Infrared Gas Analysis Method and Device Based on Multi-Sensor Fusion

Abstract

The present invention relates to the technical field of infrared gas analysis, and discloses an infrared gas analysis method and device based on multi-sensor fusion. The method includes: performing adaptive digital filtering and multi-scale wavelet denoising operations on the multi-component gas spectral absorption data collected by a multi-band infrared sensor array to obtain a corrected multi-dimensional spectral matrix; performing partition cross-validation compensation operations to obtain a standard spectral matrix; constructing a covariance matrix, and extracting principal component features and latent variable features through eigenvalue decomposition to obtain a spectral feature vector; inputting the spectral feature vector into a deep auto-encoding network model for non-linear mapping and residual reconstruction calculations to obtain a compensated feature mapping vector; performing information fusion to obtain a fusion decision vector; respectively inputting the fusion decision vector into a Gaussian kernel support vector classifier and a partial least squares regression model, and outputting the type identification and concentration value of the target gas. This method realizes high-precision identification and quantitative analysis of gas components.

Description

Technical Field

[0001] The present invention relates to the field of infrared gas analysis technology, and in particular to a multi-sensor fusion infrared gas analysis method and device. Background Art

[0002] The traditional single infrared sensor gas analysis method is easily affected by cross-interference and environmental noise in complex environments, resulting in unstable measurement results and making it difficult to meet the needs of industrial sites for high-precision detection of gas components.

[0003] Although the application of multi-band infrared sensor arrays can obtain richer gas spectral feature information, how to effectively deal with the data fusion problem of multiple sensors and how to eliminate the interference of environmental factors such as temperature and pressure are still technical problems that need to be solved in this field. At the same time, the synchronous acquisition and real-time processing of multi-sensor data also put forward higher requirements on the computational efficiency of the system. In addition, in practical applications, the spectral characteristics of gas components often show strong nonlinear and coupling characteristics. Traditional linear analysis methods are difficult to accurately extract and characterize these characteristics. Although the feature extraction method based on deep learning has strong nonlinear mapping capabilities, its network structure design and parameter optimization lack specificity and it is difficult to fully utilize the physical characteristics of spectral data. Summary of the invention

[0004] The present invention provides a multi-sensor fusion infrared gas analysis method and device, which realizes high-precision identification and quantitative analysis of gas components.

[0005] In a first aspect, the present invention provides a multi-sensor fusion infrared gas analysis method, the multi-sensor fusion infrared gas analysis method comprising:

[0006] Adaptive digital filtering and multi-scale wavelet denoising are performed on the multi-component gas spectral absorption data collected by the multi-band infrared sensor array to obtain a corrected multi-dimensional spectral matrix.

[0007] The corrected multidimensional spectrum matrix is ​​subjected to a partition cross-validation compensation operation with a segmented temperature compensation coefficient matrix and a segmented pressure compensation coefficient matrix to obtain a standard spectrum matrix;

[0008] Constructing a covariance matrix according to the standard spectral matrix, extracting principal component features and latent variable features by eigenvalue decomposition, and obtaining a spectral feature vector;

[0009] Inputting the spectral feature vector into a deep autoencoder network model for nonlinear mapping and residual reconstruction calculation to obtain a compensated feature mapping vector;

[0010] Performing information fusion on the compensated feature mapping vector to obtain a fusion decision vector;

[0011] The fusion decision vector is input into a Gaussian kernel support vector classifier and a partial least squares regression model respectively, and the type identification and concentration value of the target gas are output.

[0012] In a second aspect, the present invention provides a multi-sensor fusion infrared gas analysis device, the multi-sensor fusion infrared gas analysis device comprising:

[0013] A denoising module is used to perform adaptive digital filtering and multi-scale wavelet denoising operations on the multi-component gas spectral absorption data collected by the multi-band infrared sensor array to obtain a corrected multi-dimensional spectral matrix;

[0014] A compensation module, used for performing a partition cross-validation compensation operation on the corrected multi-dimensional spectrum matrix, the segmented temperature compensation coefficient matrix and the segmented pressure compensation coefficient matrix, respectively, to obtain a standard spectrum matrix;

[0015] A construction module is used to construct a covariance matrix according to the standard spectral matrix, extract principal component features and latent variable features by eigenvalue decomposition, and obtain a spectral feature vector;

[0016] A calculation module, used for inputting the spectral feature vector into a deep autoencoder network model for nonlinear mapping and residual reconstruction calculation to obtain a compensated feature mapping vector;

[0017] A fusion module, used for performing information fusion on the compensated feature mapping vector to obtain a fusion decision vector;

[0018] The output module is used to input the fusion decision vector into the Gaussian kernel support vector classifier and the partial least squares regression model respectively, and output the type identification and concentration value of the target gas.

[0019] The third aspect of the present invention provides a multi-sensor fusion infrared gas analysis device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the multi-sensor fusion infrared gas analysis device performs the above-mentioned multi-sensor fusion infrared gas analysis method.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, in which instructions are stored, and when the computer-readable storage medium is run on a computer, the computer executes the above-mentioned multi-sensor fusion infrared gas analysis method.

[0021] In the technical solution provided by the present invention, a signal preprocessing method combining multi-scale wavelet denoising and adaptive digital filtering is constructed to effectively suppress the influence of environmental noise; a segmented temperature and pressure compensation and cross-validation optimization strategy are adopted to significantly improve the measurement stability of the system under complex working conditions; a nonlinear feature mapping structure based on a deep autoencoder network is designed, and the adaptive fusion of multi-source information is realized in combination with the improved DS evidence theory, solving the problem of feature extraction and decision optimization in multi-sensor data fusion; and a support vector classification and partial least squares regression model with mixed regularization constraints is introduced to realize high-precision identification and quantitative analysis of gas components. The present invention has strong environmental adaptability and analysis accuracy.

[0022] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0023] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of an embodiment of a multi-sensor fusion infrared gas analysis method in an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of an embodiment of an infrared gas analysis device with multi-sensor fusion in an embodiment of the present invention;

[0026] Figure 3 It is a schematic diagram of an embodiment of an infrared gas analysis device with multi-sensor fusion in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0029] To facilitate understanding of this embodiment, a multi-sensor fusion infrared gas analysis method disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps:

[0030] 101. Perform adaptive digital filtering and multi-scale wavelet denoising operations on the multi-component gas spectral absorption data collected by the multi-band infrared sensor array to obtain a corrected multi-dimensional spectral matrix;

[0031] It is understandable that the execution subject of the present invention may be a multi-sensor fusion infrared gas analysis device, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0032] Specifically, the infrared sensor array acquisition signals of the 3-5μm band, 7-9μm band and 10-12μm band are synchronously sampled at 10Hz. These bands cover the absorption characteristics of most gases. In the signal acquisition stage, the signal of each band is sampled in real time using a synchronous sampling frequency of 10Hz to obtain the spectral absorption data of multi-component gases. For the collected data, the maximum and minimum values ​​are normalized to ensure that the data of different bands are compared on the same scale, which helps to remove the influence of factors such as sensor differences and environmental changes. The normalized data are sorted into a data array according to the order of sensor bands to form an initial spectral data matrix. In order to improve the signal quality and remove noise, the initial spectral data matrix is ​​input into the db4 wavelet transform function for 3-layer wavelet decomposition. Wavelet transform can effectively decompose the signal into low-frequency approximate coefficients and high-frequency detail coefficients. Through decomposition, the main components and noise components of the signal are processed respectively. The noise variance of each high-frequency sub-band is calculated based on the wavelet decomposition coefficient matrix, and the adaptive soft threshold is set according to the size of the noise variance, and then the high-frequency detail coefficient is threshold-shrinked. The noise is suppressed by adjusting the threshold while retaining the useful information in the signal. The denoised wavelet coefficient matrix is ​​reconstructed by inverse wavelet transform to obtain the preliminary denoised spectral data and restore the real gas absorption spectrum as much as possible. After denoising, the preliminary denoised spectral data enters the baseline drift identification stage. Due to instrument deviation or environmental changes, the spectral data will have baseline drift, which affects the accurate measurement of gas concentration. The baseline drift is identified and fitted using a third-order polynomial fitting model. By calculating the difference between the fitting curve and the original spectral data, the baseline drift part is effectively subtracted from the multi-component gas spectral absorption data to obtain the baseline-corrected spectral data. The channel response characteristics of the baseline-corrected spectral data are compensated. By compensating the response characteristics of each sensor channel, the response differences between different sensors are eliminated, making the data of each band more consistent and improving the accuracy of the data. The compensated spectral data are reorganized into a multidimensional matrix, in which the rows represent the data of each sampling point and the columns represent the output signal of each sensor, forming the final corrected multidimensional spectral matrix.

[0033] 102. Performing a partition cross-validation compensation operation on the corrected multidimensional spectrum matrix, the segmented temperature compensation coefficient matrix and the segmented pressure compensation coefficient matrix, respectively, to obtain a standard spectrum matrix;

[0034] Specifically, the standard gas sample is subjected to multi-point spectral acquisition within a preset temperature range, and these temperature points are evenly distributed within the temperature range, covering possible working environment conditions. The temperature compensation coefficient corresponding to each temperature point is fitted by using the least squares method to collect data at each temperature point. The temperature compensation coefficient reflects the change in the spectral absorption characteristics of the gas under different temperature conditions, and can effectively eliminate the influence of temperature fluctuations. The temperature compensation coefficient is reorganized according to the size of the temperature interval to construct a three-dimensional array structure, which can reflect the change in the compensation coefficient under different temperature intervals, and on this basis, a segmented interpolation calculation is performed to ensure that the compensation coefficient can be accurately obtained in different temperature intervals, and a segmented temperature compensation coefficient matrix is ​​obtained. Similarly, the standard gas sample is subjected to spectral acquisition at multiple pressure points within a preset pressure range, and the pressure compensation coefficient corresponding to each pressure point is obtained by fitting the least squares method, which reflects the change in the spectrum under different pressure conditions and can eliminate the influence of pressure fluctuations on the measurement results. The pressure compensation coefficient is reorganized according to the size of the pressure interval, a three-dimensional array structure is constructed, and these coefficients are segmented interpolation calculations are performed to obtain a segmented pressure compensation coefficient matrix. The corrected multidimensional spectral matrix is ​​partitioned according to the temperature and pressure range. Each partition corresponds to a combination of temperature and pressure, representing the spectral data under specific environmental conditions. The partition data is input into the K-fold cross-validation model to verify the compensation effect of each data interval. K-fold cross-validation is an effective validation method. By dividing the data set into K subsets, K-1 subsets are used for training in turn, and the remaining subset is used for validation to obtain the compensation effect of each data partition. Through cross-validation, the effect of each partition data after compensation is evaluated to ensure the accuracy and robustness of the model. According to the deviation obtained by cross-validation, the temperature and pressure compensation coefficients are iteratively optimized to improve the compensation accuracy. The optimized compensation coefficients are used to perform compensation operations on the spectral data after preliminary compensation to obtain the optimized compensated spectral data. The optimized compensated data are reorganized and spliced, and recombined into a complete spectral matrix according to the previous partition structure. After the reorganization is completed, a normalization operation is performed to ensure that the data collected by different sensors are compared at the same scale and eliminate the systematic errors caused by sensor differences and changes in the measurement environment. Through the above processing and optimization, the final standard spectral matrix contains accurate temperature and pressure compensation data.

[0035] 103. Construct a covariance matrix based on the standard spectral matrix, extract the principal component features and latent variable features through eigenvalue decomposition, and obtain the spectral feature vector;

[0036] Specifically, the standard spectrum matrix is ​​centered. The standard spectrum matrix contains the absorption spectrum data of the gas in each band, where each column represents the data of different sensors in different bands, and the rows represent different sampling points. In order to remove the offset in the spectrum data, the spectrum data of each band is processed, and the mean of each band is subtracted from the data of each band to obtain a centralized spectrum data matrix. In this way, the overall trend in the spectrum data is removed, so that the variance of the data is more concentrated in the fluctuation part. Based on the centralized spectrum data matrix, the covariance values ​​between the bands are calculated. The covariance matrix reflects the correlation between different bands. A symmetric covariance matrix is ​​constructed through matrix operations to describe the coordinated change relationship between the bands in the data. The covariance matrix is ​​input into the eigenvalue decomposition algorithm. Eigenvalue decomposition diagonalizes the covariance matrix to obtain the eigenvalue sequence and the corresponding eigenvector matrix. The eigenvalue reflects the size of the variance in each direction of the data, while the eigenvector represents the specific distribution of these directions. By arranging the eigenvalue sequence in descending order, the direction with the largest variance is found, and the cumulative contribution rate of each eigenvalue is calculated. In order to ensure that the extracted data is representative enough, the eigenvalues ​​and corresponding eigenvectors with a cumulative contribution rate greater than 95% are selected to retain the variance information of most of the data. After screening, the eigenvector matrix obtained is the principal component feature matrix, which represents the main change direction of the data. Principal component analysis can project the original data from high-dimensional space to low-dimensional space while retaining most of the information. The standard spectral matrix is ​​input into the partial least squares decomposition (PLS) algorithm for feature extraction. The PLS method extracts potential variables from the standard spectral matrix, which reflect the variation in the spectral data that is not captured by principal component analysis. The loading matrix and the score matrix are calculated by the PLS algorithm. The loading matrix represents the weight of each spectral variable, while the score matrix represents the potential characteristics of each sample. In order to ensure that the extracted latent variables can effectively explain the variance of the spectral data, the latent variables with an explained variance greater than 90% are screened to obtain the latent variable feature matrix. The principal component feature matrix and the latent variable feature matrix are concatenated by column vectors to obtain a combined feature matrix. The column vectors of the combined feature matrix are Schmidt orthogonalized. Schmidt orthogonalization is a method of converting a set of linearly dependent vectors into a set of linearly independent vectors. For each column vector in the combined feature matrix, the difference operation is performed with the previously orthogonalized vector in turn, and the difference is normalized to ensure that each feature vector is orthogonal and standardized. This process effectively eliminates the correlation between vectors, making the final spectral feature vector more representative and independent.

[0037] 104. Input the spectral feature vector into the deep autoencoder network model for nonlinear mapping and residual reconstruction calculation to obtain a compensated feature mapping vector;

[0038] Specifically, the spectral feature vector is input into the first coding layer in the deep autoencoder network model for feature compression. The number of neurons contained in the first coding layer is twice the sum of the number of principal component features and the number of latent variable features. The input of each neuron is nonlinearly transformed by the ReLU activation function. At the same time, a batch normalization operation with a weight of 0.7 is used to stabilize the training process, and a dropout layer with a weight of 0.3 is used for feature screening to reduce the risk of overfitting. Through the processing of this layer, the first layer of encoded features obtained helps to extract important low-dimensional features in spectral data. The first layer of encoded features is input into the second coding layer for dimensionality reduction mapping. The number of neurons in the second coding layer is the sum of the number of principal component features and latent variable features. The input of this layer is processed by the ReLU activation function, and batch normalization with a weight of 0.7 is applied to ensure the stability of model training. The second coding layer introduces jump connections, that is, the original feature information in the input data is retained and transmitted together with the features processed by the activation function, thereby enhancing the ability of feature representation and avoiding the loss of key information in the dimensionality reduction process, and obtaining the second layer of encoded features. The second layer of encoded features is input into the third coding layer in the deep autoencoder network, and feature aggregation is performed at this layer. The third encoding layer contains a third neuron with 1 / 2 of the sum of the number of principal component features and the number of latent variable features. This design can further extract the core information in the spectral data. The neuron input of the third layer is processed by the ReLU activation function and a batch normalization operation with a weight of 0.7 is used. At the same time, the third layer introduces an attention mechanism, which makes the model pay more attention to the key features in the spectral data and improves the effectiveness of feature extraction. After processing, the third layer encoding features are obtained. After the encoding process is completed, the decoding process is carried out, and the third layer encoding features are input to the first decoding layer for feature reconstruction. The number of neurons in the first decoding layer is the same as that in the third encoding layer. The input of the fourth neuron is processed by the ReLU activation function and a batch normalization with a weight of 0.7 is used. In this layer, the decoding process is opposite to the encoding process. Reconstruction is performed according to the extracted features, and the features of the decoding layer are combined with the corresponding encoding layer features through residual connections to enhance the recovery ability of information and obtain the first layer decoding features. The first layer decoding features are passed to the second decoding layer for dimensional expansion. The number of neurons in the second decoding layer is the same as that in the second encoding layer, and the input of the fifth neuron is also processed by the ReLU activation function and batch normalization. The second decoding layer introduces a reverse jump connection, which can fuse the information of the encoding layer with the features of the decoding layer, help the decoding process recover more original feature information, and obtain the second layer of decoding features. The second layer of decoding features enters the third decoding layer for feature restoration. The number of neurons in the third decoding layer is the same as that in the first encoding layer. The input of the sixth neuron is operated by the Sigmoid activation function and combined with a 0.3 Dropout layer to avoid overfitting. Through the processing of this layer, the reconstructed feature vector is finally obtained, which represents the complete restoration of the spectral features.In order to optimize the feature mapping process, a loss function is constructed based on the reconstructed feature vector and the original spectral feature vector. The loss function consists of two parts: reconstruction error and feature correlation constraint. The weight of the reconstruction error is set to 0.7, and the weight of the feature correlation constraint is set to 0.3. When calculating the loss, the cosine similarity is used to measure the similarity between the original spectral feature vector and the reconstructed feature vector. In this way, the features are effectively compensated and the reconstruction error is reduced. By calculating the residual error, the compensated feature mapping vector is obtained. This vector represents the spectral data after nonlinear mapping, residual reconstruction and feature compensation, which can effectively improve the accuracy and robustness of gas analysis.

[0039] 105. Perform information fusion on the compensated feature mapping vector to obtain a fusion decision vector;

[0040] Specifically, the compensated feature mapping vectors are grouped according to the sensor bands to construct the basic probability allocation matrix of each sensor. The feature mapping vectors of each sensor band are regarded as independent information sources and constitute the basic evidence matrix of the sensor layer. In this process, the initial trust weights are set based on the signal-to-noise ratio. These trust weights can reflect the reliability of each sensor and help the subsequent evidence fusion. Sensors with higher signal-to-noise ratios will be given higher trust, thereby enhancing their influence in the fusion process. The sensor layer basic evidence matrix is ​​input into the conflict factor calculation module. In this module, the conflict between evidences is calculated using the Dempster combination rule. This rule helps identify which evidences have large differences by quantifying the consistency and conflict between evidences. For evidence with a conflict degree greater than 0.8, the penalty factor is adjusted to reduce the negative impact of conflicting evidence on the final decision. Through this adjustment, the obtained sensor layer fusion evidence will be more stable and consistent, and can better reflect the consensus information in the multi-sensor system. The sensor layer fusion evidence is feature decomposed, and the evidence of different feature subspaces is calculated by the mutual information criterion. The mutual information criterion helps evaluate the contribution of each feature subspace to the target variable, and then determine which features are more important to the decision process. By evaluating the importance of features, the trust function of the feature layer is constructed to obtain the basic evidence matrix of the feature layer. The evidence matrix of the feature layer reflects the reliability of each feature under different environmental conditions. The basic evidence matrix of the feature layer is graded according to the contribution of the features. Features with a contribution less than 0.1 are pruned to eliminate the interference of redundant and low-contribution features on the fusion process. The remaining features are reconstructed through evidence accumulation operations to obtain the update of feature support. This process effectively enhances the information support of high-contribution features and forms a feature layer fusion evidence matrix, which represents the information of all features after effective screening and reconstruction. Based on the sensor layer fusion evidence and the feature layer fusion evidence, the decision layer evidence matrix is ​​constructed. The decision layer evidence matrix integrates information from different levels to form the final decision basis. In this process, the trust threshold judgment rules are set to determine which evidence is reliable enough to affect the final decision. Through these rules, the initial decision support is obtained and optimized. According to the maximum trust principle, the evidence with a trust difference less than 0.05 is combined into the same category, thereby reducing uncertainty and improving the accuracy of decision-making, and obtaining the decision layer fusion evidence. In order to improve the accuracy and robustness of decision-making, all evidence is input into the evidence chain optimization module. This module dynamically adjusts the trust weight of each layer of evidence through the back-propagation algorithm and performs fine-tuning. Through this process, the model can better adapt to different data sources and environmental conditions to ensure the optimal effect of the information fusion process. Based on the optimized multi-layer fusion evidence, evidence comprehensive calculation is performed, and the Dempster combination rule is used to integrate multi-source information to obtain the final fusion decision vector.

[0041] 106. The fused decision vector is input into the Gaussian kernel support vector classifier and the partial least squares regression model respectively, and the type identification and concentration value of the target gas are output.

[0042] Specifically, the fusion decision vector is preprocessed by data standardization to eliminate the dimension difference between different feature dimensions, so that each feature has the same scale, and avoids some features from causing unnecessary deviations in model training due to too large or too small value range. The fusion decision vector is standardized to convert the data into a standardized feature vector. The standardized feature vector is input into the first input channel of the Gaussian kernel support vector classifier (SVC). In the Gaussian kernel support vector classifier, the low-dimensional input data is mapped to the high-dimensional feature space through feature space mapping, so that the nonlinear data that is originally inseparable in the low-dimensional space can become linearly separable in the high-dimensional space. Through this mapping, the Gaussian kernel SVC can effectively perform multi-category classification operations and output the probability distribution vector of the gas type. The probability value corresponding to each category reflects the possibility of belonging to the category, and the maximum probability judgment is performed based on the probability distribution vector to determine the most likely gas type. By performing type mapping with the preset gas type label library, the type identification of the target gas is output. The standardized feature vector is input into the second input channel of the partial least squares regression model. The partial least squares regression model selects the corresponding regression model according to the type identification of the target gas. Since different types of gases have different concentration variation patterns, each type of gas requires a different regression model for accurate prediction. Therefore, according to the previously determined gas type, the most suitable regression model for the current gas type is selected from the pre-trained model library, and then the data to be predicted is processed according to this regression model. A mixed regularization constraint is imposed on the data to be predicted. Feature selection and parameter adjustment are performed by introducing a combination of L1 norm and L2 norm, where the coefficient of L1 norm is set to 0.3 and the coefficient of L2 norm is set to 0.7. The L1 norm helps to produce a sparse solution. By sparsely processing some features in the regression model, unimportant features are removed and the model complexity is reduced. The L2 norm helps to avoid overfitting and ensure the generalization ability of the model by smoothing the regression coefficient. Under the action of mixed regularization, the most useful features for prediction are selected, and the parameters of the model are adjusted to obtain a more accurate prediction effect, and a regularized feature vector is obtained. The regularized feature vector is input into the selected regression model for latent variable mapping and linear regression calculation. The role of latent variable mapping is to convert the input data into a latent feature space related to the target gas concentration, and the linear regression calculation predicts the concentration of the target gas based on this latent feature space. Through the training of the regression model, the concentration value of the gas is calculated based on the input feature vector to obtain the predicted concentration value of the target gas.

[0043] In the embodiments of the present invention, a signal preprocessing method combining multi-scale wavelet denoising and adaptive digital filtering is constructed to effectively suppress the influence of environmental noise; a segmented temperature and pressure compensation and cross-validation optimization strategy are adopted to significantly improve the measurement stability of the system under complex working conditions; a nonlinear feature mapping structure based on a deep autoencoder network is designed, and the adaptive fusion of multi-source information is realized in combination with the improved DS evidence theory, solving the problem of feature extraction and decision optimization in multi-sensor data fusion; and a support vector classification and partial least squares regression model with mixed regularization constraints is introduced to realize high-precision identification and quantitative analysis of gas components. The present invention has strong environmental adaptability and analysis accuracy.

[0044] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0045] The infrared sensor array acquisition signals in the 3-5μm band, 7-9μm band and 10-12μm band are synchronously sampled at 10Hz to obtain multi-component gas spectral absorption data;

[0046] The maximum and minimum normalization transformation of the multi-component gas spectral absorption data is performed, and the data array is constructed according to the sensor band order to obtain the initial spectral data matrix. The initial spectral data matrix is ​​input into the db4 wavelet transformation function for three-layer wavelet decomposition, and the low-frequency approximate coefficients and high-frequency detail coefficients are extracted to obtain the wavelet decomposition coefficient matrix.

[0047] The noise variance of each high-frequency sub-band is calculated based on the wavelet decomposition coefficient matrix, and an adaptive soft threshold is set according to the noise variance. The high-frequency detail coefficients are subjected to threshold shrinkage processing to obtain the denoised wavelet coefficient matrix.

[0048] The denoised wavelet coefficient matrix is ​​reconstructed by inverse wavelet transform to obtain preliminary denoised spectral data, and the preliminary denoised spectral data is input into a third-order polynomial fitting model for baseline drift identification. The fitted baseline is subtracted from the multi-component gas spectral absorption data to obtain baseline-corrected spectral data.

[0049] The channel response characteristic compensation is performed on the baseline-corrected spectral data to obtain response-corrected spectral data, and the response-corrected spectral data are reorganized in a manner in which rows represent the number of sampling points and columns represent the number of sensors to obtain a corrected multidimensional spectral matrix.

[0050] Specifically, the infrared sensor array acquisition signals of the 3-5μm band, 7-9μm band and 10-12μm band are synchronously sampled at 10Hz. Assume that the data corresponding to each band of the sensor array is ,in Represents the signal in the 3-5μm band, Represents the signal in the 7-9μm band, Represents the signal in the 10-12μm band, Indicates time. Synchronous sampling ensures that the signals of the three bands are collected at the same time. The collected spectral absorption data are normalized to eliminate the scale difference of data in different bands and facilitate subsequent processing. The basic formula for the maximum and minimum value normalization transformation is:

[0051] ;

[0052] in, Representative In the group data The value of the feature, and are the minimum and maximum values ​​of the feature respectively. The normalized data It falls between 0 and 1, ensuring that the data of different sensor bands are compared and processed under the same standard. By normalizing the spectral data of different bands and organizing the data array according to the band order, the initial spectral data matrix is ​​obtained. , the rows of the matrix represent different sampling points, and the columns represent data from different sensor bands. The initial spectral data matrix is ​​processed using wavelet transform. The db4 wavelet function is used for three-layer wavelet decomposition. The wavelet transform decomposes the signal into low-frequency approximate coefficients and high-frequency detail coefficients, where the low-frequency approximate coefficients represent the main information of the signal, and the high-frequency detail coefficients contain the noise component of the signal. The mathematical formula for wavelet decomposition is expressed as:

[0053] ;

[0054] ;

[0055] in, and are the low-frequency and high-frequency wavelet coefficients, respectively. and It is the signal data at different levels of wavelet transform. The low-frequency approximate coefficients obtained by wavelet decomposition and high frequency detail factor Composition of wavelet decomposition coefficient matrix , which contains the components of the signal in different frequency bands. Based on the wavelet decomposition coefficient matrix, the noise variance of each high-frequency sub-band is calculated. The estimation formula of the noise variance is:

[0056] ;

[0057] in, is the noise variance of the high frequency subband, is the high frequency coefficient values, is the mean of the high frequency coefficients, is the total number of high-frequency coefficients. According to the noise variance, an adaptive soft threshold is set , which is used to perform threshold shrinkage on the high-frequency detail coefficients. This process is performed using the following formula:

[0058] ;

[0059] in, is the original value of the high-frequency detail coefficient, is the high frequency coefficient after denoising, is the threshold value, Indicates the sign of the coefficient. In this way, the high-frequency noise components are removed and the denoised wavelet coefficient matrix is ​​obtained. The denoised coefficient matrix is ​​subjected to inverse wavelet transform to reconstruct a preliminary denoised version of the original signal. The formula for inverse wavelet transform is:

[0060] ;

[0061] in, and is the denoised wavelet coefficient, and is the corresponding frequency component. After inverse wavelet transform, the preliminary denoised spectral data is obtained , the data is closer to the real gas spectrum characteristics. The preliminary denoised spectral data is input into the third-order polynomial fitting model for baseline drift identification. Baseline drift is caused by environmental factors or deviations of the equipment itself, and is removed by fitting the model. Use a third-order polynomial for baseline fitting. The mathematical model of baseline fitting is:

[0062] ;

[0063] in, is the fitted baseline, are the polynomial coefficients, is the time or sample point. By fitting the polynomial, the mathematical expression of baseline drift is obtained, and it is subtracted from the preliminary denoised spectral data to obtain the spectral data after baseline correction. After the baseline correction is completed, the channel response characteristic compensation is performed. Since the different bands of the sensor have inconsistent responses, the response correction is performed for each sensor. By constructing the response compensation function Adjust the sensor response for each band:

[0064] ;

[0065] in, is the response-corrected spectral data, is the response compensation function. By compensating the signal of each band, the response differences of different sensors are eliminated to obtain more accurate spectral data. The response-corrected spectral data are reorganized in such a way that rows represent the number of sampling points and columns represent the number of sensors to obtain the final corrected multidimensional spectral matrix. Matrix Each row in represents a different sampling point, and each column represents the correction data of a different band. In this way, the final multidimensional spectral matrix can accurately reflect the spectral characteristics of the gas.

[0066] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0067] The spectrum of the standard gas sample is collected at multiple temperature points within the preset temperature range, and the temperature compensation coefficient corresponding to each temperature point is obtained by least squares calculation based on multiple sets of standard sample test data;

[0068] The temperature compensation coefficients are reorganized according to the size of the temperature interval to construct a three-dimensional array structure, and the compensation coefficients are calculated by segmented interpolation to obtain a segmented temperature compensation coefficient matrix;

[0069] The spectrum of the standard gas sample is collected at multiple pressure points within a preset pressure range, and the pressure compensation coefficient corresponding to each pressure point is obtained by least squares calculation based on multiple sets of standard sample test data;

[0070] The pressure compensation coefficients are reorganized according to the pressure interval size to construct a three-dimensional array structure, and the compensation coefficients are calculated by segmented interpolation to obtain a segmented pressure compensation coefficient matrix;

[0071] Partitioning the corrected multidimensional spectral matrix according to temperature and pressure ranges, and inputting the partitioned data into a K-fold cross-validation model to obtain a partitioned data group to be compensated;

[0072] The partition data group to be compensated is respectively subjected to compensation operation with the segmented temperature compensation coefficient matrix and the segmented pressure compensation coefficient matrix of the corresponding interval to obtain the spectral data after preliminary compensation;

[0073] The spectral data after preliminary compensation are cross-validated and evaluated, and the compensation coefficients are iteratively optimized according to the validation deviation to obtain the spectral data after optimized compensation. The spectral data after optimized compensation are reorganized, spliced ​​and normalized to obtain the standard spectral matrix.

[0074] Specifically, within a preset temperature range, spectra are collected at multiple temperature points. Spectral tests are performed at different temperatures to obtain corresponding spectral absorption data ,in Indicates Test temperature points, Represents the spectral data at this temperature. Based on these data, the spectral data at each temperature point are fitted using the least squares method to obtain the temperature compensation coefficient. The least squares method determines the compensation coefficient by minimizing the fitting error. The specific formula is:

[0075] ;

[0076] in, It is the temperature The corresponding fitting function is, is the parameter vector of the fitting. By solving this optimization problem, the compensation coefficient of each temperature point is obtained , these compensation coefficients reflect the effect of temperature on spectral data. The compensation coefficients of each temperature point are reorganized into a three-dimensional array structure according to the size of the temperature range. Assuming the temperature range is arrive The temperature step size is , the compensation coefficient for each temperature point Organized into a matrix or 3D array , which is of the form:

[0077] ;

[0078] This three-dimensional array contains the compensation coefficients at different temperature points. On this basis, a continuous temperature compensation coefficient matrix is ​​obtained through the piecewise interpolation method, which can perform compensation operations at any temperature point. The basic formula of piecewise interpolation is:

[0079] ;

[0080] in, At any temperature The interpolation compensation coefficient under is the interpolation function weight, which is an interpolation coefficient based on the temperature range, ensuring that there is a suitable compensation coefficient in different temperature ranges. Similarly, the spectrum of the gas sample is collected within the preset pressure range, and the pressure compensation coefficient is obtained by least squares fitting. Assume that at the pressure point The spectral data obtained is The pressure compensation coefficient is obtained by the least squares method. The fitting formula used is the same as the calculation method of the temperature compensation coefficient, in which the pressure variable On:

[0081] ;

[0082] The pressure compensation coefficient obtained Describe the effect of pressure on spectral data. Reorganize the pressure compensation coefficient into a three-dimensional array according to the size of the pressure range. , which is similar to the three-dimensional array of temperature compensation coefficients, and a continuous pressure compensation coefficient matrix is ​​obtained by segmented interpolation. Data partitioning is performed on the corrected multidimensional spectral matrix. Assume that the corrected spectral data matrix is , which contains the sample values ​​of each band at different time points. In order to perform temperature and pressure compensation, these data are partitioned according to the temperature and pressure range. For example, set a temperature interval and a pressure range Divide the data into multiple small intervals, corresponding to different temperature and pressure combinations. , the data in this interval is input into the K-fold cross validation model, and the partitioned data group to be compensated is obtained through the model. K-fold cross validation is a model evaluation method used to test whether the data meets expectations. subsets, and use them cyclically The cross-validation model helps the system determine whether the data in each interval meets the requirements of temperature and output compensation, and obtains the partition data group to be compensated. The partition data group to be compensated is respectively compared with the corresponding segmented temperature compensation coefficient matrix And the segmented pressure compensation coefficient matrix Perform compensation operation. The compensation operation formula is:

[0083] ;

[0084] in, is the compensated spectral data, and are the temperature and pressure compensation coefficients, respectively. In this process, according to the temperature and pressure values ​​of each data point, the compensation coefficients are extracted from the corresponding temperature compensation coefficient matrix and pressure compensation coefficient matrix, and the corresponding compensation operations are performed. The spectral data after preliminary compensation is cross-validated and evaluated. The compensation coefficients are optimized according to the verification deviation. The verification deviation is calculated by comparing the spectral data after preliminary compensation with the original data. If the verification deviation exceeds the preset threshold, the compensation coefficients are iteratively optimized according to the deviation to adjust the temperature and pressure compensation coefficients. The optimized compensation coefficients are obtained by minimizing the error:

[0085] ;

[0086] The spectral data after optimization and compensation are reorganized, spliced ​​and normalized to obtain a standard spectral matrix The normalization operation uses methods such as maximum and minimum value normalization or z-score standardization to ensure that the data of all bands are on the same scale to facilitate subsequent analysis.

[0087] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0088] The standard spectral matrix is ​​centralized, and the mean value of each band is subtracted from the spectral data of the band to obtain a centralized spectral data matrix. The covariance values ​​between the bands are calculated based on the centralized spectral data matrix, and a symmetric covariance matrix is ​​constructed through matrix operations.

[0089] The covariance matrix is ​​input into the eigenvalue decomposition algorithm, and the eigenvalue sequence and the corresponding eigenvector matrix are obtained by diagonal decomposition. The eigenvalue sequence is arranged in descending order, and the cumulative contribution rate is calculated. The eigenvalues ​​and eigenvectors corresponding to the cumulative contribution rate greater than 95% are selected to obtain the principal component characteristic matrix.

[0090] The standard spectral matrix is ​​input into the partial least squares decomposition algorithm to calculate the loading matrix and the score matrix, and the latent variables with explained variance greater than 90% are screened to obtain the latent variable feature matrix. The principal component feature matrix and the latent variable feature matrix are then concatenated by column vectors to obtain the combined feature matrix.

[0091] The column vectors of the combined characteristic matrix are Schmidt orthogonalized in turn, and each characteristic vector is subtracted from the previously orthogonalized vector and normalized to obtain the spectral characteristic vector.

[0092] Specifically, the original standard spectral data is preprocessed to make it suitable for subsequent feature extraction and analysis. Assume there is a standard spectral matrix , where each row represents the measurement data of a spectral sampling, and each column represents the observation value of a spectral band. The dimension of the matrix is ,in represents the number of samples, Indicates the number of spectral bands. When performing centering, the mean of each band is calculated, and then the mean of each band is subtracted from the observed data of each band. It can be expressed by the following formula:

[0093] ;

[0094] in, Represents the first Row, No. The elements of the column, is the value of the corresponding element in the original matrix, It is Column (i.e. The calculation formula is:

[0095] ;

[0096] Through this step, the data of each band is converted into a form with a mean value of zero, which helps to eliminate the errors caused by baseline drift or offset between different bands. Based on the spectral data matrix after centralization Calculate the covariance matrix. The covariance matrix is ​​an important tool to measure the linear relationship between different bands. The calculation formula of the covariance matrix is:

[0097] ;

[0098] in, is the covariance matrix, is the transpose of the centered matrix, is the centralized data matrix. The covariance matrix is ​​symmetric and contains the covariance information between each band. Each element Indicates Band and The covariance between the bands reflects the linear correlation between the two. The covariance matrix is ​​input into the eigenvalue decomposition algorithm for diagonalization to obtain the eigenvalue sequence and the corresponding eigenvector matrix. The goal of eigenvalue decomposition is to convert the covariance matrix Decomposed into the following form:

[0099] ;

[0100] in, is the eigenvector matrix, is a diagonal matrix containing the eigenvalues ​​of the covariance matrix. represents the direction of each principal component in the covariance matrix, and the eigenvalue In order to select the main components, they are arranged in descending order of eigenvalues ​​and the cumulative contribution rate is calculated:

[0101] ;

[0102] in, Indicates eigenvalues, is the total number of spectral bands. Select eigenvalues ​​and eigenvectors with cumulative contribution greater than 95%. These principal components retain most of the data information. The selected eigenvector matrix That is, the principal component feature matrix, which contains the most important components in the data. Enter the partial least squares (PLS) algorithm. PLS is a technique used to extract latent variables by constructing a loading matrix and a score matrix to extract the relationship between variables. In the PLS model, the loading matrix is ​​obtained by solving the following optimization problem and the score matrix

[0103] ;

[0104] in, is the score matrix, representing the latent structure of the data, is the loading matrix, which reflects the contribution of each band to the latent variable. By minimizing the sum of squared errors, a set of latent variables is obtained, each of which describes some implicit characteristics of the input data. In PLS, latent variables with an explained variance greater than 90% are screened out. These latent variables have a strong ability to explain data changes, so they have a higher contribution in subsequent analysis. The latent variable feature matrix is ​​derived from the loading matrix and the score matrix Extracted from, expressed as:

[0105] Latent variable feature matrix ;

[0106] The matrix retains the representation of the data in the latent variable space and can capture the main structural information of the data. and the latent variable feature matrix Concatenate column vectors to obtain a combined feature matrix , which contains the principal component characteristics and latent variable characteristics of the data. The concatenated matrix It is a matrix that combines the most important components and potential structures of the data and can effectively represent the key information in the original data. The column vectors of the combined feature matrix are sequentially Schmidt orthogonalized. Schmidt orthogonalization is a method of converting a set of linearly dependent vectors into a set of linearly independent vectors. Suppose there is a set of vectors , the orthogonalization process is carried out through the following steps:

[0107] ;

[0108] ;

[0109] ;

[0110]

[0111] in, represents the orthogonalized vector, represents the original vector, and represents the dot product of the vector. Through orthogonalization, each eigenvector is ensured to be linearly independent, which helps to improve the stability and interpretability of data analysis. The obtained spectral eigenvector Used to represent the main features of spectral data.

[0112] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0113] The spectral feature vector is input into the first encoding layer of the deep autoencoder network model for feature compression, where the first encoding layer contains twice the number of first neurons as the sum of the number of principal component features and the number of latent variable features. The input of the first neuron is processed by the ReLU activation function and batch normalization with a weight of 0.7, and feature screening is performed through a Dropout layer of 0.3 to obtain the first layer of encoding features;

[0114] The first-layer encoding features are input into the second encoding layer in the deep autoencoder network model for dimensionality reduction mapping, where the second encoding layer contains the second neuron with the sum of the number of principal component features and the number of latent variable features. The input of the second neuron is processed by the ReLU activation function and the batch normalization operation with a weight of 0.7. The original feature information is retained by adding skip connections to obtain the second-layer encoding features;

[0115] The second layer encoding features are input into the third encoding layer in the deep autoencoder network model for feature aggregation, where the third encoding layer contains third neurons whose number is 1 / 2 of the sum of the number of principal component features and the number of latent variable features. The input of the third neuron is processed by the ReLU activation function and batch normalization with a weight of 0.7, and the attention mechanism is introduced to enhance the key features to obtain the third layer encoding features;

[0116] The third layer encoding features are input into the first decoding layer in the deep autoencoder network model for feature reconstruction, where the number of neurons in the first decoding layer is the same as that in the third encoding layer, and the input of the fourth neuron in the first decoding layer is processed by the ReLU activation function and batch normalization with a weight of 0.7, and a residual connection is established with the corresponding encoding layer features to obtain the first layer decoding features;

[0117] The first layer decoding features are input into the second decoding layer in the deep autoencoder network model for dimension expansion, where the number of neurons in the second decoding layer is the same as that in the second encoding layer. The input of the fifth neuron in the second decoding layer is processed by the ReLU activation function and the batch normalization operation with a weight of 0.7. The encoding layer information is fused through the reverse jump connection to obtain the second layer decoding features;

[0118] The second layer of decoded features are input into the third decoding layer in the deep autoencoder network model for feature restoration, where the number of neurons in the third decoding layer is the same as that in the first encoding layer. The input of the sixth neuron in the third decoding layer is operated by the Sigmoid activation function, and combined with a 0.3 Dropout layer to prevent overfitting, to obtain a reconstructed feature vector;

[0119] A loss function is constructed based on the reconstructed feature vector and the spectral feature vector. The reconstruction error weight is set to 0.7 and the feature correlation constraint weight is set to 0.3. The similarity between features is calculated by cosine similarity and feature compensation is performed to obtain the initial feature mapping vector. The residual error is calculated for the initial feature mapping vector and the reconstructed feature vector to obtain the compensated feature mapping vector.

[0120] Specifically, the spectral feature vector is input into the first encoding layer of the deep autoencoder network for feature compression. Assume that the input spectral feature vector is , which contains the combined representation of the principal component features and the latent variable features. The number of neurons in the first encoding layer is set to twice the sum of the number of principal component features and the number of latent variable features, that is:

[0121] ;

[0122] in, represents the number of principal component features, Represents the number of latent variable features. The input of each neuron is processed by the ReLU activation function and normalized by batch normalization. The batch normalization operation is implemented by the following formula:

[0123] ;

[0124] in, is the current input, and are the mean and standard deviation of the current batch, respectively. and are learnable scaling and offset parameters, is a small constant to prevent division by zero. A Dropout layer is used for feature screening to avoid overfitting, and the Dropout ratio of this layer is 0.3. The mathematical expression of the Dropout operation is:

[0125] ;

[0126] in, is an indicator function, which means that the feature value is "kept" with a certain probability, and the others are set to zero. The first layer of encoded features obtained It is expressed as:

[0127] ;

[0128] in, is the weight matrix of the first coding layer, is the input spectral feature vector. The first layer encoding feature Input the second coding layer for dimensionality reduction mapping. The number of neurons in the second coding layer is set to the sum of the number of principal component features and the number of latent variable features, that is:

[0129] ;

[0130] Similarly, the input of each neuron is processed by the ReLU activation function and batch normalized. In order to retain the original feature information, the second encoding layer introduces a jump connection. The jump connection is implemented as follows:

[0131] ;

[0132] in, is the weight matrix of the second coding layer, is the output of the previous layer. Through the jump connection, the network can learn new features and retain the information of the original features, thereby reducing the information loss in training. Input to the third coding layer for feature aggregation. The number of neurons in the third coding layer is set to half the sum of the number of principal component features and the number of latent variable features, that is:

[0133] ;

[0134] In this layer, the input is also processed using the ReLU activation function and normalized by batch normalization. At the same time, in order to enhance key features, the third encoding layer introduces the attention mechanism. The attention mechanism weights the features by calculating the importance of each feature. The formula is as follows:

[0135] ;

[0136] in, is the weight matrix of the third encoding layer. The Softmax function is used to calculate the weight of the feature. In this way, the network can automatically focus on the most important features. Get the third layer encoding feature :

[0137] ;

[0138] After the encoding process is completed, decoding is performed to reconstruct the original spectral data. Input to the first decoding layer for feature reconstruction. The number of neurons in the first decoding layer is the same as that in the third encoding layer, that is:

[0139] ;

[0140] Similar to the encoding process, the input of the first decoding layer passes through the ReLU activation function and establishes a residual connection with the features of the corresponding encoding layer. The residual connection is implemented by the following formula:

[0141] ;

[0142] in, is the weight matrix of the first decoding layer, is the third layer encoding feature. Through residual connection, the network can better retain the original information during feature reconstruction. Input to the second decoding layer for dimension expansion. The number of neurons in the second decoding layer is the same as that in the second encoding layer, namely:

[0143] ;

[0144] This layer also passes through the ReLU activation function, and integrates the encoding layer information into the decoding process through the reverse jump connection:

[0145] ;

[0146] The second layer decoding features Input to the third decoding layer for feature restoration. The number of neurons in the third decoding layer is the same as that in the first encoding layer, namely:

[0147] ;

[0148] The activation function of this layer is Sigmoid, which is used to output a reconstructed feature vector between 0 and 1:

[0149] ;

[0150] After the reconstruction is completed, a loss function is constructed to train the deep autoencoder network. The loss function includes the reconstruction error and the feature correlation constraint, where the reconstruction error weight is set to 0.7 and the feature correlation constraint weight is set to 0.3. The reconstruction error is calculated using the mean square error (MSE), and the formula is:

[0151] ;

[0152] in, is the original spectral data, is the reconstructed spectral data. The feature correlation constraint calculates the similarity between features through cosine similarity, and the formula is:

[0153] ;

[0154] The loss function is expressed as:

[0155] ;

[0156] By minimizing the loss function, the network can compensate for the spectral features and obtain a compensated feature mapping vector.

[0157] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0158] The compensated feature mapping vectors are grouped according to the sensor bands, the basic probability distribution matrix of each sensor is constructed, and the initial trust weight is set based on the signal-to-noise ratio to obtain the basic evidence matrix at the sensor layer;

[0159] The sensor layer basic evidence matrix is ​​input into the conflict factor calculation module, and the conflict degree between evidences is calculated by Dempster combination rule. The penalty factor is adjusted for evidence with a conflict degree greater than 0.8 to obtain sensor layer fusion evidence.

[0160] The sensor layer fusion evidence is decomposed, the feature importance of the evidence in different feature subspaces is calculated by the mutual information criterion, and the feature layer trust function is constructed to obtain the feature layer basic evidence matrix;

[0161] The basic evidence matrix of the feature layer is graded according to the feature contribution, and the features with a contribution less than 0.1 are pruned. The feature support is reconstructed through evidence accumulation operation to obtain the feature layer fusion evidence;

[0162] Based on the sensor layer fusion evidence and the feature layer fusion evidence, the decision layer evidence matrix is ​​constructed, and the trust threshold judgment rule is set to obtain the initial decision support. The initial decision support is clustered according to the maximum trust principle, and the evidence with a trust difference of less than 0.05 is combined into the same category to obtain the decision layer fusion evidence.

[0163] The decision-layer fusion evidence is input into the evidence chain optimization module, and the trust weights of each layer are updated and fine-tuned through back propagation to obtain the optimized multi-layer fusion evidence. Based on the optimized multi-layer fusion evidence, evidence comprehensive calculation is performed, and the Dempster combination rule is used to integrate multi-source information to obtain the fusion decision vector.

[0164] Specifically, the compensated feature mapping vectors are grouped according to the different bands of the sensors, and the basic probability allocation matrix of each sensor is constructed. According to the band characteristics of each sensor, the feature mapping vectors are assigned to different band groups, and the preliminary probability allocation is calculated for each group. These probability allocation matrices represent the observation evidence of each sensor in a specific band. In actual operation, each band corresponds to the response value of a sensor, usually in the form of Indicates that Represents the band number of the sensor, and each element Indicated in The probability distribution value of each feature. In order to accurately evaluate the reliability of each sensor, the initial trust weight is set based on the signal-to-noise ratio of each sensor, and the trust of each sensor in different environments is quantified. Sensors with higher signal-to-noise ratios have higher trust. Calculated by the following formula:

[0165] ;

[0166] in, For the The signal-to-noise ratio of the sensor, is the total number of sensors. Based on the initial trust weight, the sensor layer basic evidence matrix is ​​constructed , which expresses the evidence distribution of different sensors for each feature. Each element in this matrix Indicates sensor In the The sensor layer basic evidence matrix is ​​input into the conflict factor calculation module, and the conflict degree between different evidences is calculated using the Dempster combination rule. In the Dempster combination, the conflict degree Determined by calculating the intersection between evidence matrices:

[0167] ;

[0168] If the conflict If it is greater than 0.8, it is considered that there is a strong conflict between the evidences and the weight of the evidence needs to be adjusted through the penalty factor. The adjusted evidence matrix is ​​obtained by introducing the adjustment factor To achieve:

[0169] ;

[0170] in, is the adjustment factor, usually chosen Balance the conflicting evidence to obtain sensor-level fusion evidence The sensor layer fusion evidence is decomposed into features, and the importance of each feature is calculated using the mutual information criterion. The mutual information metric measures the degree of information sharing between two random variables. The calculation formula is:

[0171] ;

[0172] According to the calculation results of mutual information, the features with high importance are selected, and the feature layer trust function is constructed according to the contribution of the features. It represents the contribution of each feature to the sensor information and is weighted according to the contribution. The feature layer basic evidence matrix is ​​graded and features with a contribution less than 0.1 are pruned. This process helps to remove redundant or irrelevant features and improve the fusion effect. The pruned feature support is reconstructed through evidence accumulation operations to obtain the feature layer fusion evidence matrix Each element Indicates At the decision layer, the decision layer evidence matrix is ​​constructed based on the sensor layer fusion evidence and the feature layer fusion evidence. By setting a trust threshold , classify different evidences into two categories: high confidence and low confidence, and obtain the initial decision support , which is defined as the weighted average of all evidence. The support is expressed as:

[0173] ;

[0174] in, and are the trust weights of sensors and features, respectively. Applying the maximum trust principle, the evidence with a trust difference of less than 0.05 is clustered into the same category to obtain the fusion evidence of the decision layer. In order to optimize the multi-layer fusion process, the fusion evidence of the decision layer Input to the evidence chain optimization module. This module updates the trust weights of each layer through back propagation and makes fine adjustments to make the trust at different levels more accurate. The optimized evidence matrix Used to integrate multi-source information based on Dempster combination rules to obtain the final fusion decision vector , represents the decision result obtained by integrating the information of all sensors and feature layers. Through the multi-layer evidence fusion process, the accuracy and robustness of gas analysis are effectively improved, ensuring the adaptability to environmental changes and different measurement conditions.

[0175] In a specific embodiment, the process of executing step 106 may specifically include the following steps:

[0176] Perform data standardization preprocessing on the fusion decision vector, organize the data structure according to the sample feature dimension, and obtain the standardized feature vector;

[0177] The standardized feature vector is input into the first input channel of the Gaussian kernel support vector classifier, and the probability distribution vector of the gas type is obtained through feature space mapping and multi-category classification operations. The probability distribution vector is subjected to maximum probability judgment, and type mapping is performed based on a preset gas type label library to output the type identification of the target gas.

[0178] Input the standardized feature vector into the second input channel of the partial least squares regression model, select the corresponding regression model according to the type identification of the target gas, and obtain the data to be predicted;

[0179] A mixed regularization constraint is imposed on the predicted data. Feature selection and parameter adjustment are performed through the L1 norm coefficient of 0.3 and the L2 norm coefficient of 0.7 to obtain the regularized feature vector. The regularized feature vector is input into the selected regression model, and latent variable mapping and linear regression calculation are performed to obtain the concentration value of the target gas.

[0180] Specifically, the fusion decision vector is preprocessed by data standardization to eliminate the differences in the dimensions of different features and ensure that each feature has the same scale. The mean of each feature is removed and scaled to a distribution with a standard deviation of 1. , each element All are standardized, and the standardization formula is:

[0181] ;

[0182] in, It is The mean of the features, It is The standard deviation of the feature, is the standardized eigenvalue. Through this step, the standardized eigenvector is obtained , which is used as the input for subsequent models. The standardized feature vector is input to the first input channel of the Gaussian kernel support vector classifier. The Gaussian kernel support vector classifier maps the input feature space to a higher-dimensional feature space through a nonlinear mapping to facilitate more complex category division. In the Gaussian kernel support vector classifier, the similarity between data points is calculated using the Gaussian kernel function, which is in the form of:

[0183] ;

[0184] in, is the input feature vector and The square of the Euclidean distance between is the parameter of the kernel function, which determines the "width" or "range" after mapping to the high-dimensional space. Through this Gaussian kernel function, the features of the input data are mapped to the high-dimensional space, so that better classification can be performed through the hyperplane in the new feature space. After Gaussian kernel mapping, the Gaussian kernel support vector classifier performs multi-class classification operations and generates a probability distribution vector of gas types. ,in is the number of gas types, Indicates that the input data belongs to The Gaussian kernel support vector classifier performs the maximum probability judgment on the probability distribution vector and selects the gas type with the highest probability. is the maximum probability value, the model outputs the gas type identifier According to the preset gas type label library, the gas type identifier is mapped to the specific gas name, and the type identifier of the target gas is output. At the same time, the standardized feature vector is input into the second input channel of the partial least squares regression model. Partial least squares regression is a method that reduces the dimension of the input feature matrix and the output target matrix simultaneously to maximize the covariance between the input features and the output. Assume that the standardized feature vector is The goal of the regression model is to identify Select the corresponding regression model for training. The regression model calculates the gas concentration by minimizing the following loss function:

[0185] ;

[0186] in, is the actual gas concentration, is the concentration predicted by the regression model, and the loss function Measures the difference between the model prediction and the true value. Identified by the type of target gas , select the corresponding regression model to get the data to be predicted. When processing the data to be predicted, apply hybrid regularization constraints to improve the generalization ability of the model and prevent overfitting. Hybrid regularization constraints combine L1 regularization (Lasso) and L2 regularization (Ridge), usually expressed as:

[0187] ;

[0188] in, and are the coefficients of L1 and L2 regularization, respectively. is the weight vector in the regression model, represents the L1 norm (i.e., the sum of the absolute values ​​of the weight vector), represents the L2 norm (i.e., the sum of the squares of the weight vector). In this process, the L1 norm helps with feature selection, while the L2 norm helps adjust the complexity of the model. Through these regularization constraints, the model can better fit the training data while avoiding overfitting. The regularized feature vector Input into the selected regression model, perform latent variable mapping and linear regression calculation, and obtain the final predicted value of gas concentration. The partial least squares regression model calculates the potential linear relationship and uses the regression coefficient Make a prediction and get the concentration value of the target gas ,Right now:

[0189] ;

[0190] in, is the input feature matrix, is the weight vector of the regression model, is the final predicted gas concentration value.

[0191] The above describes the infrared gas analysis method of multi-sensor fusion in the embodiment of the present invention. The following describes the infrared gas analysis device of multi-sensor fusion in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a multi-sensor fusion infrared gas analysis device includes:

[0192] De-noising module 201, used for performing adaptive digital filtering and multi-scale wavelet denoising operations on multi-component gas spectral absorption data collected by a multi-band infrared sensor array to obtain a corrected multi-dimensional spectral matrix;

[0193] The compensation module 202 is used to perform a partition cross-validation compensation operation on the corrected multi-dimensional spectrum matrix and the segmented temperature compensation coefficient matrix and the segmented pressure compensation coefficient matrix to obtain a standard spectrum matrix;

[0194] A construction module 203 is used to construct a covariance matrix according to a standard spectral matrix, extract principal component features and latent variable features by eigenvalue decomposition, and obtain a spectral feature vector;

[0195] The calculation module 204 is used to input the spectral feature vector into the deep autoencoder network model for nonlinear mapping and residual reconstruction calculation to obtain a compensated feature mapping vector;

[0196] A fusion module 205 is used to perform information fusion on the compensated feature mapping vector to obtain a fusion decision vector;

[0197] The output module 206 is used to input the fusion decision vector into the Gaussian kernel support vector classifier and the partial least squares regression model respectively, and output the type identification and concentration value of the target gas.

[0198] Through the coordinated cooperation of the above-mentioned components, a signal preprocessing method combining multi-scale wavelet denoising and adaptive digital filtering is constructed to effectively suppress the influence of environmental noise; the segmented temperature and pressure compensation and cross-validation optimization strategies are adopted to significantly improve the measurement stability of the system under complex working conditions; a nonlinear feature mapping structure based on a deep autoencoder network is designed, and the adaptive fusion of multi-source information is realized in combination with the improved DS evidence theory, solving the problem of feature extraction and decision optimization in multi-sensor data fusion; by introducing support vector classification and partial least squares regression models with mixed regularization constraints, high-precision identification and quantitative analysis of gas components are achieved, and the present invention has strong environmental adaptability and analysis accuracy.

[0199] above Figure 2 The infrared gas analysis device with multi-sensor fusion in the embodiment of the present invention is described in detail from the perspective of modular functional entity. The infrared gas analysis device with multi-sensor fusion in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0200] Figure 3 : is a structural schematic diagram of a multi-sensor fusion infrared gas analysis device provided by an embodiment of the present invention. The multi-sensor fusion infrared gas analysis device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the multi-sensor fusion infrared gas analysis device 300. Furthermore, the processor 310 can be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the multi-sensor fusion infrared gas analysis device 300 to implement the steps of the above-mentioned multi-sensor fusion infrared gas analysis method.

[0201] The multi-sensor fusion infrared gas analysis device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will appreciate that Figure 3 The structure of the multi-sensor fusion infrared gas analysis device shown does not constitute a limitation on the multi-sensor fusion infrared gas analysis device provided by the present invention, and may include more or less components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0202] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the multi-sensor fusion infrared gas analysis method.

[0203] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0204] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0205] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-sensor fusion infrared gas analysis method, characterized in that: The method comprises: Adaptive digital filtering and multi-scale wavelet denoising are performed on the multi-component gas spectral absorption data collected by the multi-band infrared sensor array to obtain a corrected multi-dimensional spectral matrix. The corrected multidimensional spectrum matrix is ​​subjected to a partition cross-validation compensation operation with a segmented temperature compensation coefficient matrix and a segmented pressure compensation coefficient matrix to obtain a standard spectrum matrix; Constructing a covariance matrix according to the standard spectral matrix, extracting principal component features and latent variable features by eigenvalue decomposition, and obtaining a spectral feature vector; Inputting the spectral feature vector into a deep autoencoder network model for nonlinear mapping and residual reconstruction calculation to obtain a compensated feature mapping vector; Performing information fusion on the compensated feature mapping vector to obtain a fusion decision vector; The fusion decision vector is input into a Gaussian kernel support vector classifier and a partial least squares regression model respectively, and the type identification and concentration value of the target gas are output.

2. The infrared gas analysis method of multi-sensor fusion according to claim 1 is characterized in that: The adaptive digital filtering and multi-scale wavelet denoising operation are performed on the multi-component gas spectral absorption data collected by the multi-band infrared sensor array to obtain a corrected multi-dimensional spectral matrix, including: The infrared sensor array acquisition signals in the 3-5μm band, 7-9μm band and 10-12μm band are synchronously sampled at 10Hz to obtain multi-component gas spectral absorption data; Performing maximum and minimum value normalization transformation on the multi-component gas spectral absorption data, and constructing a data array according to the sensor band order to obtain an initial spectral data matrix, and inputting the initial spectral data matrix into the db4 wavelet transformation function to perform a three-layer wavelet decomposition, extracting low-frequency approximate coefficients and high-frequency detail coefficients, and obtaining a wavelet decomposition coefficient matrix; The noise variance of each high-frequency sub-band is calculated based on the wavelet decomposition coefficient matrix, and an adaptive soft threshold is set according to the noise variance, and a threshold shrinkage process is performed on the high-frequency detail coefficients to obtain a denoised wavelet coefficient matrix; Performing inverse wavelet transform reconstruction on the denoised wavelet coefficient matrix to obtain preliminary denoised spectral data, inputting the preliminary denoised spectral data into a third-order polynomial fitting model to identify baseline drift, and subtracting the fitting baseline from the multi-component gas spectral absorption data to obtain baseline-corrected spectral data; Channel response characteristic compensation is performed on the baseline-corrected spectral data to obtain response-corrected spectral data, and the response-corrected spectral data is reorganized in a manner in which rows represent the number of sampling points and columns represent the number of sensors to obtain a corrected multidimensional spectral matrix.

3. The infrared gas analysis method of multi-sensor fusion according to claim 2 is characterized in that: The corrected multidimensional spectrum matrix is ​​subjected to a partition cross-validation compensation operation with a segmented temperature compensation coefficient matrix and a segmented pressure compensation coefficient matrix to obtain a standard spectrum matrix, including: The spectrum of the standard gas sample is collected at multiple temperature points within the preset temperature range, and the temperature compensation coefficient corresponding to each temperature point is obtained by least squares calculation based on multiple sets of standard sample test data; The temperature compensation coefficients are reorganized to construct a three-dimensional array structure according to the size of the temperature interval, and the compensation coefficients are calculated by segmented interpolation to obtain a segmented temperature compensation coefficient matrix; The spectrum of the standard gas sample is collected at multiple pressure points within a preset pressure range, and the pressure compensation coefficient corresponding to each pressure point is obtained by least squares calculation based on multiple sets of standard sample test data; The pressure compensation coefficients are reorganized to construct a three-dimensional array structure according to the pressure interval size, and the compensation coefficients are calculated by segmented interpolation to obtain a segmented pressure compensation coefficient matrix; Partitioning the corrected multidimensional spectral matrix according to temperature and pressure ranges, and inputting the partitioned data into a K-fold cross-validation model to obtain a partitioned data group to be compensated; Perform compensation operations on the partition data group to be compensated and the segmented temperature compensation coefficient matrix and the segmented pressure compensation coefficient matrix of the corresponding interval respectively to obtain spectral data after preliminary compensation; The spectral data after preliminary compensation is cross-validated and evaluated, and the compensation coefficient is iteratively optimized according to the validation deviation to obtain the spectral data after optimized compensation, and the spectral data after optimized compensation is reorganized, spliced ​​and normalized to obtain a standard spectral matrix.

4. The infrared gas analysis method of multi-sensor fusion according to claim 3 is characterized in that: The method of constructing a covariance matrix according to the standard spectral matrix, extracting principal component features and latent variable features by eigenvalue decomposition, and obtaining a spectral feature vector includes: The standard spectral matrix is ​​centralized, the mean value of each band is subtracted from the spectral data of the band to obtain a centralized spectral data matrix, and the covariance values ​​between the bands are calculated based on the centralized spectral data matrix, and a symmetric covariance matrix is ​​constructed through matrix operations; Input the covariance matrix into the eigenvalue decomposition algorithm, diagonalize and decompose to obtain an eigenvalue sequence and a corresponding eigenvector matrix, arrange the eigenvalue sequence in descending order, calculate the cumulative contribution rate, select the eigenvalues ​​and eigenvectors corresponding to the cumulative contribution rate greater than 95%, and obtain the principal component characteristic matrix; Input the standard spectral matrix into the partial least squares decomposition algorithm, calculate the loading matrix and the score matrix, screen the latent variables with explained variance greater than 90%, obtain the latent variable feature matrix, and perform column vector splicing on the principal component feature matrix and the latent variable feature matrix to obtain a combined feature matrix; The column vectors of the combined characteristic matrix are sequentially subjected to Schmidt orthogonalization calculations, and each characteristic vector is subjected to a difference operation with a previously orthogonalized vector and normalized to obtain a spectral characteristic vector.

5. The infrared gas analysis method of multi-sensor fusion according to claim 4 is characterized in that: The step of inputting the spectral feature vector into a deep autoencoder network model for nonlinear mapping and residual reconstruction calculation to obtain a compensated feature mapping vector includes: Inputting the spectral feature vector into the first coding layer in the deep autoencoder network model for feature compression, wherein the first coding layer includes first neurons twice the sum of the number of principal component features and the number of latent variable features, the input of the first neuron is processed by the ReLU activation function and the batch normalization operation with a weight of 0.7, and feature screening is performed through a Dropout layer of 0.3 to obtain the first layer of coding features; Inputting the first layer of coding features into the second coding layer in the deep autoencoder network model for dimensionality reduction mapping, wherein the second coding layer comprises a second neuron having the sum of the number of principal component features and the number of latent variable features, the input of the second neuron is processed by a ReLU activation function and a batch normalization operation with a weight of 0.7, and the original feature information is retained by adding a skip connection to obtain the second layer of coding features; Inputting the second layer of encoded features into the third encoding layer in the deep autoencoder network model for feature aggregation, wherein the third encoding layer includes third neurons whose number is 1 / 2 of the sum of the number of principal component features and the number of latent variable features, and the input of the third neurons is processed by a ReLU activation function and a batch normalization operation with a weight of 0.7, and an attention mechanism is introduced to enhance key features, so as to obtain the third layer of encoded features; Input the third layer encoding features into the first decoding layer in the deep autoencoder network model for feature reconstruction, wherein the number of neurons in the first decoding layer is the same as that in the third encoding layer, and the input of the fourth neuron in the first decoding layer is processed by the ReLU activation function and the batch normalization operation with a weight of 0.7, and a residual connection is established with the corresponding encoding layer features to obtain the first layer decoding features; Inputting the first layer decoding features into the second decoding layer in the deep autoencoder network model for dimension expansion, wherein the number of neurons in the second decoding layer is the same as that in the second encoding layer, the input of the fifth neuron in the second decoding layer is processed by the ReLU activation function and the batch normalization operation with a weight of 0.7, and the encoding layer information is fused through the reverse jump connection to obtain the second layer decoding features; Input the second layer of decoded features into the third decoding layer in the deep autoencoder network model for feature restoration, wherein the number of neurons in the third decoding layer is the same as that in the first encoding layer, and the input of the sixth neuron in the third decoding layer is operated by a Sigmoid activation function, and combined with a 0.3 Dropout layer to prevent overfitting, to obtain a reconstructed feature vector; A loss function is constructed based on the reconstructed feature vector and the spectral feature vector, the reconstruction error weight is set to 0.7, the feature correlation constraint weight is set to 0.3, the similarity between features is calculated by cosine similarity and feature compensation is performed to obtain an initial feature mapping vector, and the residual error is calculated for the initial feature mapping vector and the reconstructed feature vector to obtain a compensated feature mapping vector.

6. The infrared gas analysis method of multi-sensor fusion according to claim 5 is characterized in that: The performing information fusion on the compensated feature mapping vector to obtain a fusion decision vector includes: The compensated feature mapping vectors are grouped according to the sensor bands, a basic probability distribution matrix of each sensor is constructed, and an initial trust weight is set based on the signal-to-noise ratio to obtain a sensor layer basic evidence matrix; The sensor layer basic evidence matrix is ​​input into the conflict factor calculation module, the conflict degree between evidences is calculated by Dempster combination rule, and the penalty factor is adjusted for evidence with a conflict degree greater than 0.8 to obtain sensor layer fusion evidence; Performing feature decomposition on the sensor layer fusion evidence, calculating feature importance of evidence in different feature subspaces by mutual information criterion, and constructing a feature layer trust function to obtain a feature layer basic evidence matrix; The feature layer basic evidence matrix is ​​graded according to feature contribution, features with contribution less than 0.1 are pruned, feature support is reconstructed through evidence accumulation operation, and feature layer fusion evidence is obtained; Based on the sensor layer fusion evidence and the feature layer fusion evidence, a decision layer evidence matrix is ​​constructed, and a trust threshold judgment rule is set to obtain an initial decision support degree, and the initial decision support degree is clustered according to the maximum trust principle, and evidence with a trust difference of less than 0.05 is combined into the same category to obtain decision layer fusion evidence; The decision-layer fusion evidence is input into the evidence chain optimization module, and the trust weights of each layer are updated and fine-tuned through back propagation to obtain optimized multi-layer fusion evidence. Evidence comprehensive calculation is performed based on the optimized multi-layer fusion evidence, and Dempster combination rule is used to integrate multi-source information to obtain a fusion decision vector.

7. The infrared gas analysis method of multi-sensor fusion according to claim 6 is characterized in that: The fusion decision vector is input into the Gaussian kernel support vector classifier and the partial least squares regression model respectively, and the type identification and concentration value of the target gas are output, including: Performing data standardization preprocessing on the fusion decision vector, and organizing the data structure according to the sample feature dimension to obtain a standardized feature vector; Input the standardized feature vector into the first input channel of the Gaussian kernel support vector classifier, obtain the probability distribution vector of the gas type through feature space mapping and multi-category classification operations, perform maximum probability judgment on the probability distribution vector, perform type mapping based on a preset gas type label library, and output the type identification of the target gas; Inputting the standardized feature vector into the second input channel of the partial least squares regression model, selecting a corresponding regression model according to the type identifier of the target gas, and obtaining data to be predicted; A mixed regularization constraint is applied to the data to be predicted, and feature selection and parameter adjustment are performed through the L1 norm coefficient of 0.3 and the L2 norm coefficient of 0.7 to obtain a regularized feature vector, and the regularized feature vector is input into the selected regression model, and latent variable mapping and linear regression calculation are performed to obtain the concentration value of the target gas.

8. A multi-sensor fusion infrared gas analysis device, characterized in that: The device for performing the infrared gas analysis method of multi-sensor fusion according to any one of claims 1 to 7 comprises: A denoising module is used to perform adaptive digital filtering and multi-scale wavelet denoising operations on the multi-component gas spectral absorption data collected by the multi-band infrared sensor array to obtain a corrected multi-dimensional spectral matrix; A compensation module, used for performing a partition cross-validation compensation operation on the corrected multi-dimensional spectrum matrix, the segmented temperature compensation coefficient matrix and the segmented pressure compensation coefficient matrix, respectively, to obtain a standard spectrum matrix; A construction module is used to construct a covariance matrix according to the standard spectral matrix, extract principal component features and latent variable features by eigenvalue decomposition, and obtain a spectral feature vector; A calculation module, used for inputting the spectral feature vector into a deep autoencoder network model for nonlinear mapping and residual reconstruction calculation to obtain a compensated feature mapping vector; A fusion module, used for performing information fusion on the compensated feature mapping vector to obtain a fusion decision vector; The output module is used to input the fusion decision vector into the Gaussian kernel support vector classifier and the partial least squares regression model respectively, and output the type identification and concentration value of the target gas.

9. A multi-sensor fusion infrared gas analysis device, characterized in that: The multi-sensor fusion infrared gas analysis device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the multi-sensor fusion infrared gas analysis device to perform the multi-sensor fusion infrared gas analysis method according to any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the infrared gas analysis method of multi-sensor fusion as described in any one of claims 1-7 is implemented.

Images