers to reduce the dimensionality of the stitched features and interact with cross-channel information, spectral deformation features of a background continuum with a global receptive field are generated.

[0016] According to a second aspect of this application, a passive infrared gas analytical inversion system based on a convolutional decoupling architecture and a background parameterization model is provided, comprising: The data preprocessing module is used to acquire the passive infrared hyperspectral data cube of the target monitoring area, and to perform radiometric calibration and spectral alignment on the passive infrared hyperspectral data cube to obtain the original observation tensor of standard radiance. The background parameterization module is used to input the original observation tensor into a preset background parameterization model, extract the radiation distribution features of the environmental background through a multi-scale spatial background estimation network, and output the corresponding baseline background feature vector. The convolutional decoupling module is used to input the original observation tensor and the baseline background feature vector into a preset convolutional decoupling architecture, and to perform bidirectional orthogonal constraint processing on the background decoupling branch and the gas feature extraction branch in the convolutional decoupling architecture to separate the pure target gas spectral feature map. The concentration inversion module is used to input the spectral feature map of the pure target gas into a preset concentration inversion mapping network to obtain the column concentration distribution matrix of the target gas in the target monitoring area; The model training module is used to build a training dataset and update the weight parameters of the convolutional decoupling architecture through a joint loss function and backpropagation algorithm. The meteorological embedding module is used to acquire synchronous meteorological state parameters and map them into meteorological feature vectors, which are then embedded into the fully connected layer of the concentration inversion mapping network to correct the inversion results. The mask generation module is used to calculate the spectral difference value and generate an initial gas mask, which is then used as a prior attention map input to the gas feature extraction branch.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a dynamic baseline reference through a background parameterization model and combines it with the bidirectional orthogonal constraints of a convolutional decoupling architecture to forcibly separate the background continuum from gas absorption features at the algorithmic level. This completely solves the problem of entanglement between dynamic background and weak gas signals in complex industrial environments, significantly reducing the system's false alarm rate and making gas detection under weak thermal contrast possible. 2. By focusing on the gas characteristic bands and potential regions through the spectral spatial attention fusion mechanism and combining the strong nonlinear fitting capability of the fully connected regression residual network, the problem of missed detection of trace gas signals by traditional methods is effectively solved. At the same time, the concentration underestimation at spectral saturation is avoided, and high-precision quantitative inversion of gas column concentration is achieved.

[0018] 3. By introducing time-series dynamic background updates, cross-scale feature aggregation, and meteorological parameter embedding mechanisms, the system can adaptively cope with the effects of sudden changes in industrial heat sources, multi-scale background textures, and extreme weather conditions without relying on a fixed reference blackbody, significantly improving long-term operational stability and all-weather monitoring capabilities in complex scenarios.

[0019] 4. By pre-screening with an initial gas mask, the convolution decoupling operation is focused on the suspected leak area, which greatly reduces the load pressure on edge computing hardware and improves the real-time processing frame rate. It can be directly deployed on edge servers in industrial sites to meet the real-time monitoring needs of scenarios such as chemical industrial parks.

[0020] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the steps of a passive infrared gas analytical inversion method based on a convolutional decoupling architecture and a background parameterization model provided in an exemplary embodiment of this application. Figure 2 This is a schematic diagram of a passive infrared gas analytical inversion system based on a convolutional decoupling architecture and a background parameterization model provided in an exemplary embodiment of this application; Figure labeling: 201, Data preprocessing module; 202, Background parameterization module; 203, Convolution decoupling module; 204, Concentration inversion module; 205, Model training module; 206, Meteorological embedding module; 207, Mask generation module. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the protection scope of this application.

[0024] This application provides a passive infrared gas analytical inversion method based on a convolutional decoupling architecture and a background parameterization model. Please refer to [link to relevant documentation]. Figure 1 The passive infrared gas analytical inversion method based on a convolutional decoupling architecture and a background parameterization model provided in this application includes the following steps: Step S101: Obtain the passive infrared hyperspectral data cube of the target monitoring area, and perform radiometric calibration and spectral alignment on the passive infrared hyperspectral data cube to obtain the original observation tensor of standard radiance. Step S102: Input the original observation tensor into the preset background parameterization model, extract the radiation distribution features of the environmental background through the multi-scale spatial background estimation network, and output the corresponding baseline background feature vector. Step S103: Input the original observation tensor and the baseline background feature vector into a preset convolutional decoupling architecture, and use the background decoupling branch and the gas feature extraction branch in the convolutional decoupling architecture to perform bidirectional orthogonal constraint processing to separate the pure target gas spectral feature map. Step S104: Input the spectral feature map of the pure target gas into a preset concentration inversion mapping network to obtain the column concentration distribution matrix of the target gas in the target monitoring area.

[0025] The above scheme synchronously collects passive infrared hyperspectral data of the target monitoring area. After radiometric calibration and spectral alignment, a standardized original observation tensor is constructed, providing a high-quality data foundation for subsequent analytical inversion. Dynamic background radiation features are accurately extracted using a background parameterization model to generate a stable baseline background reference, solving the problem of background radiation fluctuations in complex environments. Through the bidirectional orthogonal constraint of the convolutional decoupling architecture, the mechanism of background and gas features is decoupled, removing background interference from weak gas signals. Based on the concentration inversion mapping network, the accurate mapping of spectral features to column concentrations is completed, outputting the spatial distribution information of the target gas. This effectively solves the problems of high false alarm rate and low inversion accuracy caused by the entanglement of background and gas signals in traditional methods, significantly improving the accuracy, robustness, and real-time performance of passive infrared gas analytical inversion, and adapting to the application needs of complex scenarios such as industrial leak monitoring and environmental pollution assessment.

[0026] In some embodiments, acquiring a passive infrared hyperspectral data cube of the target monitoring area and performing radiometric calibration and spectral alignment on the passive infrared hyperspectral data cube to obtain the original observation tensor of standard radiance includes: By deploying a passive infrared hyperspectral imager in the target monitoring area, the broadband infrared radiation signal of the target area is continuously captured. After being split by an interferometer or dispersive element, the optical signal is converted into an electrical signal by an infrared focal plane array detector, forming a three-dimensional data structure containing two spatial dimensions and one spectral dimension, namely the passive infrared hyperspectral data cube. Based on the blackbody calibration coefficient matrix preset by the passive infrared hyperspectral imager, the electrical signal response values ​​of each channel in the passive infrared hyperspectral data cube are subjected to linear or nonlinear radiometric calibration processing, and converted into absolute radiance values ​​with clear physical dimensions. A spectral axis interpolation algorithm is used to align and correct the spectral wavelength shift caused by hardware dispersion nonlinearity, ensuring that the center wavelength of each pixel in the same channel is absolutely consistent. By integrating the radiometrically calibrated and spectrally aligned data, a standard radiance raw observation tensor with standard format, numerical stability, and dimensionless bias is generated. The data dimension of this raw observation tensor is typically width × height × number of bands, which fully preserves the spatial, spectral, and radiance information of the target area.

[0027] Specifically, the passive infrared hyperspectral imager's spectral coverage must match the characteristic absorption band of the target gas, and the sampling frequency must be no less than 50Hz to ensure the capture of dynamic changes in gas concentration. The infrared focal plane array detector has a pixel resolution of no less than 640×512, enabling precise capture of spatial details in the target area. During radiometric calibration, high-temperature and low-temperature dual blackbody calibration is used, with calibration accuracy controlled within ±2% to ensure the accuracy of radiance values. The spectral axis interpolation algorithm uses cubic spline interpolation with an interpolation accuracy of 0.1nm, effectively compensating for gaps in discrete sampling and avoiding feature extraction distortion caused by spectral alignment deviations. The original observation tensor is stored in HDF5 format, facilitating subsequent data reading and feature extraction operations while preserving the spatial coordinates and spectral dimension correlation of the data.

[0028] In some embodiments, the preset background parameterization model includes a time sliding window sampling module, a local contrast calculation unit, and a multi-scale spatial background estimation network connected in sequence; step S102 specifically includes: The reference frame sequence is extracted from the raw observation tensor of continuous time sequence using the time sliding window sampling module. The local contrast calculation unit is used to calculate the pixel-level radiometric difference between the current original observation tensor and the reference frame sequence to generate a background variation map; The current original observation tensor is merged with the background variation map and then input into the multi-scale spatial background estimation network to extract the radiation distribution features of the environmental background and generate the baseline background feature vector.

[0029] Specifically, the time-sliding window sampling module maintains a fixed-length time-sliding window in memory, preferably 5-10 frames long. It stores multiple raw observation tensors continuously acquired at historical moments as an initial reference frame sequence and optimizes the reference frame sequence using a dynamic update strategy based on structural similarity to ensure the effectiveness of the reference frames. The local contrast calculation unit calculates the absolute radiometric difference between the current raw observation tensor and each frame tensor in the reference frame sequence in each band, using pixels as the unit. It takes the average of all reference frame differences as the final radiometric difference of the pixel, counts the radiometric differences of all pixels, and generates a background variation map reflecting the degree of drastic change in spatial radiometrics. The dimension of the background variation map is consistent with the spatial dimension of the raw observation tensor. The larger the pixel value, the more drastic the change in background radiometrics in that area.

[0030] The current original observation tensor and background variation map are merged along the channel dimension to obtain input data that integrates spatial, spectral, and background variation information. This data is then input into a multi-scale spatial background estimation network. This network is a deep neural network pre-trained on a large-scale outdoor gas-free background dataset. It contains multiple three-dimensional convolutional layers with different dilation rates to extract radiation features from large-area smooth backgrounds and high-frequency textured backgrounds, respectively. The network maps and fuses the multi-scale background features through fully connected layers to generate a low-dimensional baseline background feature vector. This vector eliminates weak high-frequency fluctuations in the spectral dimension and retains the low-frequency continuum component that determines the overall thermal radiation distribution of the background. The preferred dimension is 256, which can be directly used for subsequent convolutional decoupling processing.

[0031] In some embodiments, the internal operating logic of the convolutional decoupling architecture specifically includes: The background decoupling branch receives the original observation tensor and the baseline background feature vector, and extracts the spectral deformation features of the background continuum through a three-dimensional convolutional layer; The gas feature extraction branch receives the original observation tensor and extracts the fine absorption peak features of gas molecules through a one-dimensional spectral hole convolution layer. The cosine similarity between the feature tensor output by the background decoupling branch and the feature tensor output by the gas feature extraction branch is calculated, and the network parameters are updated based on the objective of minimizing the cosine similarity to achieve the bidirectional orthogonal constraint processing.

[0032] Specifically, the convolutional decoupling architecture consists of a feature input layer, a background decoupling branch, a gas feature extraction branch, an orthogonal constraint layer, and a feature output layer. It is deployed on the core accelerated graphics card of the edge computing server to ensure real-time computing performance. The background decoupling branch uses a 3×3×3 three-dimensional convolutional kernel that slides synchronously in the spatial and spectral dimensions to capture the slowly changing background continuum correlation between adjacent pixels and adjacent bands, extracting the spectral deformation features of the background continuum. The feature tensor dimension is consistent with the spatial dimension of the original observation tensor. The gas feature extraction branch uses a one-dimensional spectral hole convolutional layer with an expansion rate of 2-4 that slides along the spectral dimension to expand the spectral receptive field without increasing the computational parameters. This accurately captures the fine absorption peak features of gas molecules at specific wavelengths, avoiding the omission of spectral absorption peaks.

[0033] The orthogonal constraint layer calculates the cosine similarity between the background feature tensor output by the background decoupling branch and the gas feature tensor output by the gas feature extraction branch in the multidimensional feature space. The weight parameters of the convolutional decoupling architecture are updated in real time through the backpropagation algorithm, forcing the background features and gas features to achieve orthogonal separation, ensuring that the gas feature tensor does not contain background radiation interference, and providing pure feature support for subsequent concentration inversion.

[0034] In some embodiments, the gas feature extraction branch is further provided with a spectral spatial attention fusion mechanism, and the specific steps for separating the pure target gas spectral feature map include: The spectral attention module is used to redistribute the channel-dimensional weights of the features output by the one-dimensional spectral hole convolutional layer, thereby enhancing the response weights of specific gas absorption bands. The spatial attention module is used to perform spatial dimension masking on the features after weight redistribution to suppress background interference in non-gas distribution areas. The features processed by the spectral attention module and the spatial attention module are fused by element-wise multiplication to generate the pure target gas spectral feature map.

[0035] Specifically, the spectral attention module compresses the feature map output by the one-dimensional spectral hole convolutional layer into channel-level feature vectors through global average pooling. Then, it calculates the weight coefficient of each spectral channel through two layers of multilayer perceptron, assigning high weights (approaching 1) to gas feature bands and low weights to redundant bands without gas absorption, thus achieving weight redistribution of spectral channels. The spatial attention module extracts spatial features through two-dimensional global average pooling and global max pooling. After dimensionality reduction by convolutional layers, it generates a spatial weight mask. The mask has high weight values ​​for regions with potential gas presence and high weight values ​​for non-gas regions, thus achieving background interference suppression in the spatial dimension. The non-gas regions include background devices and the sky.

[0036] The spectral attention-weighted feature tensor is fused with the spatial attention mask through element-wise multiplication, integrating the dual effects of spectral enhancement and spatial focusing. This generates a pure target gas spectral feature map with the same spatial dimension as the original observation tensor. This feature map retains only the absorption peak features of the target gas, completely eliminating background radiation and irrelevant interference, and significantly improving the accuracy of subsequent concentration inversion.

[0037] In some embodiments, the preset concentration inversion mapping network is a fully connected regression residual network, and the specific steps for obtaining the column concentration distribution matrix of the target gas within the target monitoring area include: Flatten the spectral feature map of the pure target gas into a one-dimensional feature vector; The one-dimensional feature vector is input into a multilayer fully connected regression residual network, and the corresponding column concentration value is output through nonlinear mapping layer by layer by a multilayer perceptron. The column concentration values ​​are reconstructed into a two-dimensional matrix based on the original spatial coordinate information to obtain the column concentration distribution matrix.

[0038] Specifically, the pure target gas spectral feature map is a two-dimensional spatial feature map (width × height), which is flattened and reshaped into a one-dimensional feature vector with a length of width × height × number of feature channels; the fully connected regression residual network contains 3-5 hidden layers, with the number of neurons in each hidden layer being 1024, 512, and 256 respectively, and adopts the ReLU activation function. The network internally sets up a residual structure with skip connections to effectively avoid the gradient vanishing problem caused by deep networks and improve the nonlinear mapping capability.

[0039] After the one-dimensional feature vector is input into the network, it is nonlinearly mapped layer by layer through a multilayer perceptron to convert the abstract spectral features into physically meaningful column concentration values. The column concentration value is the product of the gas concentration and the optical path, which can directly reflect the degree of leakage of the target gas. Finally, using the spatial coordinate index of the original observation tensor record, the one-dimensional column concentration value is refilled into the corresponding two-dimensional empty grid to generate a column concentration distribution matrix that is completely corresponding to the target monitoring field of view. The value of the matrix element is the target gas column concentration at the corresponding spatial location, which can be directly mapped into a pseudo-color image for display on the user terminal, making it convenient for staff to intuitively view the gas leakage distribution.

[0040] In some embodiments, the method further includes a training step of the preset convolutional decoupling architecture: Construct a training dataset containing mixed radiation spectrum samples, known background radiation spectrum samples, and known gas concentration labels; The mixed radiation spectrum samples are input into the convolutional decoupling architecture to be trained; The output of the background decoupling branch is constrained by known background radiation spectrum samples, the output of the concentration inversion mapping network is constrained by known gas concentration labels, and a joint loss function is constructed by combining a cosine similarity penalty term. The weight parameters of the convolutional decoupling architecture are updated using the backpropagation algorithm based on the joint loss function until the model converges.

[0041] Specifically, the training dataset is constructed through two methods: controlled laboratory collection and radiative transfer model synthesis. The sample size is no less than 100,000 sets, covering scenarios with different driving conditions, different background temperatures, and different gas concentrations. The mixed radiation spectrum samples are mixed signals of background radiation and gas radiation, the known background radiation spectrum samples are pure background signals without gas interference, and the known gas concentration labels are the actual column concentration values ​​of the corresponding mixed samples.

[0042] The joint loss function consists of three parts: background constraint loss, concentration inversion loss, and cosine similarity penalty term. The calculation formula is as follows:

[0043] In the formula, Background constraint loss, For concentration inversion loss, This is a cosine similarity penalty term. , , The weights are used as coefficients. The Adam optimizer is used during training, with a learning rate of 1e-4 and a training epoch of no less than 50 epochs. When the fluctuation range of the joint loss function value is less than 1e-5, the model is considered to have converged. The trained network weight parameters are saved for analytical inversion in real-world scenarios.

[0044] In some embodiments, the time-sliding window sampling module employs a dynamic update strategy based on structural similarity when extracting the reference frame sequence: Calculate the structural similarity index between the historical frames and the current original observation tensor within the sliding window; Remove historical frames whose structural similarity index is below a preset threshold; The retained historical frames are weighted and averaged according to the time decay weight to generate a reference frame sequence for characterizing the transient background.

[0045] Specifically, the structural similarity index is used to evaluate the structural consistency between historical frames and the current frame. The preset structural similarity threshold is set to 0.7. When the structural similarity index of a historical frame is lower than this threshold, it indicates that the background corresponding to the historical frame has changed drastically, and the frame is removed from the sliding window.

[0046] For the retained valid historical frames, a weighted average is performed according to the time decay weight. The more recent the historical frame, the greater the weight. The weight allocation follows the exponential decay law. A reference frame sequence representing the transient background is generated by weighted averaging to ensure that the reference frame sequence can adapt to the dynamic changes of the background in real time and improve the extraction accuracy of the baseline background feature vector.

[0047] In some embodiments, the convolutional decoupling architecture further includes an initial gas mask generation step based on a dynamic threshold before feature extraction: Calculate the spectral difference between the original observation tensor and the reference band in the main absorption band of the gas; An adaptive floating threshold is determined based on the distribution of spectral differences in the historical gas-free state of the target monitoring area. The calculated spectral difference values ​​are binarized using the adaptive floating threshold to generate an initial gas mask that marks the potential gas presence region. The initial gas mask is input as a prior attention map into the gas feature extraction branch.

[0048] Specifically, the main absorption band of the gas is the characteristic absorption wavelength range of the target gas, the reference band is the adjacent non-gas absorption band, and the spectral difference is the difference in radiance between the original observation tensor in the main absorption band and the reference band. The larger the difference, the higher the probability that there is gas in the region.

[0049] Based on the spectral differences in the target monitoring area under historical gas-free conditions, their normal distribution parameters (mean μ, standard deviation σ) are statistically analyzed, and an adaptive floating threshold is set to μ+2σ. A binarization algorithm is used to determine the spectral differences. When the difference is greater than the adaptive floating threshold, the area is determined to be a potential gas-containing region and assigned a value of 1; otherwise, it is assigned a value of 0. A two-dimensional initial gas mask is generated, with the mask dimension consistent with the spatial dimension of the original observation tensor. The initial gas mask is used as a prior attention map input to the gas feature extraction branch, guiding the network to focus on potential gas regions, reducing invalid computation in gas-free areas, and improving the system's real-time processing frame rate.

[0050] In some embodiments, the background decoupling branch introduces a cross-scale spatial feature aggregation mechanism when extracting the spectral deformation features of the background continuum: The baseline background feature vector is processed in parallel using multiple two-dimensional dilated convolutions with different dilation rates to obtain spatial texture features under different receptive fields; Spatial texture features from different receptive fields are spliced ​​together along the channel dimension; By using point-to-point convolutional layers to reduce the dimensionality of the stitched features and interact with cross-channel information, spectral deformation features of a background continuum with a global receptive field are generated.

[0051] Specifically, the cross-scale spatial feature aggregation mechanism adopts a multi-channel dilated convolutional layer with a pyramid-like structure, setting up three parallel two-dimensional dilated convolutional channels with dilation rates of 1, 3, and 5 respectively; the channel with a dilation rate of 1 extracts small-scale background texture, the channel with a dilation rate of 3 extracts medium-scale background features, and the channel with a dilation rate of 5 extracts large-scale background features. Each channel uses a 3×3 convolutional kernel to ensure the integrity of feature extraction.

[0052] The spatial texture features extracted from the three channels with different receptive fields are concatenated along the channel dimension, and the number of feature channels after concatenation is the sum of the three channels. The concatenated features are then dimensionality-reduced by a 1×1 point-directed convolutional layer, reducing the number of channels to 64. At the same time, cross-channel information interaction is achieved, multi-scale background features are integrated, and finally, a background continuum spectral deformation feature with a global receptive field is generated. This ensures that the background decoupling branch can accurately fit the radiation distribution law of complex non-uniform backgrounds and improve the decoupling effect between background and gas features.

[0053] Please see Figure 2 The second embodiment of the present invention provides a passive infrared gas analytical inversion system based on a convolutional decoupling architecture and a background parameterization model, comprising: The data preprocessing module 201 is used to: acquire a passive infrared hyperspectral data cube of the target monitoring area, and perform radiometric calibration and spectral alignment on the passive infrared hyperspectral data cube to obtain the original observation tensor of standard radiance; Background parameterization module 202 is used to: input the original observation tensor into a preset background parameterization model, extract the radiation distribution features of the environmental background through a multi-scale spatial background estimation network, and output the corresponding baseline background feature vector; The convolution decoupling module 203 is used to: input the original observation tensor and the baseline background feature vector into a preset convolution decoupling architecture, and use the background decoupling branch and the gas feature extraction branch in the convolution decoupling architecture to perform bidirectional orthogonal constraint processing to separate the pure target gas spectral feature map; The concentration inversion module 204 is used to: input the spectral feature map of the pure target gas into a preset concentration inversion mapping network to obtain the column concentration distribution matrix of the target gas in the target monitoring area; The model training module 205 is used to: construct a training dataset and update the weight parameters of the convolutional decoupling architecture through a joint loss function and a backpropagation algorithm; The meteorological embedding module 206 is used to: acquire synchronous meteorological state parameters and map them into meteorological feature vectors, and embed them into the fully connected layer of the concentration inversion mapping network to correct the inversion results; The mask generation module 207 is used to: calculate the spectral difference value and generate an initial gas mask, which is then used as a prior attention map input to the gas feature extraction branch.

[0054] Specifically, the data preprocessing module 201 is connected to the passive infrared hyperspectral imager and data acquisition card, and is responsible for real-time data acquisition and preprocessing to ensure the accuracy and consistency of the original observation tensor; the background parameterization module 202 is deployed on the edge computing server and has a built-in pre-trained multi-scale spatial background estimation network, which can quickly extract baseline background feature vectors; the convolution decoupling module 203 is the core module of the system, which relies on GPU acceleration to achieve efficient decoupling of background and gas features; the concentration inversion module 204 converts spectral features into quantitative concentration data and outputs an intuitive column concentration distribution matrix; the model training module 205 is used for model training and parameter optimization before system deployment to ensure that the model is adapted to different application scenarios; the meteorological embedding module 206 communicates with the field meteorological station to obtain meteorological parameters such as temperature and humidity, and realizes environmental compensation for the inversion results; the mask generation module 207 generates an initial gas mask to guide the convolution decoupling module to focus on key areas and improve computational efficiency.

[0055] The modules communicate with each other via a high-speed data bus, with transmission latency controlled within 10ms to ensure the real-time performance of the system. The modules adopt a modular design, which allows for flexible adjustment of the parameters of each module according to the actual application scenario, thereby improving the versatility and scalability of the system and making it suitable for passive infrared gas monitoring needs of different scales and environments.

[0056] It should be noted that the passive infrared gas analysis and inversion system based on convolutional decoupling architecture and background parameterization model provided in this embodiment of the invention is used to execute all the process steps of the passive infrared gas analysis and inversion method based on convolutional decoupling architecture and background parameterization model in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0057] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0058] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0059] The embodiments, implementation methods, and related technical features of this application can be combined and substituted for each other without conflict.

[0060] The above are merely preferred embodiments of this application and are not intended to limit this application in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this application without departing from the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A passive infrared gas analytical inversion method based on a convolutional decoupling architecture and a background parameterization model, characterized in that, include: Step S101: Obtain the passive infrared hyperspectral data cube of the target monitoring area, and perform radiometric calibration and spectral alignment on the passive infrared hyperspectral data cube to obtain the original observation tensor of standard radiance. Step S102: Input the original observation tensor into the preset background parameterization model, extract the radiation distribution features of the environmental background through the multi-scale spatial background estimation network, and output the corresponding baseline background feature vector. Step S103: Input the original observation tensor and the baseline background feature vector into a preset convolutional decoupling architecture, and use the background decoupling branch and the gas feature extraction branch in the convolutional decoupling architecture to perform bidirectional orthogonal constraint processing to separate the pure target gas spectral feature map. Step S104: Input the spectral feature map of the pure target gas into a preset concentration inversion mapping network to obtain the column concentration distribution matrix of the target gas in the target monitoring area.

2. The method according to claim 1, characterized in that, The preset background parameterization model includes a time sliding window sampling module, a local contrast calculation unit, and a multi-scale spatial background estimation network connected in sequence; step S102 specifically includes: The reference frame sequence is extracted from the raw observation tensor of continuous time sequence using the time sliding window sampling module. The local contrast calculation unit is used to calculate the pixel-level radiometric difference between the current original observation tensor and the reference frame sequence to generate a background variation map; The current original observation tensor is merged with the background variation map and then input into the multi-scale spatial background estimation network to extract the radiation distribution features of the environmental background and generate the baseline background feature vector.

3. The method according to claim 1, characterized in that, The internal operating logic of the convolutional decoupling architecture specifically includes: The background decoupling branch receives the original observation tensor and the baseline background feature vector, and extracts the spectral deformation features of the background continuum through a three-dimensional convolutional layer; The gas feature extraction branch receives the original observation tensor and extracts the fine absorption peak features of gas molecules through a one-dimensional spectral hole convolution layer. The cosine similarity between the feature tensor output by the background decoupling branch and the feature tensor output by the gas feature extraction branch is calculated, and the network parameters are updated based on the objective of minimizing the cosine similarity to achieve the bidirectional orthogonal constraint processing.

4. The method according to claim 3, characterized in that, The gas feature extraction branch also includes a spectral spatial attention fusion mechanism. The specific steps for separating the pure target gas spectral feature map include: The spectral attention module is used to redistribute the channel-dimensional weights of the features output by the one-dimensional spectral hole convolutional layer, thereby enhancing the response weights of specific gas absorption bands. The spatial attention module is used to perform spatial dimension masking on the features after weight redistribution to suppress background interference in non-gas distribution areas. The features processed by the spectral attention module and the spatial attention module are fused by element-wise multiplication to generate the pure target gas spectral feature map.

5. The method according to claim 1, characterized in that, The preset concentration inversion mapping network is a fully connected regression residual network, and the specific steps for obtaining the column concentration distribution matrix of the target gas within the target monitoring area include: Flatten the spectral feature map of the pure target gas into a one-dimensional feature vector; The one-dimensional feature vector is input into a multilayer fully connected regression residual network, and the corresponding column concentration value is output through nonlinear mapping layer by layer by a multilayer perceptron. The column concentration values ​​are reconstructed into a two-dimensional matrix based on the original spatial coordinate information to obtain the column concentration distribution matrix.

6. The method according to claim 1, characterized in that, The method also includes a training step for the preset convolutional decoupling architecture: Construct a training dataset containing mixed radiation spectrum samples, known background radiation spectrum samples, and known gas concentration labels; The mixed radiation spectrum samples are input into the convolutional decoupling architecture to be trained; The output of the background decoupling branch is constrained by known background radiation spectrum samples, the output of the concentration inversion mapping network is constrained by known gas concentration labels, and a joint loss function is constructed by combining a cosine similarity penalty term. The weight parameters of the convolutional decoupling architecture are updated using the backpropagation algorithm based on the joint loss function until the model converges.

7. The method according to claim 2, characterized in that, The time-sliding window sampling module employs a dynamic update strategy based on structural similarity when extracting the reference frame sequence: Calculate the structural similarity index between the historical frames and the current original observation tensor within the sliding window; Remove historical frames whose structural similarity index is below a preset threshold; The retained historical frames are weighted and averaged according to the time decay weight to generate a reference frame sequence for characterizing the transient background.

8. The method according to claim 3, characterized in that, Before feature extraction, the convolutional decoupling architecture also includes an initial gas mask generation step based on a dynamic threshold: Calculate the spectral difference between the original observation tensor and the reference band in the main absorption band of the gas; An adaptive floating threshold is determined based on the distribution of spectral differences in the historical gas-free state of the target monitoring area. The calculated spectral difference values ​​are binarized using the adaptive floating threshold to generate an initial gas mask that marks the potential gas presence region. The initial gas mask is input as a prior attention map into the gas feature extraction branch.

9. The method according to claim 3, characterized in that, The background decoupling branch introduces a cross-scale spatial feature aggregation mechanism when extracting the spectral deformation features of the background continuum: The baseline background feature vector is processed in parallel using multiple two-dimensional dilated convolutions with different dilation rates to obtain spatial texture features under different receptive fields; Spatial texture features from different receptive fields are spliced ​​together along the channel dimension; By using point-to-point convolutional layers to reduce the dimensionality of the stitched features and interact with cross-channel information, spectral deformation features of a background continuum with a global receptive field are generated.

10. A passive infrared gas analytical inversion system based on a convolutional decoupling architecture and a background parameterization model, characterized in that, include: The data preprocessing module is used to acquire the passive infrared hyperspectral data cube of the target monitoring area, and to perform radiometric calibration and spectral alignment on the passive infrared hyperspectral data cube to obtain the original observation tensor of standard radiance. The background parameterization module is used to input the original observation tensor into a preset background parameterization model, extract the radiation distribution features of the environmental background through a multi-scale spatial background estimation network, and output the corresponding baseline background feature vector. The convolutional decoupling module is used to input the original observation tensor and the baseline background feature vector into a preset convolutional decoupling architecture, and to perform bidirectional orthogonal constraint processing on the background decoupling branch and the gas feature extraction branch in the convolutional decoupling architecture to separate the pure target gas spectral feature map. The concentration inversion module is used to input the spectral feature map of the pure target gas into a preset concentration inversion mapping network to obtain the column concentration distribution matrix of the target gas in the target monitoring area; The model training module is used to build a training dataset and update the weight parameters of the convolutional decoupling architecture through a joint loss function and backpropagation algorithm. The meteorological embedding module is used to acquire synchronous meteorological state parameters and map them into meteorological feature vectors, which are then embedded into the fully connected layer of the concentration inversion mapping network to correct the inversion results. The mask generation module is used to calculate the spectral difference value and generate an initial gas mask, which is then used as a prior attention map input to the gas feature extraction branch.

Images