Detection method of SF6 decomposition products by FTIR combined with CARS-ELM

Abstract

The present invention provides a SF6 decomposition product detection method combining FTIR with CARS-ELM, which includes decomposition product spectrum collection, spectrum feature extraction and establishment of a decomposition product quantitative model; the present invention deducts the interference background according to the Lambert-Beer law to extract the effective absorption spectrum of the decomposition product submerged in SF6; adopts the Monte Carlo method to eliminate abnormal samples and establish an effective SF6 decomposition product data set; selects spectral features by the CARS method and establishes an ELM quantitative model of the decomposition product; calculates the relative error and R of the model prediction result 2 To verify the feasibility of the method; the present invention solves the problem of difficulty in quantification caused by severe submergence of the absorption peaks of the decomposition products and overlapping spectra of different decomposition products, and finally realizes high-precision measurement of SF6 decomposition products. At the same time, this method is also suitable for rapid analysis of a large amount of sample data, laying a foundation for the application of FTIR spectroscopy in online quantitative detection of multi-component SF6 decomposition products.

Description

Technical Field

[0001] The invention belongs to the technical field of gas detection and relates to a method for detecting SF6 decomposition products by combining FTIR with CARS-ELM. Background Art

[0002] With the development of UHV technology, sulfur hexafluoride (SF6) has been widely used as an excellent insulating and arc-extinguishing medium in gas insulated switchgear (GIS), the core equipment of UHV power grid. However, the occurrence of partial discharge inside the equipment will cause SF6 to decompose and produce highly corrosive toxic gases. The long-term accumulation of these products will aggravate the occurrence of partial discharge levels and form insulation faults. Studies have shown that different fault types will generate corresponding decomposition products. Therefore, the type and severity of GIS insulation faults can be evaluated by qualitative and quantitative analysis of SF6 decomposition products.

[0003] Fourier transform infrared spectroscopy (FTIR) has been widely used in the field of gas measurement. It has the advantages of fast detection speed, high sensitivity, and no damage to samples. In particular, this method can realize online and repeated measurement of multi-component substances at the same time. Many researchers have carried out relevant research in this regard. Zhang Xiaoxing et al. used FTIR to determine the positions of the absorption peaks of four decomposition products, SOF2, SO2F2, SO2 and CO, and calculated the corresponding absorption coefficients; Liu Xiaobo et al. used this technology to measure the purity of SF6, and the measurement error was kept at 0.04%; Visa et al. designed four different faults and detected the concentrations of different products through this technology; Zheng et al. proposed online detection of SO2 and water vapor in GIS based on FTIR spectroscopy, and realized quantitative analysis of the two substances through partial least squares (PLS).

[0004] Current research results all use traditional PLS to establish a quantitative correction model when using FTIR technology to detect SF6 products. However, SF6 not only has a wide distribution of absorption bands in the mid-infrared region, but also has extremely strong absorption and even reaches saturation, resulting in the absorption peak of the product being seriously submerged. This is the biggest problem in using infrared absorption spectroscopy technology to quantitatively detect SF6 decomposition products. At the same time, the overlapping absorption peaks between different products result in a simple linear relationship between material properties and spectral data. Summary of the invention

[0005] The present invention aims to solve the above problems and defects, and proposes a method for detecting SF6 decomposition products by combining FTIR with CARS-ELM.

[0006] In order to achieve its purpose, the technical solution adopted by the present invention is:

[0007] The SF6 decomposition product detection method using FTIR combined with CARS-ELM includes the following steps:

[0008] Step 1: Collect and process the spectrum data of the SF6 decomposition product to be tested to obtain the effective absorption spectrum data of the mixture;

[0009] Step 2: using Monte Carlo sampling parameters, abnormal samples are screened according to the effective absorption spectrum data of the mixture obtained in step 1 to obtain a valid SF6 decomposition product data set;

[0010] Step 3: Using competitive adaptive re-weighted sampling (CARS) to select spectral features of the data set obtained in step 2;

[0011] Step 4: According to step 3, the optimal variable subset generated by feature selection is obtained, and a spectral data and concentration quantitative analysis model is constructed, and the SF6 decomposition products are quantitatively detected through the model.

[0012] The further improvement of this method is that the step 1 specifically includes the following steps:

[0013] Step 11: Collect multiple sets of data of the product to be tested with SF6 as the background spectrum and different concentrations to obtain effective absorption spectrum data of the mixture;

[0014] Step 12: Perform background subtraction on the spectral data with SF6 as the background spectrum obtained in step 11 according to the Lambert-Beer law to extract the effective characteristic absorption spectrum of the mixture to be tested, which is expressed as follows:

[0015]

[0016] Where S, B and A represent the single-channel absorption spectrum of the standard sample, the single-channel absorption spectrum of the SF6 background and the sample absorbance, respectively.

[0017] Step 13: The absorption spectrum of the product obtained in step 12 is baseline corrected by the concave rubbreband method. The entire spectrum is divided into equal intervals, and the minimum value in the segment is found by sorting according to the absorbance value. This value is regarded as the base point of each segment spectrum. Finally, the baseline point is preliminarily selected in the entire spectrum interval to estimate the baseline by interpolation method to obtain the corrected baseline, and the spectrum data obtained in step 12 is subtracted from the corrected baseline to obtain the effective absorption spectrum data of the mixture.

[0018] The further improvement of this method is that the step 2 specifically includes the following steps:

[0019] Step 21: using a cross-validation method to determine the number of principal components (characteristic variables after dimensionality reduction) for the spectral data obtained after baseline correction in step 13 in a partial least squares method or a principal component regression method;

[0020] Step 22: Divide the reduced-dimensional data set into training and validation sets at random according to the corresponding ratio. Use all sample groups in the training set to fit the regression equation, and substitute the samples in the validation set into the regression equation for prediction, and you can get the prediction error of each validation sample.

[0021] Step 23: Repeat step 22 for a predetermined number of times, and finally obtain the mean and standard deviation of the prediction error of each sample. The calculation expression is:

[0022]

[0023]

[0024] Where k is the total number of times the j-th sample in the validation set is run; error(i) is the prediction error of the j-th sample in the i-th run.

[0025] According to the calculation results, a mean-standard deviation spatial distribution map is constructed. The sample groups that are far away from the sample space and have a large prediction error mean or standard deviation are identified as abnormal samples. These abnormal samples are eliminated to establish a valid SF6 decomposition product data set.

[0026] The further improvement of this method is that the step 3 specifically includes the following steps:

[0027] Step 31: Use the Monte Carlo sampling method to sample the sample characteristic variables in the data set obtained in step 23 N times, and randomly select samples at a ratio of not less than 80% to establish a partial least squares regression model. For example, the spectral data matrix is ​​X (n × p), the sample concentration matrix is ​​Y (n × 1), n ​​is the number of samples, and p is the number of variables, then the partial least squares regression model is:

[0028] Y=Xb+e

[0029] In the formula, the regression coefficient b is a p-dimensional coefficient vector, and e is the residual. The absolute value of the regression coefficient of the i-th element in b | b i |(1≤i≤p) represents the contribution of the i-th variable to the regression model. The larger the value, the more important the corresponding variable is in the prediction of component concentration.

[0030] Step 32: Based on the result calculated in step 31, the exponential decay function is used to forcibly remove |b i|The wavelength with relatively small value. After the i-th sampling operation, the preservation rate calculation expression of the wavelength variable point is as follows:

[0031] r i =ae -ki

[0032] Where a and k are constants.

[0033]

[0034]

[0035] Step 33: Use the adaptive reweighted sampling technique to adjust the weight w of each wavelength variable i Evaluation is performed for variable screening, and the calculation expression of the weight value is:

[0036]

[0037] The wavelength with a larger absolute value weight of the regression coefficient is more likely to be selected, and the wavelength with a smaller absolute value weight is more likely to be eliminated.

[0038] Step 34: After N samplings, N variable subsets are obtained. By calculating and comparing the cross-validation root mean square error of the variable subsets generated by each sampling, the variable subset with the smallest error value is taken as the optimal variable subset.

[0039] The improvement of this method is that the spectrum data and concentration quantitative analysis model in step 4 includes:

[0040] Step 41: Determine the number of neurons in the hidden layer, and randomly set the connection weight a between the input layer and the hidden layer and the bias b of the neurons in the hidden layer;

[0041] N arbitrary data (x i ,t i ), where x i =[x i1 ,x i2 ,…,x xm ] T ∈R n , t i =[t i1 ,t i2 ,…,t im ]∈R m , the extreme learning machine regression model with N hidden layer nodes and activation function G can be expressed as:

[0042]

[0043] In the formula, a i is the weight vector of the i-th hidden layer node and the input node; β iis the weight vector of the i-th hidden node and output node; b i is the bias of the i-th hidden node; is the number of hidden layer nodes.

[0044] Step 42: Select an infinitely differentiable function as the activation function of the hidden layer neurons, and then calculate the hidden layer output matrix H. The expression in step 41 can be simplified as;

[0045] Hβ=T

[0046]

[0047]

[0048] Where H is the hidden layer output matrix, and the i-th column corresponds to the input x1, x2, …, x n The i-th hidden layer outputs the vector.

[0049] Step 43: Output weights according to the hidden layer output results in step 42 It can be obtained by solving the least squares solution of the following linear equations:

[0050]

[0051] The solution is:

[0052]

[0053] In the formula, H + is the Moore-Penrose generalized inverse of the hidden layer output matrix H.

[0054] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is:

[0055] By effectively extracting and selecting features from the infrared absorption spectra of the decomposition products according to the Lambert-Beer law, an extreme learning machine quantitative model was established to achieve quantitative analysis of multiple groups of SF6 products:

[0056] (1) The absorption peaks of decomposition products submerged in high-concentration SF6 were effectively extracted, and an effective SF6 decomposition product data set was established.

[0057] (2) It can realize high-precision quantitative detection of multi-component SF6 decomposition products and quickly analyze a large amount of sample data. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0059] Figure 1 This is a flow chart of a method for detecting SF6 decomposition products by combining FTIR with CARS-ELM according to an embodiment of the present invention;

[0060] Figure 2 A schematic diagram of a gas measuring device according to an embodiment of the present invention;

[0061] Figure 3 Spectral comparison diagram of SF6 decomposition product data before and after preprocessing in the embodiment of the present invention;

[0062] Figure 4 It is a mean-standard deviation distribution diagram of the abnormal sample screening process in the example of the present invention;

[0063] Figure 5 It is a diagram of the feature selection process of the absorption spectrum in the example of the present invention;

[0064] Figure 6 This is a diagram of the prediction results of the traditional PLS quantitative model in the example of the present invention;

[0065] Figure 7 This is a diagram of the prediction results of the CARS-ELM quantitative model in an example of the present invention;

[0066] Among them, 1 is SO2 gas tank; 2, 4, 6 are mass flow meters; 3 is SO2F2 gas tank; 5 is CS2 gas tank: 7 is four-way valve; 8 is infrared light source; 9 is reflector; 10 is aperture; 11, 15 is collimator; 12 is fixed mirror; 13 is beam splitter; 14 is moving mirror; 16, 17 is gas valve; 18 is gas absorption cell; 19 is condenser; 20 is detector; 21 is computer. DETAILED DESCRIPTION

[0067] The present invention is further described in detail below in conjunction with embodiments:

[0068] The embodiment of the present invention provides a SF6 decomposition product detection method combining FTIR with competitive adaptive re-weighted sampling-extreme learning machine (CARS-ELM). The flow chart of the measurement method and the gas measurement device diagram are shown in FIG. Figure 1 , Figure 2 As shown, the specific preferred embodiment of the present invention is described below in conjunction with the flow chart and the measurement device diagram.

[0069] like Figure 1 As shown in the figure, the SF6 decomposition product detection process of FTIR combined with CARS-ELM includes: acquisition and processing of absorption spectrum data of multi-component SF6 mixed gas to be tested, abnormal sample detection, absorption spectrum feature selection, establishment of quantitative calibration model and calculation of relative measurement error.

[0070] In the following examples, the gas to be tested is a mixture of three gases, namely, SO2, SO2F2 and CS2, with SF6 as the background. 141 sets of decomposition spectral data with different concentration gradients were collected, and the model was used to predict the concentration, which was compared with the actual value to verify the accuracy of the model.

[0071] In the following implementation case, the Figure 2 The optical experimental device shown in the figure consists of three parts: gas configuration, gas measurement and data acquisition. The gas configuration part includes: SO2 gas storage tank 1, mass flow meters 2, 4 and 6, SO2F2 gas storage tank 3, CS2 gas storage tank 5, four-way valve 7, the gas measurement part includes: infrared light source 8, reflector 9, diaphragm 10, collimating mirrors 11 and 15, fixed mirror 12, beam splitter 13, moving mirror 14, gas valves 16 and 17, gas absorption cell 18, condenser 19, detector 20, and the data acquisition part includes computer 21.

[0072] 141 groups of SF6 decomposition gas with different concentration gradients are prepared using standard gases with concentrations of 86.9ppm SO2, 99.99ppm SO2F2 and 83.4ppm CS2. The light emitted by the infrared light source 8 is reflected by the collimator 11 as parallel light to the interferometer, and then reflected by the collimator 15, the light passes through the 10cm gas pool 18 filled with the gas to be measured, and finally the light is converged by the condenser 19 to the detector 20 to generate interference spectrum, and the computer 21 completes the Fourier transform, spectrum data processing and model building.

[0073] Based on the above, the specific steps include:

[0074] Step 1: Collection and processing of characteristic absorption spectrum data of the gas to be tested, specifically including the following steps:

[0075] Step 11: Build an optical experimental device for gas measurement based on FTIR;

[0076] Step 12: 141 groups of mixed gases with different concentration gradients are respectively introduced into the gas pool of the constructed gas measurement device. The characteristic absorption spectrum data of the special absorption spectrum of the gas to be measured are collected through the detection of the signal by the detector. Each group of samples is collected 4 times, and then the average value is taken to obtain the initial spectrum of the gas to be measured. Figure 3 As shown in (a), the absorption spectrum of the decomposition product is severely flooded.

[0077] Step 13: Subtract the interference background of the above-mentioned decomposition gas to be tested, take the characteristic absorption spectrum of high-concentration SF6 as interference, and use the Lambert-Beer law to subtract it, the expression is:

[0078]

[0079] Where S, B and A represent the single-channel absorption spectrum of the standard sample, the single-channel absorption spectrum of the SF6 background and the sample absorbance, respectively.

[0080] Step 14: Perform concave rubberband baseline correction based on the product absorption spectrum extracted in step 13. The entire spectrum is divided into equal intervals and sorted according to the absorbance value to find the minimum value in the segment. This value is regarded as the base point of each segment spectrum. Finally, the baseline point is preliminarily selected in the entire spectrum interval and the baseline is estimated by interpolation method to obtain the corrected baseline. The spectrum data obtained in step 13 is subjected to the corrected baseline subtraction to obtain the effective absorption spectrum of the mixture, that is, the effective absorption spectrum of the mixture is as shown in FIG. Figure 3 (b) as shown.

[0081] Step 2: Set the Monte Carlo parameters to detect abnormal sample groups, including the following steps:

[0082] Step 21: using a cross-validation method to determine the number of principal components (characteristic variables after dimensionality reduction) for the spectral data obtained after baseline correction in step 13 in a partial least squares method or a principal component regression method;

[0083] Step 22: The reduced-dimensional data set is randomly divided into a training set and a validation set at a ratio of 80%. All sample groups in the training set are used to fit the regression equation, and the samples in the validation set are substituted into the regression equation for prediction to obtain the prediction error of each validation sample.

[0084] Step 23: Repeat step 22 for a predetermined number of times, and finally obtain the mean and standard deviation of the prediction error of each sample. The calculation expression is:

[0085]

[0086]

[0087] Where k is the total number of times the j-th sample in the validation set is run; error(i) is the prediction error of the j-th sample in the i-th run.

[0088] The spatial distribution of the mean-standard deviation of the 141 groups of three decomposition samples obtained above is as follows: Figure 4 As shown, the sample points outside the dotted line are far away from the sample space or the mean, and the points with large standard deviation are judged as abnormal samples, which are eliminated and a usable data set is established.

[0089] Step 3: According to the competitive adaptive re-weighted sampling (CARS) method, the three decomposition product concentrations are used as labels for the above-mentioned data set in turn, and feature selection is performed using SO2F2 as an example. The calculation process is as follows: Figure 5 As shown. Specifically including the following steps:

[0090] Step 31: Use the Monte Carlo sampling method to sample the characteristic points of each group of samples N times, and randomly select samples at a ratio of 80-90% to establish a partial least squares regression model. Assuming that the spectral data matrix is ​​X (n × p), the sample concentration matrix is ​​Y (n × 1), n ​​is the number of samples, and p is the number of variables, the partial least squares regression model is:

[0091] Y=Xb+e#(4)

[0092] In the formula, the regression coefficient b is a p-dimensional coefficient vector, and e is the residual. The absolute value of the regression coefficient of the i-th element in b | b i |(1≤i≤p) represents the contribution of the i-th variable to the regression model. The larger the value, the more important the corresponding variable is in the prediction of component concentration.

[0093] Step 32: Based on the calculation result of step 31, the exponential decay function is used to forcibly remove |b i |The wavelength with a relatively small value. After the i-th sampling operation, the preservation rate calculation expression of the wavelength variable point is as follows:

[0094] r i =ae -ki #(5)

[0095] Where a and k are constants.

[0096]

[0097] Step 33: Use the adaptive reweighted sampling technique to adjust the weight W of each wavelength variable i The evaluation is performed to select variables, and the weight value is calculated as shown in formula (7):

[0098]

[0099] The wavelength with a larger absolute value weight of the regression coefficient is more likely to be selected, and the wavelength with a smaller absolute value weight is more likely to be eliminated.

[0100] Step 34: After N samplings, N variable subsets are obtained. By calculating and comparing the cross-validation root mean square error of the variable subsets generated by each sampling, the variable subset with the smallest error value is taken as the optimal variable subset.

[0101] Step 4: Establish an extreme learning machine quantitative correction model for spectral data and concentration, which specifically includes the following steps:

[0102] Step 41: Determine the number of neurons in the hidden layer, and randomly set the connection weight a between the input layer and the hidden layer and the bias b of the neurons in the hidden layer;

[0103] N arbitrary data (x i ,t i ), where x i =[x i1 ,x i2 ,…,x in ] T ∈R n , t i =[t i1 ,t i2 ,…,t im ] T ∈R m , the extreme learning machine regression model with N hidden layer nodes and activation function G can be expressed as:

[0104]

[0105] In the formula, a i is the weight vector of the i-th hidden layer node and the input node; β i is the weight vector of the i-th hidden node and output node; b i is the bias of the i-th hidden node; is the number of hidden layer nodes.

[0106] Step 42: Select an infinitely differentiable function as the activation function of the hidden layer neurons, and then calculate the hidden layer output matrix H. The expression of step 41 above can be abbreviated as:

[0107] Hβ=T#(9)

[0108]

[0109]

[0110] Where H is the hidden layer output matrix, and the i-th column corresponds to the input x1, x2, …, x n The i-th hidden layer outputs the vector.

[0111] Step 43: Output weights according to the hidden layer output results in step 42 It can be obtained by solving the least squares solution of the following linear equations:

[0112]

[0113] The solution is:

[0114]

[0115] In the formula, H + is the Moore-Penrose generalized inverse of the hidden layer output matrix H.

[0116] A specific embodiment is provided to illustrate the technical solution and technical effect of the present invention. In the traditional PLS process of measuring gas concentration, only the absorption band peak is used to invert the gas concentration. The measurement result is as follows: Figure 6 The measurement results of FTIR combined with CARS-ELM in this paper are shown in Figure 7 As shown, the comparison of the prediction results of the two methods fully proves that the CARS-ELM method proposed in the present invention meets the requirement of the SF6 decomposition product detection standard that the relative error does not exceed 2%.

[0117] The concentration prediction results of the SF6 decomposition product detection method combined with FTIR and CARS-ELM are shown in Table 1. It shows the measurement results of 17 groups of different concentration gradients in the test set and verifies the effectiveness of the method; in the above embodiment, the maximum relative error of the 17 groups of different concentration groups measured by the present invention does not exceed 1.7%, and the maximum absolute error does not exceed 1.06ppm, and the relevant determination coefficient R 2 The experimental results show that the SF6 decomposition product measurement method of FTIR combined with CARS-ELM proposed in the present invention effectively extracts the absorption peak of the decomposition product, obtains more accurate measurement results of multi-component trace SF6 decomposition products, and realizes the rapid analysis of a large amount of sample data.

[0118] Table 1: Prediction results of SF6 decomposition products by FTIR combined with CARS-ELM

[0119]

[0120] The embodiments described above are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the protection scope determined by the claims of the device of the present invention.

Claims

1. FTIR combined with CARS-ELM SF6 decomposition product detection method, characterized in that: It includes the following steps: Step 1: Collect and process the spectrum data of the SF6 decomposition product to be tested to obtain the effective absorption spectrum data of the mixture; Step 2: using Monte Carlo sampling parameters, perform abnormal sample screening based on the effective absorption spectrum data of the mixture obtained in step 1 to obtain a valid SF6 decomposition product data set; Step 3: Using a competitive adaptive reweighted sampling method to select spectral features of the data set obtained in step 2; Step 4: According to step 3, the optimal variable subset generated by feature selection is obtained, a spectrum data and concentration quantitative analysis model is constructed, and the SF6 decomposition products are quantitatively detected through the model; The step 1 specifically includes the following steps: Step 11: Collect multiple sets of data of the product to be tested with SF6 as the background spectrum and different concentrations to obtain the characteristic absorption spectrum of the mixed gas; Step 12: Perform background subtraction on the spectral data with SF6 as the background spectrum obtained in step 11 according to the Lambert-Beer law to extract the effective characteristic absorption spectrum of the mixture to be tested, which is expressed as follows: In the formula, S, B and A represent the single-channel absorption spectrum of the standard sample, the single-channel absorption spectrum of the SF6 background and the absorbance of the sample, respectively; Step 13: The absorption spectrum of the product obtained in step 12 is baseline corrected by a concave rubbreband method; the entire spectrum is divided into equal intervals, and the minimum value in the segment is found according to the size of the absorbance value, and this value is regarded as the base point of each segment spectrum; finally, the baseline point is preliminarily selected in the entire spectrum interval and the baseline is estimated by interpolation method to obtain a corrected baseline; the spectrum data obtained in step 12 is subjected to the corrected baseline subtraction to obtain the effective absorption spectrum data of the mixture; The spectrum data and concentration quantitative analysis model in step 4 includes: Step 41: Determine the number of neurons in the hidden layer, and randomly set the connection weight a between the input layer and the hidden layer and the bias b of the neurons in the hidden layer; N arbitrary data (x i ,t i ), where x i =[x i1 ,x i2 ,…,x xm ] T ∈R n , t i =[t i1 ,t i2 ,…,t im ]∈R m , the extreme learning machine regression model with N hidden layer nodes and activation function G can be expressed as: In the formula, a i is the weight vector of the i-th hidden layer node and the input node; β i is the weight vector of the i-th hidden node and output node; b i is the bias of the i-th hidden node; is the number of hidden layer nodes; Step 42: Select an infinitely differentiable function as the activation function of the hidden layer neurons, and then calculate the hidden layer output matrix H. The expression in step 41 can be simplified as; Hβ=T Where H is the hidden layer output matrix, and the i-th column corresponds to the input x1, x2, …, x n The i-th hidden layer output vector; Step 43: Output weights according to the hidden layer output results in step 42 It can be obtained by solving the least squares solution of the following linear equations: The solution is: In the formula, H + is the Moore-Penrose generalized inverse of the hidden layer output matrix H.

2. The SF6 decomposition product detection method of FTIR combined with CARS-ELM according to claim 1, characterized in that: The step 2 specifically includes the following steps: Step 21: Determine the number of principal components using a cross-validation method on the spectral data obtained after baseline correction in step 13 in a partial least squares method or a principal component regression method; Step 22: Randomly divide the reduced-dimensional data set into a training set and a validation set according to the corresponding ratio; use all sample groups in the training set to fit the regression equation, and substitute the samples in the validation set into the regression equation for prediction, so as to obtain the prediction error of each validation sample; Step 23: Repeat step 22 for a predetermined number of times, and finally obtain the mean and standard deviation of the prediction error of each sample. The calculation expression is: Where k is the total number of times the jth sample in the validation set is run; error(i) is the prediction error of the jth sample in the i-th run; According to the calculation results, a mean-standard deviation spatial distribution map is constructed. The sample groups that are far away from the sample space and have a large prediction error mean or standard deviation are identified as abnormal samples. These abnormal samples are eliminated to establish a valid SF6 decomposition product data set.

3. The SF6 decomposition product detection method of FTIR combined with CARS-ELM according to claim 2, characterized in that: The step 3 specifically includes the following steps: Step 31: The sample characteristic variables in the data set obtained in step 23 are sampled N times using the Monte Carlo sampling method, and samples are randomly selected at a ratio of not less than 80% to establish a partial least squares regression model; the spectral data matrix is ​​X (n × p), the sample concentration matrix is ​​Y (n × 1), n ​​is the number of samples, and p is the number of variables, then the partial least squares regression model is: Y=Xb+e In the formula, the regression coefficient b is a p-dimensional coefficient vector, e is the residual; the absolute value of the regression coefficient of the i-th element in b | b i |(1≤i≤p) represents the contribution of the i-th variable to the regression model. The larger the value, the more important the corresponding variable is in the prediction of component concentration. Step 32: Based on the result calculated in step 31, the exponential decay function is used to forcibly remove |b i |The wavelength with a relatively small value; after the i-th sampling operation, the preservation rate calculation expression of the wavelength variable point is as follows: r i =ae -ki In the formula, a and k are constants; Step 33: Use the adaptive reweighted sampling technique to adjust the weight w of each wavelength variable i Evaluation is performed for variable screening, and the calculation expression of the weight value is: The wavelength with a larger absolute value weight of the regression coefficient is more likely to be selected, and the wavelength with a smaller absolute value weight is more likely to be eliminated; Step 34: After N samplings, N variable subsets are obtained. By calculating and comparing the cross-validation root mean square error of the variable subsets generated by each sampling, the variable subset with the smallest error value is taken as the optimal variable subset.

Images