Spectral aliasing decoupling and concentration inversion method under cross influence of multi-source environmental factors

Abstract

The invention discloses a spectrum aliasing decoupling and concentration inversion method under the cross influence of multi-source environmental factors, belongs to the field of industrial process control and environment monitoring, and constructs an environment-spectrum collaborative fusion decoupling model for concentration prediction. The method specifically comprises the following steps: respectively collecting absorption spectrum signals of specified mixed gas at different temperatures, pressures and known concentrations, meanwhile, collecting environmental parameter data, constructing a multi-source data set, and carrying out denoising, dimension reduction and preprocessing on the multi-source data set; constructing a self-supervised feature extraction network for adaptive modulation of environmental parameters to realize deep fusion of spectrum and environmental information; the feature expression capability and generalization performance of the self-supervised feature extraction network are improved by using a self-supervised learning mechanism; and constructing a BPBO-GRNN self-adaptive concentration inversion optimization model for realizing inversion of mixed gas concentration and self-adaptive optimization of model parameters. According to the invention, high-precision concentration inversion and stable detection of the aliasing gas can be realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of industrial process control and environmental monitoring, and particularly relates to a spectrum aliasing decoupling and concentration inversion method under the cross influence of multiple source environmental factors. BACKGROUND

[0002] In the field of industrial process control and environmental monitoring, especially in the complex and harsh application scenarios such as continuous monitoring of flue gas emission of thermal power plants, high technical requirements are put forward for accurate and real-time online measurement of gas concentration. In order to meet this challenge, tunable diode laser absorption spectroscopy (TDLAS) technology is widely regarded as a key technology in this field due to its inherent advantages of high sensitivity, high selectivity and fast online response. However, when this technology is applied to the above-mentioned complex industrial environment, its measurement accuracy and reliability still face severe technical challenges, which limit the full play of its performance.

[0003] In the flue gas denitrification process of thermal power plants, in order to reduce the pollutant nitrogen oxide to harmless nitrogen, ammonia is injected into high-temperature flue gas as a reducing agent. In order to ensure the denitrification efficiency, the injection amount of ammonia is usually slightly excessive, resulting in a part of unreacted NH3 being discharged with the flue gas, forming "ammonia escape". At the same time, incomplete combustion of fuel in the boiler will inevitably produce carbon monoxide. Therefore, in the final emission flue, CO and NH3 must coexist and fully mix. These two gases have partial overlapping of absorption spectra in the commonly used near-infrared monitoring band, forming serious spectral aliasing.

[0004] The dynamic nature of industrial processes results in dramatic and frequent fluctuations in temperature and pressure within the flue. These environmental parameters are not simply background noise, but rather they profoundly change the absorption spectral line shape of gas molecules in a coupled and nonlinear manner. Studies have shown that pressure changes the spectral line width through collision broadening, and temperature changes the spectral line width through Doppler broadening. Temperature more critically modulates the integral absorption intensity of the spectral line by directly changing the molecular energy level population. The key is that these effects have strong molecular specificity, and the spectral line shape of NH3 is extremely sensitive to pressure and temperature, while CO is relatively sluggish. Therefore, the final observed spectral shape is a complex result of the nonlinear coupling of concentration, pressure and temperature. The difference and nonlinearity of the response of different components to environmental parameters further exacerbate the complexity of spectral aliasing decoupling.

[0005] High-precision spectral analysis of mixed gas under multi-source environmental interference has always been a research hotspot and difficulty in the field of industrial sensing. Early research mainly relied on the Lambert-Beer law, and the concentration inversion was performed by measuring the absorption peak at a specific wavelength. However, in the presence of CO and NH3, the absorption spectra of which are severely overlapped in the near-infrared band, there is almost no independent absorption peak that is not disturbed, which fundamentally renders this method ineffective. To solve the problem of spectral overlap, domestic and foreign scholars introduced chemometrics algorithms such as multiple linear regression, principal component regression, and partial least squares. This kind of method regards the mixed spectrum as a linear superposition of the pure component spectrum, which can separate the overlapping signals to some extent. However, the core linear assumption is severely shaken in dynamic industrial environments, and the dramatic fluctuations in pressure and temperature can cause significant nonlinear broadening and deformation of the spectral lines of molecules such as NH3. Linear models such as PLS cannot effectively model the nonlinear spectral changes driven by environmental parameters, and when the working conditions deviate from the calibration conditions, the model accuracy will decrease sharply, and the robustness is poor. Although subsequent research shifted to nonlinear models such as neural networks, the current mainstream strategy of simply concatenating high-dimensional spectra with low-dimensional environmental data has introduced deep flaws such as information flooding and mismatched feature attributes, and has not effectively learned the key interference mechanisms.

[0006] In summary, the existing spectral inversion methods are generally sensitive to temperature and pressure fluctuations, the linear assumption is invalid, and the multi-source feature fusion is insufficient, making it difficult to achieve high-precision concentration inversion and stable detection of overlapping gases in complex industrial conditions. SUMMARY

[0007] To solve the problem of nonlinear spectral deformation specific to molecules caused by dramatic fluctuations in temperature, pressure, and other environmental parameters in complex industrial scenarios such as continuous monitoring systems for flue gas emissions in thermal power plants, the present invention proposes a method for spectral overlap decoupling and concentration inversion under the cross-influence of multi-source environmental factors. This method realizes cross-modal fusion from static environmental input to dynamic regulation of spectral features through multi-modal environmental parameter deep characterization and adaptive modulation mechanism; and combined with the adaptive optimization model of generalized regression neural network (GRNN) based on peregrine optimization algorithm (BPBO), it breaks through the limitations of traditional algorithms that are strongly dependent on parameters and prone to local optimization. Compared with existing technologies, the present invention has made innovative breakthroughs in adaptive compensation of spectral nonlinear deformation and dynamic modeling of features, and has realized adaptive compensation of spectral nonlinear deformation under temperature and pressure coupling interference and dynamic modeling of features, providing an original technical path for high-precision detection of multi-component gas in complex industrial environments.

[0008] The technical solutions of the present invention are as follows: A method for spectral aliasing decoupling and concentration inversion under the combined influence of multiple environmental factors is proposed. An environment-spectral collaborative fusion decoupling model is constructed for concentration prediction. This model includes a multi-source data processing module, a self-supervised feature extraction network, a self-supervised auxiliary decoding module, and a BPBO-GRNN adaptive concentration inversion optimization model. The method specifically includes the following steps: Step 1: Collect the absorption spectral signals of the specified mixed gas at different temperatures, pressures, and known concentrations, and simultaneously collect environmental parameter data to construct a multi-source dataset; Step 2: Denoise, reduce dimensionality, and preprocess the multi-source dataset using the multi-source data processing module; Step 3: Construct a self-supervised feature extraction network with adaptive modulation of environmental parameters to achieve deep fusion of spectral and environmental information and obtain fused feature vectors; Step 4: Construct a self-supervised auxiliary decoding module. This module utilizes a self-supervised learning mechanism to improve the feature representation ability and generalization performance of the self-supervised feature extraction network, and fully explore unlabeled data. Step 5: Construct a BPBO-GRNN adaptive concentration inversion optimization model to realize the inversion of mixed gas concentration and adaptive optimization of model parameters; Step 6: Collect the absorption spectrum signal of the unknown concentration of the gas to be measured under the actual operating conditions of the thermal power plant, and at the same time collect the actual environmental parameter data to construct a multi-source dataset; import the multi-source dataset into the environmental-spectral synergistic fusion decoupling model trained by the laboratory calibration data for concentration prediction.

[0009] Further, the specific process of step 1 is as follows: First, under different temperature and pressure conditions, the absorption spectrum signals of the CO and NH3 mixture with known concentration gradients are collected, and the corresponding environmental parameter data are recorded to construct a multi-source dataset; the multi-source dataset is used as the training set of the model; then, the aliased spectrum signals under unknown concentration conditions are collected within the same temperature and pressure control range and used as the model validation set.

[0010] Furthermore, in step 2, the data processing module includes an intelligent noise reduction module, a preprocessing module, and a dimensionality reduction module; the specific process is as follows: Step 2.1: Construct an intelligent denoising module. The intelligent denoising module adopts a pre-trained one-dimensional U-Net convolutional neural network structure. The multi-source dataset is input into the one-dimensional U-Net convolutional neural network for denoising, and the mapping relationship is as follows: ; in, It is the first High-fidelity spectra of individual samples; It is the first The original spectra of each sample; trainable parameters of the one-dimensional U-Net convolutional neural network; a function mapping corresponding to the one-dimensional U-Net convolutional neural network structure; Step 2.2, introduce a non-parametric distribution alignment mechanism of quantile mapping to construct a preprocessing module; the preprocessing module includes a distribution feature analysis unit and a non-parametric distribution alignment unit, and the specific working process is as follows: Step 2.2.1, based on the distribution feature analysis unit, the empirical distribution estimation is carried out, and the empirical cumulative distribution function of each spectral feature and environmental parameter is calculated: ; wherein, is the empirical cumulative distribution function; represents the total number of samples; is a threshold value of the feature value; represents the feature value of the i th sample; is an indicator function, which takes 1 when is less than or equal to , otherwise 0; Step 2.2.2, based on the non-parametric distribution alignment unit, quantile mapping is carried out, and the quantile value of the high-fidelity spectrum is mapped through the inverse cumulative distribution function of the target distribution to obtain the standardization result : ; wherein, is the inverse cumulative distribution function of the standard normal distribution; Step 2.3, construct a dimension reduction module.

[0011] Further, in the step 2.3, the dimension reduction module includes an encoder submodule, a decoder submodule and a feature reconstruction verification unit, and the specific working process is as follows: Step 2.3.1, the encoder submodule maps the standardization result to a low-dimensional latent space through an autoencoder mapping function to obtain a low-dimensional feature vector : ; wherein, represents the autoencoder mapping function; is a set of trainable parameters; is the compressed spectral dimension; Step 2.3.2, the decoder submodule reconstructs the low-dimensional feature vector back to the original space through the inverse mapping function of the decoder: ; wherein,​ This represents the inverse mapping function of the decoder; This is the set of trainable parameters for the decoder; For the reconstruction result; Step 2.3.3: The goal of the feature reconstruction verification unit is to minimize the mean square error between the input and the reconstructed spectrum. By minimizing the loss function, the network can learn the nonlinear feature structure of the spectral data; the loss function... for: ; in, and They represent the first Standardization and reconstruction results for each sample; This represents the L2 norm.

[0012] Furthermore, in step 3, the self-supervised feature extraction network includes an environment modulation branch, a spectral backbone network, and a feature fusion module; the specific process is as follows: Step 3.1: Construct an environment modulation branch to obtain high-level environment feature vectors. Furthermore, a multi-head self-attention mechanism and a gating control structure are introduced; Step 3.2: Construct a spectral backbone network consisting of multiple environmentally adaptive residual blocks stacked in series, and extract environmentally adaptive features from the spectral data step by step to obtain the environmentally adaptive spectral feature vector. ; Step 3.3: Construct a feature fusion module to perform feature fusion, specifically by fusing the environmental adaptive spectral feature vector. High-level environmental feature vector The features are concatenated to obtain a fused feature vector. : ; in, This refers to the vector concatenation operation along the feature dimension.

[0013] Furthermore, in step 3.1, the environment modulation branch includes a multimodal environment parameter self-attention representation submodule, an environment feature deep reconstruction submodule, and an adaptive control unit. The specific working process is as follows: Step 3.1.1: Construct a multimodal environment parameter self-attention representation submodule based on a multi-head self-attention mechanism to obtain environmental feature representations. : ; in, For the Softmax function; , , These are the query vector, key vector, and value vector corresponding to each attention head; For transpose; The dimension of the key vector; Step 3.1.2: Construct a deep environmental feature reconstruction submodule using a multi-layer residual network structure; As the input to the deep reconstruction submodule of environmental features, let the initial feature vector be denoted as . This submodule consists of multiple stacked residual blocks; for the first... The residual blocks are calculated as follows: ; in, For the first The feature vector output by the residual block is used as the first... Input of each residual block; For the first The feature vectors output by each residual block; It is the ReLU activation function; For the first One residual block; The total number of residual blocks; after The hierarchical feature extraction and superposition of layer residual blocks yields the high-level environmental feature vector. : ; in, For the first The feature vectors output by each residual block; Step 3.1.3: Construct an adaptive control unit composed of multiple fully connected layers. For the first layer in the spectral backbone network... An environment adaptive residual block that needs to be modulated; the adaptive control unit uses high-level environmental feature vectors. Generate the corresponding environmental modulation parameters; specifically as follows: First, the high-level environmental feature vector The inputs are fed into three independent fully connected layers to generate preliminary environmental modulation parameters and a gating signal. The preliminary environmental modulation parameters include a preliminary scaling factor and a preliminary offset factor. Then, based on the gated signal, the preliminary environmental modulation parameters are weighted and fused to obtain the final environmental modulation parameters: ; ; in, For the first Gating signals for an environment-adaptive residual block; , The first a preliminary scaling factor and a preliminary offset factor of the environment adaptive residual block; 、 a final scaling factor and a final offset factor of the nth environment adaptive residual block, respectively; denotes element-wise multiplication; is a unit vector; Finally, the obtained environment modulation parameters are applied to the feature map of the convolution block to realize adaptive affine transformation.

[0014] Further, in step 3.2, each environment adaptive residual block adopts a residual structure composed of a main path and a shortcut path; the calculation formula of the main path is: ; ; wherein, is the output feature map of the first stage; is the first batch normalization operation; denotes one-dimensional convolution operation; is the output of the previous environment adaptive residual block; is the final output of the main path; is the second batch normalization operation; denotes one-dimensional convolution operation; The calculation formula of the shortcut path is: ; wherein, is the output of the shortcut path; denotes one-dimensional convolution operation with a 1x1 convolution kernel; is the dimension of the feature map; After element-wise addition of the output of the main path and the output of the shortcut path, the spectral feature map is obtained: ; is the output of a single environment adaptive residual block, and for the nth environment adaptive residual block: ; wherein, is the spectral feature map output by the nth environment adaptive residual block; is the nth environment adaptive residual block; is the spectral feature map output by the nth environment adaptive residual block; is obtained by After sequential processing of each environment-adaptive residual block, the final spectral feature map is obtained. ; A global average pooling operation is used for transformation to obtain the final environment-adaptive spectral feature vector. : ; in, This is a global average pooling operation.

[0015] Furthermore, in step 4, the self-supervised auxiliary decoding module includes an auxiliary decoder, a self-supervised constraint unit, and a feature optimization and verification unit. The specific working process is as follows: Step 4.1: Fuse the feature vectors The input auxiliary decoder extracts physically consistent deep feature representations. The auxiliary decoder reconstructs the spectral data through a multi-layer fully connected structure, remapping the highly compressed features back to the original spectral space, and finally generating the reconstructed spectrum. Step 4.2: Define the reconstruction loss function in the self-supervised constraint unit. The difference between the original spectrum and the reconstructed spectrum is measured: ; in, For the first The original spectra of each sample; For the first The reconstructed spectra of each sample; the number of original spectra corresponds to the number of reconstructed spectra. Step 4.3: Construct a feature optimization and verification unit to continuously monitor reconstruction performance and adjust the structural parameters of the self-supervised feature extraction network during the training phase, ensuring that the model obtains stable feature representation under unlabeled data conditions; During training, the loss function is reconstructed. As a self-supervised signal, all trainable parameters, including the auxiliary decoder, feature fusion module, environmental modulation branch, and spectral backbone network, are jointly optimized through an end-to-end backpropagation algorithm. During the training phase, an early stopping strategy is adopted to determine the optimal stopping point based on the reconstruction error of the validation set to prevent overfitting. After training, the auxiliary decoder is removed, and only the optimized self-supervised feature extraction network is retained.

[0016] Furthermore, in step 5, the BPBO-GRNN adaptive concentration inversion optimization model includes a GRNN concentration prediction submodule and a BPBO adaptive optimization submodule, the specific process of which is as follows: Step 5.1, based on the structure of the generalized regression neural network, the GRNN concentration prediction submodule is constructed, and the Gaussian kernel function is used to weight the feature similarity modeling to control the kernel bandwidth with the smoothing factor, so as to obtain the concentration prediction results sensitive to the changes of spectrum and environmental characteristics; the fusion feature vector of the first training sample is defined For any input fusion feature vector, the similarity of the first training sample is measured by the Euclidean distance, the fusion feature vector of the first training sample is , and the Gaussian kernel function is converted into kernel weight: ; Wherein, is the kernel weight between and ; is the exponential function; is the Euclidean distance function; is the smoothing factor; The concentration prediction value is: ; Wherein, is the total number of training samples; is the true concentration label of the first training sample; Step 5.2, construct the BPBO global adaptive optimization submodule, which proposes a generalized regression neural network adaptive optimization mechanism based on the peregrine optimization algorithm, simulates the global exploration and local development behavior of peregrine in the process of hunting, and dynamically searches for the optimal smoothing factor.

[0017] Further, the specific process of step 5.2 is: First, the optimization process of the smoothing factor is modeled as a minimization problem, and the smoothing factor is updated as the individual position of the peregrine optimization algorithm to find the optimal smoothing factor; for any given value, its fitness is calculated as follows: ; Wherein, is the number of samples in the validation set; is the first validation sample; is the true concentration label of the first validation sample; Indicates the use of the current The constructed GRNN concentration prediction submodule predicts the concentration values ​​of the validation samples; Set population size With maximum number of iterations Set the smoothing factor The search space is a preset range ,in , They are respectively The minimum and maximum values; random initialization within the search space. The positions of individual raptors form the initial population; the fitness value of each individual is calculated, and the position of the individual with the lowest fitness value is recorded as the current global optimum. ; In each iteration of the Raptor optimization algorithm, a random number within the interval [0, 1] is generated. Preset a threshold As a switching threshold between the exploration and development phases; if the updated individual position exceeds the search space, it is truncated and mapped back to the search space range; the specific iterative process of the Raptor optimization algorithm is as follows: like The algorithm will then enter the global exploration phase; the individual position update formula is as follows: ; in, For the first The individual in the first The position of the next iteration corresponds to the [number]th iteration. The smoothing factor value at the next iteration; The updated position; It is a random number in the range [-1, 1]; like The algorithm then enters the local development phase; this phase contains two sub-phases, determined by the number of iterations. and The relationship determines; when At that time, execute the dive-and-pursue strategy: ; in, For the first The average of all individual positions in the next iteration; It is a random number within the range [0, 1]. when At that time, execute a precise attack strategy: ; in, For the Lévy flight function; After each position update, the fitness value is recalculated, and if , the individual position is updated; When the number of iterations reaches , the final global optimal position is output as the optimal smoothing factor; After obtaining the optimal smoothing factor , the BPBO-GRNN adaptive concentration inversion optimization model is based on predicting the concentration of any input fusion feature vector to obtain the optimal concentration prediction value : .

[0018] The beneficial technical effects brought by the present application are as follows.

[0019] 1. The present application constructs a double-branch deep fusion structure of spectrum and environmental parameters, innovatively designs an environmental adaptive residual block (EAR-Block), embeds the dynamic modulation factor generated by the environmental parameters into the spectrum feature extraction process, realizes dynamic matching and weight balance of high-dimensional spectrum and low-dimensional environmental parameters, improves the robustness of feature expression, and effectively solves the problem of spectrum information drowning environmental characteristics in traditional methods.

[0020] 2. A multi-modal environmental parameter self-attention representation submodule is constructed based on a multi-head self-attention mechanism, combined with the nonlinear mapping and reconstruction of a multi-layer residual network, to generate high-dimensional feature embedding with global dependence, realizing deep fusion of spectrum and environmental information.

[0021] 3. A self-supervised auxiliary decoding module is introduced to learn the physical law of spectrum and environmental influence mode (i.e. using spectrum reconstruction constraints to realize feature learning of unlabeled samples), without the need for a large number of labeled samples to improve the adaptability and generalization ability of the model to complex temperature-pressure coupling interference.

[0022] 4. The smoothing factor of the generalized regression neural network (GRNN) is globally adaptively optimized by adopting the peregrine optimization algorithm (BPBO), avoiding the low efficiency and local optimal problem of artificial parameter adjustment, realizing intelligent optimization of model parameters, and significantly improving the precision and convergence performance of the concentration inversion model.

[0023] 5. An end-to-end trainable multi-level fusion inversion framework is constructed, which significantly improves the real-time performance and engineering deployability of the model, can realize fast feature extraction and real-time concentration prediction under complex working conditions, is convenient for embedding into an industrial monitoring system for online application, and significantly improves the overall response speed and engineering practical value of the system. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1A flow chart of the spectrum aliasing decoupling and concentration inversion method under the multi-source environmental factor cross-influence of the present application.

[0025] Figure 2 A denoising result and residual analysis chart; wherein (a) is the original spectrum, (b) is the spectrum after denoising, and (c) is the residual analysis chart.

[0026] Figure 3 A gradient comparison result chart of the real concentration and the predicted concentration of NH3 after the method of the present application is adopted.

[0027] Figure 4 A gradient comparison result chart of the real concentration and the predicted concentration of CO after the method of the present application is adopted.

[0028] Figure 5 A residual distribution chart of NH3 and CO after the method of the present application is adopted; wherein (a) is the residual distribution chart of NH3, and (b) is the residual distribution chart of CO.

[0029] Figure 6 A gradient comparison result chart of the real concentration and the predicted concentration of NH3 after the method of the present application is adopted.

[0030] Figure 7 A gradient comparison result chart of the real concentration and the predicted concentration of CO after the method of the present application is adopted.

[0031] Figure 8 A residual distribution chart of NH3 and CO after the method of the present application is adopted; wherein (a) is the residual distribution chart of NH3, and (b) is the residual distribution chart of CO. DETAILED DESCRIPTION

[0032] The present application will be further described in detail below in combination with the accompanying drawings and specific embodiments: The present application innovatively proposes a deep collaborative fusion technical scheme based on self-supervised feature extraction, deep representation of environmental parameters and BPBO-GRNN joint optimization, which first realizes a dynamic integrated modeling mechanism of environmental information, spectral features and inversion model, breaks through the technical bottlenecks of traditional methods such as split modeling of spectral features and environmental parameters, linear hypothesis failure and dependence on artificial parameter adjustment, aims to solve the problems of spectral nonlinear deformation and aliasing caused by the dramatic fluctuations of temperature, pressure and other environmental parameters, and overcome the information flooding problem of high-dimensional spectrum to low-dimensional environmental parameters, so as to realize the high robustness inversion of the target gas concentration in the seriously aliased spectrum.

[0033] In view of the spectral overlap decoupling and concentration inversion technology challenges of CO and NH3 under the condition of temperature and pressure coupling interference, the application originally constructs an intelligent spectrum processing and concentration inversion method with multi-module cooperation. The core innovation of the method mainly focuses on the following modules: a multi-source data preprocessing module, a non-parametric distribution alignment mechanism based on quantile mapping is introduced, breaking through the limitation of traditional Z-score standardization relying on Gaussian assumption; a self-supervised feature extraction network, a deep fusion mechanism of environmental parameters and spectral features is constructed, a self-attention and gate regulation structure is introduced, the self-learning and dynamic adjustment of environmental features are realized, and an environmental adaptive residual block is used to embed an environmental adaptive modulation mechanism in the spectral backbone network, realizing the dynamic compensation and fusion of spectral and environmental features; a BPBO-GRNN adaptive concentration inversion optimization model based on peregrine optimization algorithm realizes intelligent optimization and global optimal configuration of GRNN model parameters.

[0034] The application realizes high-precision inversion of gas concentration and adaptive compensation of environmental interference by integrating a deep learning model of multi-modal environmental parameters and spectral data. The spectral data is processed by an intelligent denoising module, deep features are extracted using a one-dimensional U-Net convolutional neural network, and spectral details and absorption peak structures are maintained. The dimension reduction module compresses the spectral data into low-redundancy features for subsequent processing. In the self-supervised feature extraction network block, environmental information is embedded and nonlinearly mapped into high-dimensional features through a self-attention mechanism and a residual network, generating a dynamic modulation signal, and is deeply fused with spectral features; the environmental adaptive residual block further calibrates the spectral features to realize dynamic compensation. The self-supervised auxiliary decoding module optimizes the network feature expression through self-supervised learning, pre-trains with unlabeled data, and improves the generalization ability of the model. Then, the BPBO-GRNN adaptive concentration inversion optimization model combines GRNN and peregrine optimization algorithm to realize high-precision concentration prediction by adaptively adjusting the smoothing factor. Through multi-level feature fusion and optimization adjustment, the application effectively improves the precision and robustness of concentration inversion, and has good adaptability and generalization ability in complex environments.

[0035] The application takes the complex working conditions of uncertain temperature and pressure in flue gas emission of thermal power plants as the background, selects typical mixed gas CO and NH3 as the target object, and constructs a spectral overlap decoupling and concentration inversion experimental system under the action of multi-source environmental factors. The method is not only suitable for emission monitoring of the denitrification process of thermal power plants, but also can be extended to other multi-component gas overlap spectrum measurement scenes affected by environmental interference.

[0036] Experimental conditions and test environment: The experiment uses a controllable environment spectrum acquisition system independently built by the laboratory. The system mainly includes a temperature control unit, a pressure regulation unit, a gas mixing module, and a spectrum detection module. The temperature control unit can be continuously adjustable in the range of 0-300℃; the pressure regulation unit can be stably controlled in the range of 0.5-3.0 atm. Real-time monitoring and closed-loop feedback of working condition parameters are realized through high-precision temperature controller and pressure sensor, ensuring accurate controllability and repeatability of experimental environment parameters.

[0037] The present application constructs an environment-spectrum collaborative fusion decoupling model for concentration prediction, mainly including a multi-source data processing module, a self-supervised feature extraction network, a self-supervised auxiliary decoding module, and a BPBO-GRNN adaptive concentration inversion optimization model. As shown in Figure 1 Step 1, respectively collect the absorption spectrum signals of the specified mixed gas under different temperatures, pressures and known concentrations, and collect the environmental parameter data at the same time, and construct a multi-source data set; First, under different temperatures (25-250℃) and pressures (0.8-2.5 atm), the absorption spectrum signals of CO and NH3 mixed gas with known concentration gradient are collected, and the corresponding environmental parameter data is recorded, and a multi-source data set is constructed. This part of data is used as the training set of the model for feature learning and parameter optimization. Then, under the same temperature and pressure control range, the mixed spectrum signals under unknown concentration conditions are collected as the model verification set, which is used to evaluate the inversion accuracy and generalization performance.

[0038] During the experiment, the volume fraction of CO and NH3 is controlled by a standard gas proportioning device, in which the CO concentration gradient range is 320 ppm-1200 ppm, and the NH3 concentration gradient range is 320 ppm-2000 ppm. The setting range of temperature and pressure covers the typical variation range of the flue gas discharge pipeline of the thermal power plant, which can effectively simulate the gas mixing and spectrum interference characteristics under real complex working conditions.

[0039] The collected spectrum has obvious absorption overlap in the near-infrared band, showing partial overlap of CO and NH3 spectral lines. The spectral line broadening and nonlinear change of absorption intensity caused by temperature and pressure change further aggravate the spectral interference, providing a typical verification condition for subsequent data denoising, dimensionality reduction and concentration inversion modeling.

[0040] Step 2, denoising, dimensionality reduction and preprocessing of the multi-source data set through the multi-source data processing module; the data processing module includes an intelligent denoising module, a preprocessing module and a dimensionality reduction module, and the specific process is as follows: Step 2.1, construct an intelligent denoising module for intelligent denoising of the multi-source data set; ​The intelligent denoising module adopts a pre-trained one-dimensional U-Net convolutional neural network structure, which mainly includes an encoder, a decoder, and skip connection units. The encoder extracts deep feature information of the spectrum through multi-layer one-dimensional convolution and downsampling operations. The decoder reconstructs the features layer by layer through upsampling and convolution operations. Skip connection units are used to transfer and fuse shallow and deep features between corresponding layers to maintain the integrity of spectral details and absorption peak structure.

[0041] A multi-source dataset is input into a one-dimensional U-Net convolutional neural network for denoising. This module autonomously learns the mapping relationship from the noisy spectral domain to the clean spectral manifold from the data through a highly nonlinear mapping function, transforming any input raw spectrum into a denoised high-fidelity spectrum. The mapping relationship is as follows: ; in, It is the first High-fidelity spectra of individual samples; It is the first The original spectra of each sample; These are the trainable parameters of a one-dimensional U-Net convolutional neural network; This represents the function mapping corresponding to the one-dimensional U-Net convolutional neural network structure; specifically, the internal structure of this convolutional neural network is optimized for the input data.

[0042] The encoder progressively compresses the input spectrum through multiple layers of one-dimensional convolution and downsampling operations to extract deep feature information. Assume the encoder of this invention comprises... One-dimensional convolution and downsampling operations in the first layer, with the input feature map of the first layer. By the The original spectra of each sample Direct import, i.e. ;No. The input feature map of layer 1 is the first layer. Output feature map of the layer , No. The output feature map of the layer is The specific calculation process is as follows: ; ; in, For the first Layers are used for feature maps of skip connections; This represents a one-dimensional convolution operation; It is a non-linear activation function; This is a max pooling operation used to compress and extract features.

[0043] The decoder functions in the opposite way to the encoder; it is responsible for accurately reconstructing the pure spectral signal from abstract feature representations. The decoder is configured accordingly. Layer upsampling operation. The decoder starts from the network layer... Output feature map of the layer Begin, through Layer-by-layer upsampling recovers the original dimensions of the spectrum. Let the decoder... The input feature map of the layer is Upsampling is performed through a transposed convolutional layer, and the calculation is as follows: ; in, For decoder number Upsampled feature map of the layer; This is a transposed convolutional layer.

[0044] Based on the upsampled feature maps, the skip connection unit first concatenates the high-resolution feature maps of the corresponding layers of the encoder along the channel dimension through long-distance skip connections; then, it directly supplies the shallow features of the encoder containing precise location information to the decoder, thereby achieving effective fusion of deep feature information and shallow detail information to generate the final output of the current layer of the decoder: ; ; in, Indicates the encoder's first... The feature map output by the layer, the encoder's first layer Layer and decoding layer The layers are relatively symmetrical; For decoder number Feature map after layer fusion; This refers to the vector concatenation operation along the feature dimension; For decoder number The final output of the layer; This represents a one-dimensional convolution operation, used for preliminary feature extraction and channel integration of the feature map concatenated by skip connections; This represents a one-dimensional convolution operation, used in... Further feature reconstruction is performed based on the extracted features; go through The upsampling and skip fusion of the layers, and finally the output convolutional layer, maps the multi-channel feature map back to a single channel, resulting in a denoised spectrum that is completely consistent with the input dimension and has a significantly improved signal-to-noise ratio: ; in, represents a one-dimensional convolution operation with a 1x1 kernel for channel compression of the multi-channel feature map to a single-channel denoised spectrum; The final output of the decoder is layer; The U-Net convolutional neural network adopts the structure of encoding down-sampling to extract spectral features, decoding up-sampling to reconstruct the spectrum, and jump connection to fuse multi-layer features. As shown in the residual analysis of the denoising result, Figure 2 the overall trend of the denoised spectrum is highly consistent with the original spectrum, the peak structure is well preserved, and the high-frequency noise component is significantly weakened. The residual graph is approximately zero-mean random distribution, indicating that the model effectively removes random noise and does not introduce obvious systematic bias, verifying the superiority of the proposed method in spectral detail fidelity and noise suppression.

[0045] Step 2.2, construct a preprocessing module for maintaining the intrinsic rank structure of data under non-Gaussian, skewness, and multimodal distribution conditions; this module breaks through the limitations of traditional Z-score standardization by introducing a non-parametric distribution alignment mechanism based on quantile mapping (i.e., the limitation of traditional Z-score standardization that only aligns data mean and variance, assuming that data follows Gaussian distribution).

[0046] This module mainly includes a distribution feature analysis unit and a non-parametric distribution alignment unit; the distribution feature analysis unit is used to identify and characterize the statistical distribution characteristics of different data sources, including skewness, multimodality, and distribution shape differences; the non-parametric distribution alignment unit uses a distribution alignment strategy based on quantile mapping to map different modal data to a unified target distribution space by constructing an empirical cumulative distribution function, achieving overall distribution level standardization and consistency.

[0047] The specific working process of the preprocessing module is as follows: Step 2.2.1, based on the distribution feature analysis unit, estimate the empirical distribution, calculate the empirical cumulative distribution function of each spectral feature and environmental parameter, which is used to represent the quantile position of the input feature value in the sample distribution, and the specific definition is as follows: ; wherein, is the empirical cumulative distribution function; represents the total number of samples; is the threshold value of the feature value; represents the feature value of the th sample; is an indicator function, which takes 1 when is less than or equal to , otherwise 0.

[0048] Step 2.2.2, quantile mapping is performed on the non-parametric distribution alignment unit, and the quantile value of the high-fidelity spectrum is mapped through the inverse cumulative distribution function of the target distribution to obtain a standardized result : ; is the inverse cumulative distribution function of the standard normal distribution. This method does not make distribution assumptions on the original data distribution, but directly maps the quantile information of each feature in the training set to a unified target distribution through an empirical cumulative distribution function, providing a reliable input basis for subsequent autoencoder dimension reduction and double-branch fusion modeling.

[0049] Step 2.3, for the convenience of subsequent feature extraction and fusion, the present application constructs a dimension reduction module, which is based on an autoencoder structure and performs nonlinear compression representation on high-dimensional spectral data, overcoming the limitations of traditional principal component analysis which can only extract linear features. The compressed spectral features generated by this module not only have low redundancy, but also retain the key structural properties of the spectral sequence. The module includes an encoder submodule, a decoder submodule, and a feature reconstruction verification unit. The encoder submodule is composed of multiple groups of fully connected layers and is used to compress the features of the standardized high-dimensional spectral data, autonomously learn the main information features in the spectral signal, and generate representative low-dimensional feature vectors. The decoder submodule is symmetrical to the encoder structure and is used to reconstruct spectral data with the same dimension as the original spectrum based on the low-dimensional feature vectors. The feature reconstruction verification unit automatically updates the parameters of each layer of the encoder and decoder by minimizing the reconstruction error, thereby obtaining a stable and compact low-dimensional representation while preserving the main physical features of the spectrum.

[0050] The specific working process of the dimension reduction module is as follows: Step 2.3.1, the encoder submodule maps the standardized result to a low-dimensional latent space through an autoencoder mapping function to obtain a low-dimensional feature vector , and the mapping relationship is: ; wherein, represents the autoencoder mapping function; is a set of trainable parameters; is the compressed spectral dimension.

[0051] Step 2.3.2, the decoder submodule reconstructs the low-dimensional feature vector back to the original space through the inverse mapping function of the decoder, which is used to maintain data consistency: ;​​ in, Represents the inverse mapping function of the decoder, ensuring The features retained in the original spectrum are sufficient to reflect key physical information such as the position, shape, and full width at half maximum of the absorption peaks; This is the set of trainable parameters for the decoder; This is the result of the reconstruction.

[0052] Step 2.3.3: The goal of the feature reconstruction verification unit is to minimize the mean square error (MSE) between the input and the reconstructed spectrum. By minimizing the loss function, the network can learn the nonlinear feature structure of the spectral data. Loss function It can be represented as: ; in, The total number of samples; and They represent the first Standardization and reconstruction results for each sample; Let L2 norm represent the difference between the input data and the reconstructed data. Minimizing the reconstruction loss helps the encoder preserve the local correlations and global structure of the spectral signal during dimensionality reduction, resulting in a lower-dimensional feature vector. It retains the topological properties of the spectral sequence, thus enabling it to serve as the convolutional input for subsequent environment-adaptive residual blocks (EAR-Block).

[0053] Step 3: Construct a self-supervised feature extraction network with adaptive modulation of environmental parameters to achieve deep fusion of spectral and environmental information and obtain fused feature vectors; The self-supervised feature extraction network includes an environment modulation branch, a spectral backbone network, and a feature fusion module. The environment modulation branch further includes a multimodal environment parameter self-attention representation submodule, an environment feature deep reconstruction submodule, and an adaptive control unit. The specific working process of the self-supervised feature extraction network is as follows: Step 3.1: The environment modulation branch constructs a deep fusion mechanism between multimodal environmental parameters and spectral features. By introducing a multi-head self-attention mechanism and a gating control structure, a unified framework for self-learning and dynamic allocation of environmental features is formed, which is used to realize the dynamic modulation and adaptive compensation of environmental parameters in the spectral feature extraction process. The specific working process of the environment modulation branch is as follows: Step 3.1.1: Construct a multi-modal environmental parameter self-attention representation submodule based on the multi-head self-attention mechanism. By calculating the correlation and interaction between environmental parameters, generate a high-dimensional feature embedding that can reflect the importance distribution of each parameter under different working conditions, providing an accurate and robust environmental representation basis for subsequent modulation. The application introduces a multi-head self-attention mechanism to realize dynamic correlation modeling between environment parameters, which learns the correlation between parameters and assigns context-related weights to each environment parameter, thereby obtaining embedding representation with global dependency features. The query matrix, key matrix and value matrix of the environment parameter data normalized by quantile mapping are calculated, the similarity between parameters is calculated by scaled dot-product attention, and the environment feature representation is obtained by linear transformation after splicing the multi-head attention results ; The environment feature representation explicitly encodes the interaction and relative importance between each environment parameter, thereby providing more accurate environment information representation for subsequent spectral signal modulation and concentration inversion. ; wherein, is a Softmax function; , , is the query vector, key vector and value vector corresponding to each attention head; is the transpose; is the dimension of the key vector; Step 3.1.2, an environment feature deep reconstruction submodule is constructed using a multi-layer residual network structure to perform nonlinear mapping and high-order reconstruction on the environment feature, thereby strengthening the expression ability and stability of the feature and realizing the structured transformation from static parameter representation to dynamic modulation signal, laying a foundation for the environment adaptive regulation of the model; as the input of the environment feature deep reconstruction submodule, and let the initial feature vector be . The submodule is composed of multiple residual blocks. For the th residual block, the residual information between the input and the target feature is learned through nonlinear mapping while the input feature information is preserved, thereby realizing feature enhancement and maintaining the gradient stability in the deep network, which is calculated as follows: ; wherein, is the feature vector output by the th residual block, which is used as the input of the th residual block; is the feature vector output by the th residual block; is a ReLU activation function; is the th residual block, each residual block is composed of a nonlinear transformation subnetwork composed of two fully connected layers and a ReLU activation function, which is used to perform nonlinear reconstruction on the environment feature; is the total number of residual blocks. After step-by-step feature extraction and superposition of layers of residual blocks, a high-level environment feature vector is obtained​ : ; wherein, is the eigenvector output by the th residual block; The high-level environmental eigenvector further characterizes the deep dependency relationship between different environmental parameters, providing a structured environmental characterization input for the subsequent spectral signal modulation and concentration inversion module.

[0054] Step 3.1.3, constructing an adaptive regulation unit consisting of multiple fully connected layers to generate scaling factors, offset factors, and gating signals and other modulation parameters. Through a continuous differentiable gating mechanism, this unit realizes dynamic adjustment of the spectral feature extraction process, ensuring the flexibility of feature fusion and the effectiveness of environmental compensation.

[0055] For the th environmental adaptive residual block in the spectral backbone network that needs to be modulated, the adaptive regulation unit generates the corresponding environmental modulation parameters through the high-level environmental eigenvector The process includes the following steps: Step 3.1.3.1, input the high-level environmental eigenvector into three independent fully connected layers respectively to generate preliminary environmental modulation parameters and gating signals. The preliminary environmental modulation parameters include preliminary scaling factors and preliminary offset factors.

[0056] Step 3.1.3.2, based on the gating signal, the preliminary environmental modulation parameters are weighted and fused to obtain the final environmental modulation parameters, which are calculated as follows: ; ; wherein, is the gating signal of the th environmental adaptive residual block; , are the preliminary scaling factor and the preliminary offset factor of the th environmental adaptive residual block, respectively; , are the final scaling factor and the final offset factor of the th environmental adaptive residual block, respectively; denotes element-wise multiplication; is a unit vector.

[0057] Step 3.1.3.3, apply the above obtained environmental modulation parameters to the feature mapping of the convolution block to realize adaptive affine transformation. When tends to 0, tends to 1, When the value approaches 0, modulation is turned off, and the feature map retains its original state; when... When it approaches 1, and They approach their initial values ​​respectively and That is, modulation is fully turned on.

[0058] Step 3.2: Construct a spectral backbone network consisting of multiple cascaded and stacked environment-adaptive residual blocks (EAR-Blocks). Perform step-by-step environment-adaptive feature extraction on the spectral data to obtain environment-adaptive spectral feature vectors. The spectral backbone network incorporates an environment-adaptive modulation mechanism, achieving deep fusion and dynamic compensation of spectral and environmental information. This is a key technological innovation of this invention in improving model robustness and inversion accuracy. In this invention, [the network is constructed using...] EAR-Blocks are stacked in series to form a deep spectral backbone network. Each EAR-Block adopts a residual structure, consisting of a main path and a shortcut path. As the basic unit of the spectral backbone network, the EAR-Block achieves dynamic calibration of spectral features by embedding environmental modulation mechanisms into the deep residual structure. The main path extracts spectral features step-by-step through two one-dimensional convolutions and batch normalization. After the first batch normalization output, a modulation factor generated by environmental parameters is introduced to perform a channel-by-channel affine transformation on the features, achieving environmentally adaptive adjustment of spectral features. The main path first performs the first convolution and batch normalization, outputting a feature map modulated by the environment in the first stage. Then, a second convolution transformation is performed on the already modulated feature map to learn more complex feature combination relationships. The calculation process of each EAR-Block is as follows: First, calculate the main path: ; ; in, This is the output feature map for the first stage; This is the first batch of standardized operations; This represents a one-dimensional convolution operation used to extract the first-stage features of the main path; This is the output of the previous environment-adaptive residual block, where the input of the first environment-adaptive residual block is... Low-dimensional feature vectors ,Right now . The final output of the main path; This is the second batch of standardized operations; This represents a one-dimensional convolution operation used for further feature mapping of the first-stage features.

[0059] The shortcut path is responsible for ensuring stable transmission of information and fusing with the output of the main path, and is calculated as follows: ; wherein, is the output of the shortcut path; represents a one-dimensional convolution operation with a 1x1 convolution kernel, which is used to adjust the channel number or sequence dimension in the shortcut path, so that it is consistent with the dimension of the final output of the main path ; is the dimension of the feature map; After element-wise addition of the output of the main path and the output of the shortcut path, the robust spectral feature map after environment compensation is obtained through ReLU activation, realizing direct connection of information and stable gradient; ; is the output of a single EAR-Block, which is the intermediate state of the feature after one environment adaptive transformation in the deep network. The first EAR-Block receives the low-dimensional feature vector after dimension reduction of the autoencoder as its initial input , and is modulated by the first group of parameters generated by the environment modulation branch , to obtain the spectral feature map output by the first environment adaptive residual block ; wherein, , are the final scaling factor and the final offset factor of the first environment adaptive residual block, respectively. The second EAR-Block receives the output of the previous environment adaptive residual block as its input, and is modulated by the second group of parameters , to obtain the spectral feature map output by the second environment adaptive residual block . In this way, information propagates along the network level by level. For the th EAR-Block, there are: ; wherein, is the spectral feature map output by the th environment adaptive residual block; is the th environment adaptive residual block; is the spectral feature map output by the th environment adaptive residual block; After continuous processing through EAR-Blocks, the final spectral feature map is obtained, which contains the deepest level of environment adaptive spectral information.

[0060] To convert the aforementioned one-dimensional feature map into a fixed-length one-dimensional feature vector for easier subsequent feature fusion, a global average pooling operation is used to achieve the conversion, resulting in the final environment-adaptive spectral feature vector: ; in, This is an environmentally adaptive spectral feature vector; This is a global average pooling operation. This structure ensures stable gradient propagation while enabling dynamic control and enhancement of spectral features through environmental modulation.

[0061] Step 3.3: Construct a feature fusion module to perform feature fusion, specifically by fusing the environmental adaptive spectral feature vector. High-level environmental feature vector The features are concatenated to obtain a fused feature vector. : ; in, This refers to the vector concatenation operation along the feature dimension; This dual-fusion architecture of implicit modulation and explicit splicing ensures that the output features are highly distinguishable from concentration information and environmental interference, providing a more robust feature foundation for subsequent prediction tasks.

[0062] Step 4: To enable the deep feature extraction network to learn the intrinsic physical laws and environmental influences of the spectrum without manually labeled data, this invention constructs a self-supervised auxiliary decoding module. This module utilizes a self-supervised learning mechanism to improve the feature representation ability and generalization performance of the self-supervised feature extraction network, and can be used to fully mine unlabeled data. This module mainly consists of an auxiliary decoder, a self-supervised constraint unit, and a feature optimization and verification unit. The specific working process is as follows: Step 4.1: Fuse the feature vectors The input auxiliary decoder extracts physically consistent deep feature representations. The auxiliary decoder reconstructs the spectral data through a multi-layer fully connected structure, remapping the highly compressed features back to the original spectral space, and finally generating the reconstructed spectrum. Step 4.2: To drive the parameter learning of the network, define the reconstruction loss function in the self-supervised constraint unit. The difference between the original spectrum and the reconstructed spectrum is measured to guide the parameter optimization of the feature extraction network; the calculation is as follows: ; in, For the first The original spectra of each sample; For the first reconstructed spectrum of the sample; the number of original spectra corresponds to the number of reconstructed spectra; is the total number of samples, i.e., the total number of spectra; Step 4.3, constructing a feature optimization verification unit for continuously monitoring the reconstruction performance and adjusting the structural parameters of the self-supervised feature extraction network during the training phase to ensure that the model obtains stable feature expression under the condition of no labeled data.

[0063] During the training process, the reconstruction loss function As a self-supervised signal, all trainable parameters including the auxiliary decoder, feature fusion module, environment modulation branch and spectral backbone network are jointly optimized through an end-to-end backpropagation algorithm. The early stopping strategy is adopted in the training phase to determine the optimal stopping point according to the reconstruction error of the validation set to prevent overfitting. After the training is completed, the auxiliary decoder is removed and only the optimized self-supervised feature extraction network is retained, which has stronger robustness and generalization ability and can be used as a pre-training basis for the concentration prediction model.

[0064] Step 5, to realize high-precision concentration inversion under the condition of small samples, a BPBO-GRNN adaptive concentration inversion optimization model about the fusion feature vector and the mixed gas concentration is constructed. The model is a joint decoupling model that constructs a concentration inversion network with self-learning and self-adjusting ability by deeply integrating the BPBO algorithm and the GRNN model. The model is used to realize high-precision inversion of the mixed gas concentration and adaptive optimization of the model parameters, and is the core unit for completing the final concentration prediction in the present application. The model mainly consists of a GRNN concentration prediction submodule and a BPBO adaptive optimization submodule. The robustness and prediction accuracy of GRNN are improved by using BPBO global optimization; the specific process is as follows: Step 5.1, constructing a GRNN concentration prediction submodule based on the structure of the generalized regression neural network, realizing nonlinear mapping between the input fusion feature vector and the concentration value through kernel density estimation. The submodule uses a Gaussian kernel function to model the feature similarity and uses a smoothing factor to control the kernel bandwidth, thereby obtaining a concentration prediction result that is sensitive to both spectral and environmental feature changes. GRNN is a non-parametric regression model based on kernel density estimation, and its output is obtained by weighted average of the concentration labels of all training samples. The fusion feature vector of the i-th training sample is defined as For an arbitrary input fusion feature vector, The similarity of the i-th training sample is measured by the Euclidean distance, The fusion feature vector of the i-th training sample is and is converted into kernel weight through a Gaussian kernel function: ;​​ where, is the kernel weight between and ; is the exponential function; is the Euclidean distance function; is the smoothing factor, which controls the decay rate of the similarity between samples. The higher the similarity, the greater the weight of the label in the prediction result.

[0065] The concentration prediction value of the GRNN concentration prediction submodule ; where, is the total number of training samples; is the true concentration label of the th training sample. As can be seen from the formula, the performance of GRNN mainly depends on the value of the smoothing factor . When is too small, the kernel function bandwidth tends to be narrow, which is prone to overfitting; when is too large, the kernel function bandwidth is too wide, which is prone to underfitting.

[0066] Step 5.2, build a BPBO global adaptive optimization submodule, which proposes an adaptive optimization mechanism of generalized regression neural network based on the raptor optimization algorithm, realizes intelligent optimization and global optimal configuration of the concentration inversion model parameters, and is one of the key innovations of the present application. Based on the swarm intelligence search mechanism of the raptor optimization algorithm, the GRNN smoothing factor value is dynamically adjusted through the strategy of combining global exploration and local development, avoiding the low efficiency and local optimal problem caused by traditional manual parameter adjustment or grid search, and realizing adaptive optimal configuration of the model performance.

[0067] For the problem of relying on experience and being difficult to globally optimize , the present application introduces the BPBO algorithm for adaptive optimization. The algorithm simulates the global exploration and local development behavior of raptors in the process of predation, dynamically searches for the smoothing factor, and verifies that the root mean square error of the set is the minimum optimization target. First, the optimization process of is modeled as a minimization problem, and the smoothing factor is updated as the individual position of the raptor optimization algorithm. The concentration-labeled data used in the present application is collected by a controllable gas mixing system independently built by the present application. The system can accurately control and record the true concentration of each sample during gas mixing, so each collected spectrum corresponds to a certain concentration label. The data set with concentration labels is divided into a training set and a validation set. The fitness function is defined as the root mean square error (RMSE) of the GRNN model on the validation set. For any given value, fitness is calculated as follows: ; wherein, is the number of samples in the validation set; is the th validation sample; is the th validation sample; is the true concentration label of the th validation sample;

[0068] The BPBO algorithm simulates the exploration and exploitation behavior of raptors in the process of predation to find the optimal smoothing factor. The population size and the maximum number of iterations are set; the search space of the smoothing factor is set to the preset range , wherein , are the minimum value and the maximum value of , respectively; the positions of raptor individuals are randomly initialized in the search space to form an initial population. The fitness value of each individual is calculated, and the position of the individual with the minimum fitness value is recorded as the current global optimal position . In each iteration of the raptor optimization algorithm, a random number in the interval [0, 1] is generated, and a threshold is preset as the switching threshold between exploration and exploitation stages; if the updated individual position exceeds the search space, it is truncated and mapped back within the search space range. The iterative update process of the BPBO algorithm is an adaptive and phased search strategy, and its core lies in dynamically balancing global exploration and local exploitation. In each iteration, the algorithm first generates a random number and compares it with the threshold to determine whether to execute the global exploration stage or the local exploitation stage. If , the algorithm will enter the global exploration stage, which simulates the raptors searching for prey in a wide range in the high sky, aiming to jump out of the local optimum and explore a wider search space. The individual position update formula is as follows: ; wherein, is the position of the th individual in the th iteration, corresponding to the smoothing factor value in the th iteration; is the updated position; is a random number in the range of [-1, 1].

[0069] If , the algorithm enters the local development stage, which simulates the diving and accurate pursuit of the raptor after finding the prey. This stage contains two sub-stages, determined by the relationship between the iteration number and . When , the algorithm executes the diving pursuit strategy, which quickly converges to the promising solution region by guiding the individuals to move towards the center area of the population. The diving pursuit strategy formula is as follows: ; where is the average value of all individual positions at the th iteration; is a random number in the range of [0, 1]. This strategy guides the population to gather in the center area of the current solution. When , the algorithm executes the accurate attack strategy, which simulates the raptor locking a single prey for accurate attack, increasing the randomness and diversity of search using Levy flight. The accurate attack strategy formula is as follows: ; where is the Levy flight function, and the input dimension is 1 since the optimization variable has only one smoothing factor. After each position update, the fitness value is recalculated, and if , the individual position is updated.

[0070] When the iteration number reaches , the final global optimal position is output as the optimal smoothing factor.

[0071] After obtaining the optimal smoothing factor , the BPBO-GRNN adaptive concentration inversion optimization model predicts the concentration of any input fused feature vector based on , obtaining the optimal concentration prediction value : ; By introducing the global optimal smoothing factor obtained by BPBO, the model can use the optimized Gaussian kernel function to obtain the concentration inversion result of the test sample, thereby realizing adaptive concentration prediction under the condition of high precision and small sample.

[0072] Step 6, collect the unknown concentration absorption spectrum signal of the gas to be measured under the actual working condition of the power plant, and collect the actual environmental parameter data to construct a multi-source data set; import the multi-source data set into the environment-spectrum collaborative fusion decoupling model trained by the laboratory calibration data to predict the concentration.

[0073] To verify the concentration inversion performance of the method of the application, typical gases represented by NH3 and CO were taken as examples, and the gradient comparison results of the real concentration and the predicted concentration are shown in Figure 3 and Figure 4 , and the residual distribution is shown in Figure 5 . The results show that the method of the application has high prediction accuracy under different concentration gradients, and the determination coefficients R² of NH3 and CO are 0.9994 and 0.9992 respectively, and the root mean square errors RMSE are 3.05 and 2.36 respectively, indicating that the method of the application can still maintain stable concentration inversion performance under the condition of temperature and pressure coupling interference.

[0074] To verify the effectiveness of the self-supervised feature extraction network in the application in environmental interference compensation, and further illustrate its technical effects in improving concentration inversion accuracy and robustness, a comparative ablation experiment was conducted. In the experiment, the original spectrum signal after data preprocessing in step 2 was directly input into the BPBO-GRNN adaptive concentration inversion optimization model in step 5 (i.e. without steps 3 and 4), and the test was conducted under the same temperature (100℃-250℃) and pressure (0.8-2.5 atm) conditions. The gradient comparison results of the real concentration and the predicted concentration of CO and NH3 without the compensation of the self-supervised feature extraction network are shown in Figure 6 and Figure 7 , and the residual distribution is shown in Figure 8 .

[0075] The comparison results show that the model without the compensation of the self-supervised feature extraction network has obvious deviation in the concentration prediction of NH3 and CO, and the root mean square errors RMSE are 70.22 and 36.12 respectively, and there is significant deviation between the predicted concentration and the real concentration, which makes it difficult to realize accurate inversion of the mixed gas. However, the model of the application with the introduction of the self-supervised feature extraction network has a significant reduction in the root mean square error RMSE to 3.05 and 2.36 under the same conditions, as shown in Figure 5 , and the prediction determination coefficient R² is improved to more than 0.999, and the prediction accuracy is greatly improved. The results show that the accuracy of concentration inversion is significantly improved in the model with the introduction of the self-supervised feature extraction network, and the root mean square error RMSE is reduced by about 91.6% (NH3) and 96.6% (CO), which verifies the effective compensation of the network to the spectral drift and feature aliasing problems caused by temperature and pressure coupling. The stability and robustness of the prediction are enhanced, and the performance of the model under complex working conditions is more consistent, and the error fluctuation is controlled within ±0.5% under different experimental conditions.

[0076] In summary, the spectral aliasing decoupling and concentration inversion method based on self-supervised feature extraction and BPBO-GRNN joint optimization has significant innovation and practical value in realizing adaptive compensation of spectral nonlinear deformation and dynamic modeling of features in complex industrial environments. The method breaks through the technical bottleneck of traditional models under temperature and pressure coupling conditions, such as easy loss of accuracy and insufficient anti-interference ability, and realizes high-precision, strong-robustness inversion and stable detection of multi-component gases. The present application has good system compatibility and can be embedded in a continuous emission monitoring system (CEMS) to realize high-precision concentration monitoring of aliasing gases such as CO and NH3 during the emission process of a thermal power plant, providing original and valuable technical support for industrial emission control and environmental supervision under complex conditions.

[0077] Of course, the above description is not a limitation of the present application, and the present application is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the spirit and scope of the present application should also be within the scope of the present application.

Claims

1. A method for spectral unmixing and concentration retrieval under the cross-influence of multiple environmental factors, characterized in that, The application discloses a method for predicting concentration of mixed gas based on environment-spectrum collaborative fusion decoupling model. Step 1: Collecting absorption spectrum signals of specified mixed gas under different temperatures, pressures and known concentrations, and collecting environmental parameter data to construct a multi-source data set; Step 2: Denoising, dimensionality reduction and preprocessing the multi-source data set through a multi-source data processing module; Step 3: Constructing an environment parameter self-adaptive modulation self-supervised feature extraction network to realize deep fusion of spectrum and environmental information and obtain a fusion feature vector; Step 4: Constructing a self-supervised auxiliary decoding module to improve feature expression ability and generalization performance of the self-supervised feature extraction network by using a self-supervised learning mechanism and fully mining unlabeled data; Step 5: Constructing a BPBO-GRNN self-adaptive concentration inversion optimization model to realize inversion of mixed gas concentration and self-adaptive optimization of model parameters; Step 6: Collecting absorption spectrum signals of the mixed gas under unknown concentrations in actual working conditions of a thermal power plant, collecting actual environmental parameter data to construct a multi-source data set, and inputting the multi-source data set into the environment-spectrum collaborative fusion decoupling model trained by laboratory calibration data to predict the concentration. The specific process of step 1 is as follows: first, collecting absorption spectrum signals of CO and NH3 mixed gas under different temperature and pressure conditions and recording corresponding environmental parameter data to construct a multi-source data set; and then collecting mixed spectrum signals under unknown concentration conditions in the same temperature and pressure control range as the model verification set.

2. The method of claim 1, wherein, In step 2, the data processing module includes an intelligent denoising module, a preprocessing module and a dimensionality reduction module. Step 2.1: Constructing an intelligent denoising module, which adopts a pre-trained one-dimensional U-Net convolutional neural network structure; inputting the multi-source data set into the one-dimensional U-Net convolutional neural network for denoising, and the mapping relationship is as follows:

3. The method of claim 1, wherein, Step 2.2: Introducing a quantile mapping non-parametric distribution alignment mechanism to construct a preprocessing module; the preprocessing module includes a distribution feature analysis unit and a non-parametric distribution alignment unit, and the specific working process is as follows: Step 2.2.1: Estimating an empirical distribution based on the distribution feature analysis unit, and calculating an empirical cumulative distribution function of each spectrum feature and environmental parameter: ； wherein, is a high-fidelity spectrum of the th sample; is an original spectrum of the th sample; are trainable parameters of a one-dimensional U-Net convolutional neural network; denotes a function mapping corresponding to the one-dimensional U-Net convolutional neural network structure. Step 2.3: Constructing a dimensionality reduction module. In step 2.3, the dimensionality reduction module includes an encoder submodule, a decoder submodule and a feature reconstruction verification unit, and the specific working process is as follows: ； in, The empirical cumulative distribution function; Represents the total number of samples; The threshold value is the feature value. Indicates the first Feature values ​​of each sample; For indicator functions, when Less than or equal to The value is 1 if the condition is met, otherwise it is 0. Step 2.2.

2. Quantile mapping based on non-parametric distribution alignment unit, mapping the quantile values of high-fidelity spectra through the inverse cumulative distribution function of the target distribution to obtain the normalized results :​ ； wherein is the inverse cumulative distribution function of the standard normal distribution; Step 2.3.2: The decoder submodule reconstructs the low-dimensional feature vector back to the original space through an inverse mapping function of the decoder:

4. The method of claim 3, wherein, In step 3, the self-supervised feature extraction network includes an environment modulation branch, a spectrum backbone network and a feature fusion module. Step 2.3.1, the encoder sub-module maps the standardized results to a low-dimensional latent space via an autoencoder mapping function to obtain a low-dimensional feature vector : ； wherein, represents an autoencoder mapping function; is a set of trainable parameters; is the compressed spectral dimension; The specific process is as follows: ； wherein, denotes an inverse mapping function of the decoder; is a set of trainable parameters of the decoder; is a reconstruction result; Step 2.3.3, the target of the feature reconstruction verification unit is to minimize the mean square error between the input and the reconstructed spectrum, and the network can learn the nonlinear feature structure of the spectral data by minimizing the loss function; the loss function is: ； wherein, and denote the standardized result and the reconstructed result of the th sample, respectively; denotes the L2 norm.

5. The method of claim 4, wherein, ​ ​ Step 3.1, constructing environment modulation branch to obtain high-level environment feature vector And introduce multi-head self-attention mechanism and gating control structure; Step 3.2, constructing a spectral backbone network composed of a plurality of environment-adaptive residual blocks stacked in series, performing environment-adaptive feature extraction on the spectral data level by level to obtain an environment-adaptive spectral feature vector ; Step 3.3, constructing a feature fusion module to perform feature fusion, specifically, performing splicing on the environment adaptive spectral feature vector and the high-layer environment feature vector to obtain a fused feature vector : ； wherein, is a vector concatenation operation in the feature dimension.

6. The method of claim 5, wherein, In step 3.1, the environment modulation branch further includes a multi-modal environment parameter self-attention representation submodule, an environment feature deep reconstruction submodule, and an adaptive regulation unit, and the specific working process is as follows: Step 3.1.

1. Constructing a multi-modal environmental parameter self-attention representation submodule based on a multi-head self-attention mechanism, for obtaining an environmental feature representation : ； wherein, is a Softmax function; , , are query vector, key vector and value vector corresponding to each attention head, respectively; is a transpose; is a dimension of the key vector; Step 3.1.2, constructing the environmental feature deep reconstruction sub-module using a multi-layer residual network structure; and inputting the environmental feature deep reconstruction sub-module with the initial feature vector , and taking the initial feature vector as the input of the environmental feature deep reconstruction sub-module ; the sub-module is composed of a plurality of residual block stacks; for the i-th residual block, the calculation is as follows: ​ ； wherein, is the feature vector output by the th residual block, which is input to the th residual block; is the feature vector output by the th residual block; is a ReLU activation function; is the th residual block; is the total number of residual blocks; through the step-by-step feature extraction and superposition of the layers of residual blocks, a high-level environmental feature vector is obtained. ； wherein, is the feature vector output for the i-th residual block; Step 3.1.3: Construct an adaptive control unit composed of multiple fully connected layers. For the first layer in the spectral backbone network... An environment adaptive residual block that needs to be modulated; the adaptive control unit uses high-level environmental feature vectors. Generate the corresponding environmental modulation parameters; specifically as follows: First, the high-level environment feature vector is input to three independent fully connected layers respectively to generate a preliminary environment modulation parameter and a gating signal, the preliminary environment modulation parameter including a preliminary scaling factor and a preliminary offset factor; Then, based on the gating signal, the preliminary environment modulation parameters are weighted and fused to obtain the final environment modulation parameters: ； ； wherein, is a gating signal for the th environment-adaptive residual block; , are a preliminary scaling factor and a preliminary offset factor, respectively, for the th environment-adaptive residual block; , are a final scaling factor and a final offset factor, respectively, for the th environment-adaptive residual block; denotes multiplication of the corresponding elements; is a unit vector; Finally, the obtained environment modulation parameters are applied to the feature mapping of the convolution block to realize adaptive affine transformation.

7. The method of claim 6, wherein, In step 3.2, each environment adaptive residual block adopts a residual structure composed of a main path and a shortcut path; the calculation formula of the main path is: ； ； wherein, is the output feature map of the first stage; is the first batch normalization operation; denotes a one-dimensional convolution operation; is the output of the previous environment-adaptive residual block; is the final output of the main path; is the second batch normalization operation; denotes a one-dimensional convolution operation; The calculation formula of the shortcut path is: ； wherein, is the output of the shortcut path; represents a one-dimensional convolution operation with a 1x1 convolution kernel; is the dimension of the feature map; The output of the main path and the output of the shortcut path are added element by element to obtain a spectral feature map : ； is the output of the single environment-adaptive residual block for the i-th environment-adaptive residual block: ； wherein, is the spectral feature map output by the th environment-adaptive residual block; is the spectral feature map output by the th environment-adaptive residual block; is the spectral feature map output by the th environment-adaptive residual block; After successive processing via an environment-adaptive residual block, a final spectral feature map is obtained ; The global average pooling operation is adopted for conversion to obtain a final environment adaptive spectral feature vector : ； wherein, is a global average pooling operation.

8. The method of claim 7, wherein, In step 4, the self-supervised auxiliary decoding module includes an auxiliary decoder, a self-supervised constraint unit, and a feature optimization verification unit, and the specific working process is as follows: Step 4.1, fusing feature vectors The auxiliary decoder extracts deep feature representation with physical consistency, and reconstructs the spectral data through a multi-layer fully connected structure, re-maps the highly compressed features to the original spectral space, and finally generates the reconstructed spectrum. Step 4.

2. Defining the reconstruction loss function in the self-supervised constraint unit Measuring the difference between the original and reconstructed spectra: ； wherein, is the original spectrum of the th sample; is the reconstructed spectrum of the th sample; the number of original spectra and reconstructed spectra corresponds. Step 4.3, construct a feature optimization verification unit for continuously monitoring the reconstruction performance and adjusting the structure parameters of the self-supervised feature extraction network during the training stage to ensure that the model obtains stable feature expression under the condition of no labeled data; During the training process, the reconstruction loss function As a self-supervised signal, all trainable parameters including auxiliary decoder, feature fusion module, environment modulation branch and spectral backbone network are jointly optimized by end-to-end backpropagation algorithm; Early stopping strategy is adopted in the training stage to determine the optimal stopping point according to the reconstruction error of the validation set to prevent overfitting; After training, the auxiliary decoder is removed and only the optimized self-supervised feature extraction network is retained.

9. The method of claim 8, wherein, In step 5, the BPBO-GRNN adaptive concentration inversion optimization model includes a GRNN concentration prediction submodule and a BPBO adaptive optimization submodule, and the specific process is as follows: Step 5.1: Construct a GRNN concentration prediction submodule based on the generalized regression neural network structure. Use a Gaussian kernel function to weight the feature similarity model and a smoothing factor to control the kernel bandwidth, thereby obtaining concentration prediction results that are sensitive to changes in both spectral and environmental characteristics; define the... The fused feature vector of each training sample Given a fused feature vector of any input, for With the training samples The similarity is measured by Euclidean distance. For the first The fused feature vectors of the training samples are transformed into kernel weights using a Gaussian kernel function: ； wherein, is and between the kernel weights; is an exponential function; is an Euclidean distance function; is a smoothing factor; Concentration prediction value is: ； wherein, is the total number of training samples; is the true concentration label of the th training sample; Step 5.2, construct a BPBO global adaptive optimization submodule, which proposes an adaptive optimization mechanism of generalized regression neural network based on the raptor optimization algorithm, dynamically searches for the optimal smoothing factor by simulating the global exploration and local development behavior of raptors in the process of predation.

10. The method of claim 9, wherein, The specific process of step 5.2 is as follows: First, the optimization process of smoothing factor is modeled as a minimization problem, smoothing factor is updated as the individual position of the raptor optimization algorithm, and the optimal smoothing factor is found by the raptor optimization algorithm; for any given value, the fitness is calculated as follows: ； in, The number of samples in the validation set; For the first One verification sample; For the first The true concentration label of the validation sample; Indicates the use of the current The constructed GRNN concentration prediction submodule predicts the concentration values ​​of the validation samples; Setting population size With the maximum number of iterations ; Set smoothing factor The search space is a preset range , wherein , The minimum value and the maximum value of , respectively; within the search space, the positions of raptors are randomly initialized to form an initial population; the fitness value of each individual is calculated, and the position of the individual with the minimum fitness value is recorded as the current global optimal position ; In each iteration of the algorithm, a random number in the interval [0, 1] is generated , a threshold value is preset as the switching threshold between the exploration and development stages; if the updated individual position is outside the search space, it is truncated and mapped back within the search space range; the specific iteration process of the algorithm is as follows: If , the algorithm will enter the global exploration stage; the individual position update formula is as follows: ； wherein, is the position of the i-th individual at the j-th iteration, is the position of the i-th individual at the j-th iteration, is the position of the i-th individual at the j-th iteration, is the position of the i-th individual at the j-th iteration, is the updated position of the i-th individual, is a random number in the range [-1, 1]. If , the algorithm enters a local development phase; this phase contains two sub-phases, determined by the relationship between the iteration number and ; when , a dive-and-chase strategy is executed: ； wherein, is the first iteration of the average of all individual positions; is a random number in [0, 1]. When the precise attack strategy is executed: ； wherein is the Levy flight function; After each position update, the fitness value is recalculated, and if the individual position is updated; When the number of iterations reaches the final global optimal position as the optimal smoothing factor; After obtaining the optimal smoothing factor Then, the BPBO-GRNN adaptive concentration inversion optimization model is established based on The concentration of the fusion feature vector is predicted for any input, and the optimal concentration prediction value is obtained : 。

Images