An adaptive calibration method and device for environmental impact correction of coal quality detection

Abstract

The present invention discloses an adaptive calibration method and device for correcting environmental impacts of coal quality detection, belonging to the technical field of coal quality detection. The adaptive calibration method includes (1) spectral preprocessing and (2) LIBS spectral characteristic peak intensity correction; the device includes: a control host, a spectrometer, a laser generator and a turntable sample table. The present invention solves the problems of laser energy fluctuation, unstable spectral data quality and characteristic peak signal intensity being easily disturbed by environmental factors during the use of LIBS detection equipment.

Description

Technical Field

[0001] The invention belongs to the technical field of coal quality detection, and in particular relates to an adaptive calibration method and device for correcting environmental impacts on coal quality detection. Background Art

[0002] Laser induced breakdown spectroscopy (LIBS) is a typical atomic emission spectroscopy technology with the advantages of detecting micro-damage of samples, simultaneous analysis of multiple indicators, wide applicability, and fast detection speed. It shows good application prospects, especially in the field of rapid coal quality detection. However, in actual applications, due to the complex composition of coal and the large differences in the physical and chemical properties of different types of coal, the plasma signal of coal is easily affected by various environmental factors and detection system parameters, resulting in fluctuations in spectral characteristic signals and poor spectral repeatability, resulting in low accuracy of quantitative analysis results. Existing LIBS rapid coal quality detection technologies are mostly based on the stable environment of constant temperature and humidity in the laboratory, and there are few LIBS detection equipment based on the actual complex and harsh environment of power plants.

[0003] Under laboratory conditions, the stability and repeatability of the spectrum are high, and the quantitative analysis model predictions are also relatively accurate. However, when migrating and applying LIBS detection equipment to the actual on-site environment, the LIBS detection equipment often needs to readjust the system measurement parameters and model parameter settings, which will bring great challenges to the measurement accuracy of the LIBS detection device. In addition, the actual environment of the power plant is changeable, and long-term operation will also have an impact on the spectral measurement equipment itself, causing problems such as laser energy fluctuations, spectral line drift, and changes in laser focus position. After the equipment has been running for a long time, the spectrometer needs to be recalibrated and the model parameters need to be adjusted, which is not conducive to the actual situation of long-term startup and operation of coal-fired units, and brings additional operating and maintenance costs to the enterprise. Based on this, it is necessary to improve the existing LIBS coal quality detection device. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide an adaptive calibration method and device for correcting environmental influences on coal quality detection, which solves the problems of laser energy fluctuations, unstable spectral data quality, and characteristic peak signal intensity being easily disturbed by environmental factors during the use of LIBS detection equipment.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is as follows:

[0006] In a first aspect of the present invention, the present invention provides an adaptive calibration method for environmental impact correction of coal quality detection, comprising:

[0007] (1) Spectral preprocessing:

[0008] (1.1) The spectral signal of the sample is de-noised by wavelet transform to reduce the influence of background noise on the spectral characteristic line information, and the interference of factors such as sample surface roughness, different sizes of solid particles and surface scattering on the spectral signal is reduced by standard normal transformation;

[0009] (1.2) averaging multiple spectra into one, and performing spectral averaging on the collected LIBS spectra of the samples;

[0010] (2) Correction of LIBS spectral characteristic peak intensity: The LIBS spectral data are divided into the ideal spectral data measured under experimental conditions, i.e., the original spectral data, and the actual spectral data measured at the production site. The ideal spectral data of the standard sample is measured and collected in advance, and the actual spectral data of the standard sample is measured on site by the LIBS coal quality detection device to correct the spectral data deviation of the LIBS instrument at different time and space, and obtain the corresponding correction factor parameters at different wavelength points, thereby correcting the actual spectral data of the sample to be tested;

[0011] (2.1) Select multiple standard samples, and detect the multiple standard samples using a LIBS instrument in an actual field environment to obtain actual spectral data of the multiple standard samples;

[0012] (2.2) Compare the actual spectrum data of each standard sample with its own ideal spectrum data in each dimension, i.e., wavelength point, to obtain the ratio of the ideal spectrum to the actual spectrum intensity of different standard samples at different wavelength positions, i.e., the correction factor , the calculation formula A is: , where Q represents the ideal spectral intensity of the standard sample, and q represents the actual spectral intensity of the standard sample. Represents different wavelength positions, Represents different standard samples;

[0013] (2.3) Substitute the correction factor obtained in step (2.2) The correction factor applicable to the sample to be tested in the actual environment is calculated by equal weight summation. , calculation formula B is: , where n represents the number of standard samples selected as reference, Represents different wavelength positions, i.e. different wavelength points of the spectrum;

[0014] (2.4) Perform intensity correction on the spectral data of the sample to be tested, that is, calculate formula C: , Represents the actual spectral intensity of the sample to be tested.

[0015] Preferably, the adaptive calibration method for environmental impact correction of coal quality detection also includes: (3) a modeling method based on instance transfer learning is used to retrain to calibrate the prediction model, as follows:

[0016] (3.1) Initial model training settings: select a base learner, use the ideal spectral data of the standard samples measured under experimental conditions and the standard sample information after processing and screening in step (1) as the initial training set, input it into the model for pre-training, and obtain an initial prediction model;

[0017] (3.2) collecting spectral data of a standard sample in an actual environment, obtaining a correction factor through step (2), calibrating the spectral data of the sample to be tested obtained in the actual environment, and then obtaining the test result of the sample to be tested through quantitative analysis using the initial prediction model;

[0018] (3.3) After the model runs for multiple cycles, the model has accumulated a certain amount of actual spectral data of standard samples. The model is retrained and calibrated to convert the ideal spectral data of standard samples in the initial training set, i.e., the source domain data X s The actual spectral data of the standard samples of the initial training set under the actual environment, that is, the target domain data X a Merge to form a new dataset X ;

[0019] (3.4) Constructing a weak learner f i , in the dataset X The spectral data of several standard samples randomly selected in the recent calibration process under actual environment, namely the target field data X a As a test set for the model, X A certain number of standard sample data are randomly selected from the remaining data as training sets and validation sets, and the corresponding sample data extraction information is recorded;

[0020] The training set is used to build the model, the validation set is used to adjust and optimize the model parameters, and the test set is used to measure the prediction performance of the weak learner on the actual spectral data. It is used as the basis for assigning corresponding weights to build the strong learner and complete the calibration of the prediction model. The regression determination coefficient is calculated based on the measurement results obtained by the weak learner. , the calculation formula D is: ,in, and represent the actual inspection reference value and model prediction value of the kth standard sample, respectively. and They represent the average values ​​of the actual test reference value and the model prediction value of the test set, respectively. M is the number of test sets, Representative i The regression coefficient of determination of the test set of weak learners;

[0021] (3.5) Construct N weak learners according to step (3.4) f i , record the data extraction information and regression determination coefficient of each sample R 2 ;

[0022] (3.6) Calculate the weights of each weak learner , according to the weak learners R 2 Calculate the corresponding distribution weights, test set R 2 The larger the weak learner, the greater the weight assigned to it. The calculation is as follows: ;

[0023] (3.7) Weighted ensemble strong learners, after calculating the weights of each weak learner, the individual weak learners obtained by iteration are f i Perform weighted integration to obtain the final strong regression model F(x) , the expansion of the model training set data and the calibration and retraining of the model are completed, and the calculation of the prediction result of the final model is shown in the following formula F: ,in, Representative i The weights of weak learners, Representative i The predicted value of a weak learner.

[0024] Preferably, step (1) also includes: (1.3) eliminating redundant information in the original spectral data by a continuous projection algorithm to improve the validity of the spectral data.

[0025] Preferably, in step (2), the standard samples include standard coal samples and power plant coal samples.

[0026] Preferably, in step (2.1), 10 frames of actual spectral data are collected for each standard sample and averaged into one spectral data as representative of the actual spectral data.

[0027] Preferably, in step (3), the prediction model calibration method is a method of using different data sets to train the model, and the base learner used in the process can adopt any machine learning / deep learning model.

[0028] In a second aspect of the present invention, the present invention provides a device for correcting environmental impacts on coal quality detection, which is a LIBS detection device, applicable to the above-mentioned adaptive calibration method, comprising:

[0029] Control host;

[0030] A spectrometer connected to the control host;

[0031] A laser generator connected to the control host;

[0032] The turntable sample stage comprises a main turntable, a fixed table for the sample to be tested and a calibration sample stage, wherein the main turntable is driven by a motor to rotate, and the fixed table for the sample to be tested and the calibration sample stage are both arranged on the main turntable;

[0033] The laser optical path component includes a laser reflector and a first plano-convex lens, wherein the laser reflector is arranged between the laser generator and the turntable sample stage, and the first plano-convex lens is arranged between the laser reflector and the turntable sample stage. The laser emitted by the laser generator is reflected by the laser reflector to adjust the optical path, and the reflected laser is focused by the first plano-convex lens and then emitted into the turntable sample stage.

[0034] The detection optical path component includes a dichroic mirror and a second plano-convex lens. The dichroic mirror is arranged between the laser reflector and the first plano-convex lens, and the second plano-convex lens is between the dichroic mirror and the spectrometer. The light generated after the laser penetrates the sample returns through the first plano-convex lens, and is reflected by the dichroic mirror and focused by the second plano-convex lens in turn, and then transmitted to the spectrometer by the optical fiber.

[0035] Preferably, the calibration sample stage includes a standard sample stage and a driving mechanism, the driving mechanism is connected to the standard sample stage, a plurality of standard sample positions are arranged on the top surface of the standard sample stage, and the driving mechanism drives the standard sample stage to rotate so that one of the standard samples is punctured by the laser.

[0036] More preferably, the number of the standard sample positions is greater than or equal to two, and a sample stage with a corresponding number of sample positions can be selected according to actual use requirements. Standard samples can be arbitrarily selected on the calibration sample stage and flexibly replaced or added.

[0037] More preferably, the LIBS coal quality detection device also includes an energy monitoring optical path component, which includes a laser window lens and an energy monitoring device. The energy monitoring device is connected to the control host, and the laser window lens is arranged between the laser generator and the laser reflector. The laser is irradiated on the laser window lens, part of the laser is reflected to the energy monitoring device, and the rest of the laser passes through the laser window lens.

[0038] In the third aspect of the present invention, the present invention also proposes an application of an adaptive calibration method for correcting environmental impacts on coal quality detection. The above-mentioned adaptive calibration method is suitable for fields of collecting and analyzing spectral data, including the electric power field, metallurgy field, power machinery field and petrochemical field.

[0039] Beneficial effects:

[0040] The present invention solves the problems of laser energy fluctuation, unstable spectral data quality, characteristic peak signal intensity being easily disturbed by environmental factors, and model prediction accuracy decreasing after long-term operation during use of LIBS detection equipment. With the present invention as a reference, it can also be applied from coal quality LIBS detection to LIBS detection of various objects such as metals, minerals, and biomass. The adaptive calibration scheme can be specially customized according to different detection objects and application scenarios, and then the LIBS detection scheme with high stability and accuracy applicable to multiple scenes and multiple objects can be developed, which is of great significance to the promotion and development of LIBS detection technology. The LIBS detection device with adaptive calibration function proposed by the present invention can be widely used in electric power, metallurgy, power machinery, petrochemical and other industries.

[0041] The present invention reduces the background noise of spectral data through spectral preprocessing, thereby improving the stability and effectiveness of spectral signals; the calibration and correction of spectral data can be completed during continuous operation, and the spectral intensity of coal quality LIBS data in the actual application environment can be corrected to correct the spectral difference caused by interference such as environmental fluctuation factors, thereby improving the characteristic similarity between the actual spectrum and the reference spectral data under the standard environment state, thereby expanding the applicability of the coal quality prediction model in different occasions and different environments;

[0042] The model calibration method constructed by the present invention has high accuracy and stability, and also has adaptive learning ability and expanded application ability. It can extract spectral data with high reference value from the historical spectral database under the actual measurement environment to retrain and adjust the model, thereby improving the applicability and accuracy of the algorithm model in the actual environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Shown is a schematic diagram of the LIBS coal quality detection device of the present invention;

[0044] Figure 2 Shown is a schematic diagram of the calibration method of the present invention;

[0045] Figure 3 The figure shows the comparison of the ideal spectrum in the laboratory, the original spectrum in the actual environment, and the spectrum after correction in the actual environment;

[0046] Figure 4 Shown is a flow chart of the model calibration method;

[0047] Figure 5 Shown is a comparison of the prediction effects of coal calorific value indicators before and after spectrum correction and model calibration.

[0048] Figure numerals: 1-control host, 2-laser generator, 3-spectrometer, 4-main turntable, 5-test sample fixing table, 6-calibration sample table, 7-laser reflector, 8-first plano-convex lens, 9-dichroic mirror, 10-second plano-convex lens, 11-laser window lens, 12-energy monitoring device. DETAILED DESCRIPTION

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings and other implementation methods can be obtained based on these drawings without creative work.

[0050] The existing LIBS coal quality testing equipment is susceptible to interference during the sample spectrum acquisition process. Due to the fluctuation of laser energy and changes in environmental factors such as temperature and humidity, the background spectrum of the LIBS spectrum will change greatly, which in turn affects the signal intensity of the element characteristic spectrum line, which is not conducive to the stability of the overall spectral data quality.

[0051] like Figure 1 As shown, the present invention proposes a device for correcting the environmental impact of coal quality detection, which is a LIBS coal quality detection device, which realizes automatic calibration in the actual application environment and reduces the error of the spectrum detection result. The LIBS coal quality detection device includes a control host 1 and a laser generator 2, a spectrometer 3 and a turntable sample table connected to the control host 1, and also includes a laser optical path component and a detection optical path component. Figure 1 As shown, the laser generator 2 emits a laser, which is injected into the turntable sample table through the laser optical path component, breaks through the sample surface to generate plasma and emit light, and the light generated by the breakdown enters the spectrometer 3 through the detection optical path component, and the spectrometer 3 analyzes the collected spectrum to obtain spectral data information, and controls the host 1 to receive the spectral data information to predict coal quality performance indicators such as calorific value, volatile matter, and ash content. Standard samples for calibration and samples to be tested can be placed on the turntable sample table at the same time. The turntable sample table can adjust the positions of the standard samples and the samples to be tested so that the laser breaks through the standard samples or the samples to be tested. The technical solution of the present invention is described in detail below with a specific structure.

[0052] The laser optical path assembly includes a laser reflector 7 and a first plano-convex lens 8. The laser reflector 7 is disposed between the laser generator 2 and the turntable sample stage, and the first plano-convex lens 8 is disposed between the laser reflector 7 and the turntable sample stage. The laser emitted by the laser generator 2 is reflected by the laser reflector 7 to adjust the optical path, and the reflected laser is focused by the first plano-convex lens 8 and then emitted into the turntable sample stage.

[0053] Preferably, the laser reflector 7 is arranged between the laser generator 2 and the turntable sample stage at an angle of 45°, that is, the laser enters the laser reflector 7 at an incident angle of 45° and exits from the laser reflector 7 at a reflection angle of 45°. It is easy to understand that the first plano-convex lens 8 is arranged parallel to the turntable sample stage, and the reflected laser is collected by the first plano-convex lens 8.

[0054] The detection optical path assembly includes a dichroic mirror 9 and a second plano-convex lens 10. The dichroic mirror 9 is disposed between the laser reflector 7 and the first plano-convex lens 8, and the second plano-convex lens 10 is disposed between the dichroic mirror 9 and the spectrometer 3. The light generated after the laser penetrates the sample returns through the first plano-convex lens 8, and is successively reflected by the dichroic mirror 9 and focused by the second plano-convex lens 10, and then transmitted to the spectrometer 3 by the optical fiber, and the collected spectrum is analyzed in detail.

[0055] In the present invention, the path of light along the laser generator 2, laser reflector 7, dichroic mirror 9, first plano-convex lens 8, sample, first plano-convex lens 8, dichroic mirror 9, second plano-convex lens 10 and spectrometer 3 is the main optical path of laser detection.

[0056] The dichroic mirror 9 is also called a two-color mirror or a two-phase mirror. It almost completely transmits light of a certain wavelength and almost completely reflects light of other wavelengths. Therefore, the dichroic mirror 9 is arranged between the laser reflection mirror 7 and the first plano-convex lens 8. The laser can pass through the dichroic mirror 9 normally and penetrate the sample. The light generated by penetrating the sample cannot pass through the dichroic mirror 9, but is reflected by the dichroic mirror 9 and finally enters the spectrometer 3.

[0057] The turntable sample stage includes a main turntable 4, a sample fixing stage 5 and a calibration sample stage 6. The main turntable 4 is driven by a motor to rotate. The sample fixing stage 5 and the calibration sample stage 6 are arranged on the main turntable 4. The main turntable 4 rotates to align the sample fixing stage 5 with the first plano-convex lens 8, or to align the calibration sample stage 6 with the first plano-convex lens 8.

[0058] It is easy to understand that the main turntable 4 includes a table and a motor drive device, the sample fixing table 5 and the calibration sample table 6 are arranged on the table, and the motor drive device is connected to the table to drive the table to rotate. Figure 1As shown, the top surface of the table is circular, and the axis of rotation of the table is the center of the circle. The turntable sample stage has a detection-calibration dual mode: when the sample to be tested needs to be tested, the main turntable 4 rotates the sample fixing table 5 to align the sample to be tested with the first plano-convex lens 8, and the laser penetrates the sample to be tested along the main optical path of laser detection; when calibration is required, the main turntable 4 rotates the standard sample stage to align the standard sample with the first plano-convex lens 8, and the laser penetrates the standard sample along the main optical path of laser detection, thereby achieving calibration.

[0059] Specifically, when the LIBS coal quality detection device is started for the first time or runs for a period of time, when spectral signal calibration is required, the main turntable 4 rotates to transfer the standard sample to the main optical path of laser detection, and the laser penetrates the standard sample used for calibration. The spectrum generated by the penetration of the standard sample is collected by the spectrometer 3 and input into the control host 1 through the spectral feature calibration algorithm, and the corresponding correction parameters of each wavelength point of the spectrum are obtained, which are used to correct the spectrum of the sample to be tested, so as to correct the influence of environmental factors such as temperature and humidity on the spectral signal when the spectrum calibration is completed; after completing the spectral calibration, the main turntable 4 rotates the fixed table 5 for the sample to be tested, so that the sample to be tested is transferred to the main optical path of laser detection.

[0060] It is easy to understand that when detecting the sample to be tested or the standard sample, the focus of the sample to be tested or the standard sample coincides with the focus of the laser after passing through the first plano-convex lens 8 .

[0061] Furthermore, the calibration sample stage 6 also has a rotation function. The calibration sample stage 6 includes a standard sample stage and a driving mechanism. The driving mechanism is connected to the standard sample stage. A plurality of standard sample positions are arranged on the top surface of the standard sample stage. When detecting the standard sample, the driving mechanism drives the standard sample stage to rotate so that one of the standard samples enters the main optical path of the laser detection.

[0062] In the present invention, the LIBS coal quality detection device also includes an energy monitoring optical path component for stabilizing laser energy. The energy monitoring optical path component includes a laser window lens 11 and an energy monitoring device 12. The energy monitoring device 12 is connected to the control host 1. The laser window lens 11 is arranged between the laser generator 2 and the laser reflector 7. The laser is irradiated on the laser window lens 11, and part of the laser is reflected to the energy monitoring device 12.

[0063] After the laser is emitted by the laser generator 2 and passes through the laser window lens 11, 99% of the laser is directly transmitted, and 1% of the laser is reflected into the energy monitoring device 12 to realize the monitoring of laser energy. The data measured by the energy monitoring device 12 is input into the main control computer, and the energy of the laser that penetrates the sample under test is obtained according to the ratio of the reflection and transmission of the laser in the laser window lens 11. The difference is compared with the preset energy value, and the host computer 1 is controlled to control the laser to adjust the parameters so that the emitted laser energy is controlled within a small fluctuation range.

[0064] like Figure 1 As shown, the laser window lens 11 is set between the laser generator 2 and the laser reflector 7 at an angle of 45°. The laser enters the laser window lens 11 at an incident angle of 45°, and 1% of the laser is emitted at an exit angle of 45° and enters the energy monitoring device 12.

[0065] In the present invention, the type of the energy monitoring device 12 is not limited. For example, the energy monitoring device 12 is a silicon detector energy detection module with high sensitivity to realize the monitoring of laser energy.

[0066] like Figure 2-5 As shown, the present invention proposes an adaptive calibration method for correcting environmental impacts on coal quality detection, which realizes the calibration of LIBS spectral data under different environmental conditions. It should be noted that the adaptive calibration method of the present invention is applicable not only to the above-mentioned LIBS coal quality detection device, but also to fields of collecting and analyzing spectral data, including the electric power field, metallurgy field, power machinery field and petrochemical field, as well as equipment suitable for collecting and analyzing spectral data.

[0067] The adaptive calibration method includes spectral preprocessing and LIBS spectral characteristic peak intensity correction as follows:

[0068] (1) Spectral preprocessing: To improve the quality of spectral data, the raw spectral data needs to be preprocessed before spectral calibration and quantitative analysis.

[0069] (1.1) The spectral signal of the sample is denoised by wavelet transform (WT) to reduce the influence of background noise on the spectral characteristic line information. The standard normal transform (SNV) is used to reduce the interference of factors such as sample surface roughness, different sizes of solid particles and surface scattering on the spectral signal.

[0070] (1.2) In order to improve the stability of the spectral signal, multiple spectra are averaged into one, and the LIBS spectra of the collected samples are spectrally averaged.

[0071] (1.3) The redundant information in the original spectral data is eliminated through the continuous projection algorithm (SPA), the overlap of spectral information is avoided, and the modeling efficiency and prediction accuracy are improved. Specifically, by projecting onto other wavelengths, the size of the projection vector is compared, and the wavelength with the largest projection vector is selected as the candidate wavelength. This vector is used as the projection vector of the next round and calculated and looped until the set number of wavelengths N is reached. Finally, N groups of band combinations are obtained, and multivariate linear regression models are established respectively, and the band combination with the smallest root mean square error in each model is selected as the optimal variable combination.

[0072] (2) Correction of LIBS spectral characteristic peak intensity: The LIBS spectral data are divided into ideal spectral data measured under experimental conditions and actual spectral data measured at the production site. The ideal spectral data of the standard sample is measured and collected in advance, and the actual spectral data of the standard sample is measured on site by the LIBS coal quality detection device to correct the spectral data deviation of the LIBS instrument at different time and space, and obtain the corresponding correction factor parameters at different wavelength points, thereby correcting the actual spectral data of the sample to be tested.

[0073] (2.1) Select multiple standard samples. In an actual field environment, use a LIBS instrument, such as the calibration sample stand 6 of the above-mentioned LIBS coal quality detection device, to detect the multiple standard samples and obtain actual spectral data of the multiple standard samples.

[0074] (2.2) Compare the actual spectrum data of each standard sample with its own ideal spectrum data in each dimension, i.e., wavelength point, to obtain the ratio of the ideal spectrum to the actual spectrum intensity of different standard samples at different wavelength positions, i.e., the correction factor , the calculation formula A is: , where Q represents the ideal spectral intensity of the standard sample, and q represents the actual spectral intensity of the standard sample. Represents different wavelength positions, Represents different standard samples.

[0075] (2.3) Substitute the correction factor obtained in step (2.2) The correction factor applicable to the sample to be tested in the actual environment is calculated by equal weight summation. , calculation formula B is: , where n represents the number of standard samples selected as reference, Represents different wavelength positions, i.e. different wavelength points in the spectrum.

[0076] (2.4) Perform intensity correction on the spectral data of the sample to be tested, that is, calculate formula C: , Represents the actual spectral intensity of the sample to be tested.

[0077] In step (2), the standard samples include standard coal samples and power plant coal samples.

[0078] In step (2.1), 10 frames of actual spectral data are collected for each standard sample and averaged into one spectral data as a representative of the actual spectral data.

[0079] The adaptive calibration method also includes: (3) a modeling method based on instance transfer learning for retraining to calibrate the prediction model, as follows:

[0080] (3) Modeling method based on instance transfer learning is used for retraining to calibrate the prediction model: After completing the correction and calibration of the spectral data, the model is further calibrated using the actual spectral data collected to improve the applicability and accuracy of the coal quality analysis model in the actual environment. The coal quality analysis model is calibrated by the transfer learning method, and the transfer learning can more efficiently and accurately utilize the important feature information in the training set data, so that the generalization and applicability of the model are higher. From the perspective of the transfer learning method, the data can be divided into two categories: one is a considerable amount of source domain data, which in the present invention is the ideal spectral data of the standard sample LIBS of the initial training set collected under the laboratory standard stable state; the other is the target domain data with a relatively small amount, which in the present invention is the actual LIBS spectral data of the standard sample of the initial training set collected under the actual environment of the power plant, which is called the actual spectral database. The instance-based transfer learning algorithm constructs multiple weak learners with the same number of training set samples but different types, and assigns the model weight according to the model prediction error, which indirectly assigns training weights to different samples, and data samples that are not similar and not suitable to the target domain will be assigned lower weights. After iteration, the weight of samples in the source domain that are similar to the target domain increases, while the weight of other source domain data samples decreases, thereby improving the effect of the regression model. The input is the intensity of the spectral line after variable optimization and screening, and the output is the prediction results of indicators such as calorific value and elemental carbon content.

[0081] (3.1) Initial model training settings: select a base learner, use the ideal spectral data of the standard samples measured under experimental conditions and the standard sample information processed and screened in step (1) as the initial training set, input it into the model for pre-training, and obtain an initial prediction model.

[0082] (3.2) Collect spectral data of standard samples in the actual environment, obtain the correction factor through step (2), calibrate the spectral data of the sample to be tested obtained in the actual detection environment, and then obtain the detection result of the sample to be tested through quantitative analysis using the initial prediction model.

[0083] (3.3) After the model runs for multiple cycles, the model has accumulated a certain amount of actual spectral data of standard samples. The model is retrained and calibrated to convert the ideal spectral data of standard samples in the initial training set, i.e., the source domain data X s The actual spectral data of the standard samples of the initial training set under the actual environment, that is, the target domain data X a Merge to form a new dataset X .

[0084] (3.4) Constructing a weak learner f i , in the dataset XThe spectral data of several standard samples randomly selected in the recent calibration process under actual environment, namely the target field data X a As a test set for the model, X A certain number of standard sample data are randomly selected from the remaining data as training sets and validation sets, and the corresponding sample data extraction information is recorded.

[0085] The training set is used to build the model, the validation set is used to adjust and optimize the model parameters, and the test set is used to measure the prediction performance of the weak learner on the actual spectral data. It is used as the basis for assigning corresponding weights to build the strong learner and complete the calibration of the prediction model. The regression determination coefficient is calculated based on the measurement results obtained by the weak learner. , the calculation formula D is: ,in, and represent the actual inspection reference value and model prediction value of the kth standard sample, respectively. and They represent the average values ​​of the actual test reference value and the model prediction value of the test set, respectively. M is the number of test sets. Representative i The regression coefficient of determination on the test set of the weak learner.

[0086] (3.5) Construct N weak learners according to step (3.4) f i , record the data extraction information and regression determination coefficient of each sample R 2 .

[0087] (3.6) Calculate the weights of each weak learner According to the weak learners R 2 Calculate the corresponding distribution weights, test set R 2 The larger the weak learner, the larger the weight assigned to it. The calculation is as follows: .

[0088] (3.7) Weighted ensemble strong learners, after calculating the weights of each weak learner, the individual weak learners obtained by iteration are f i Perform weighted integration to obtain the final strong regression model F(x) , complete the expansion of the model training set data and the calibration and retraining of the model. The calculation of the prediction result of the final model is shown in the following formula F: ,in, Representative i The prediction set determination coefficient of the weak learner is Representative i The predicted value of a weak learner.

[0089] In the present invention, the proposed model calibration method of instance-based transfer learning is a way of constructing a model. The base learner used in the process can be trained based on any machine learning / deep learning model, such as decision tree, support vector machine, neural network, etc.

[0090] It is easy to understand that the LIBS coal quality detection device of the present invention has high flexibility in calibration, and the calibration cycle can be set according to the actual use situation. When rapid calibration is required, the spectral data of the sample to be tested can be corrected through step (2), and the calorific value, elemental carbon content and other indicators of the sample to be tested can be measured according to the corrected spectral data. When deep calibration is required to better improve the prediction accuracy of the device, a modeling method based on transfer learning is performed through step (3), using the accumulated actual spectral data and the original initial training set spectral data to construct a configuration suitable for the actual environment, and realize automatic calibration of the LIBS detection device to improve the prediction accuracy and stability.

[0091] The embodiments provided by the present invention are described in detail above. Specific examples are used herein to describe the principles and implementation methods of the present invention, and the description of the above embodiments is only used to help understand the core idea of ​​the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified in a number of ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. An adaptive calibration method for environmental impact correction of coal quality detection, characterized in that: include: (1) Spectral preprocessing: (1.1) The spectral signal of the sample is de-noised by wavelet transform, and the interference of adverse factors on the spectral signal is reduced by standard normal transform; (1.2) averaging multiple spectra into one, and performing spectral averaging on the collected LIBS spectra of the samples; (1.3) Eliminate redundant information in the original spectral data through a continuous projection algorithm to improve the validity of the spectral data; (2) LIBS spectrum characteristic peak intensity correction: The LIBS spectrum data is divided into ideal spectrum data measured under experimental conditions and actual spectrum data measured at the production site. The ideal spectrum data of the standard sample is measured and collected in advance, and the actual spectrum data of the standard sample is measured on site by the LIBS coal quality detection device to correct the spectrum data deviation of the LIBS instrument at different time and space, and obtain the corresponding correction factor parameters at different wavelength points, thereby correcting the actual spectrum data of the sample to be tested; LIBS spectrum characteristic peak intensity correction includes the following steps: (2.1) Select multiple standard samples, and detect the multiple standard samples using a LIBS instrument in an actual field environment to obtain actual spectral data of the multiple standard samples; (2.2) Compare the actual spectrum data of each standard sample with its own ideal spectrum data in each dimension, i.e., wavelength point, to obtain the ratio of the ideal spectrum to the actual spectrum intensity of different standard samples at different wavelength positions, i.e., the correction factor The calculation formula A is: Among them, Q represents the ideal spectral intensity of the standard sample, q represents the actual spectral intensity of the standard sample, δ represents different wavelength positions, and i represents different standard samples; (2.3) Substitute the correction factor obtained in step (2.2) The correction factor applicable to the sample to be tested in the actual environment is calculated by equal weight summation. The calculation formula B is: Among them, n represents the number of standard samples selected as reference, and δ represents different wavelength positions; (2.4) Perform intensity correction on the spectral data of the sample to be tested, that is, calculate formula C: Represents the actual spectral intensity of the sample to be tested; (3) The modeling method based on instance transfer learning is used to retrain and calibrate the prediction model as follows: (3.1) Initial model training settings: select a base learner, use the ideal spectral data of the standard samples measured under experimental conditions and the standard sample information after processing and screening in step (1) as the initial training set, input it into the model for pre-training, and obtain an initial prediction model; (3.2) collecting spectral data of a standard sample in an actual environment, obtaining a correction factor through step (2), calibrating the spectral data of the sample to be tested obtained in the actual environment, and then obtaining the test result of the sample to be tested through quantitative analysis using the initial prediction model; (3.3) After the model runs for multiple cycles, the model accumulates a certain amount of actual spectral data of standard samples. The model is retrained for correction. The ideal spectral data of standard samples in the initial training set, i.e., the source domain data X s and the actual spectral data of the standard samples of the initial training set under the actual environment, i.e. the target domain data X a Merge to form a new data set X; (3.4) Construct weak learner f i , randomly select several standard samples in the data set X, and obtain the spectral data in the actual environment obtained in the recent calibration process, that is, the target field data X a As the test set of the model, a certain number of standard sample data are randomly selected from the remaining data of data set X as the training set and validation set, and the corresponding sample data extraction information is recorded; The training set is used to build the model, the validation set is used to adjust and optimize the model parameters, and the test set is used to measure the prediction performance of the weak learner on the actual spectral data. It is used as the basis for assigning corresponding weights to build the strong learner and complete the calibration of the prediction model. The regression determination coefficient is calculated based on the measurement results obtained by the weak learner. The calculation formula D is: Among them, y k and represent the actual inspection reference value and model prediction value of the kth standard sample, respectively. and They represent the average values ​​of the actual test reference value and the model prediction value of the test set, M is the number of test sets, represents the regression determination coefficient of the test set of the i-th weak learner; (3.5) According to step (3.4), construct N weak learners f i , record the data extraction information and regression determination coefficient R of each sample 2 ; (3.6) Calculate the weight W of each weak learner i , according to the R of each weak learner 2 Calculate the corresponding distribution weights, test set R 2 The larger the weak learner, the greater the weight assigned to it. i The calculation is as follows: (3.7) Weighted ensemble strong learners, after calculating the weights of each weak learner, the iterative individual weak learners f i Perform weighted integration to obtain the final strong regression model F(x), complete the expansion of the model training set data and the calibration of the model, and the calculation of the prediction result of the final model is shown in the following formula F: Among them, W i represents the weight of the i-th weak learner, z i Represents the predicted value of the i-th weak learner.

2. The adaptive calibration method according to claim 1, characterized in that: In step (2.1), 10 frames of actual spectral data are collected for each standard sample and averaged into one spectral data as a representative of the actual spectral data.

3. The adaptive calibration method according to claim 1 or 2, characterized in that: In step (2), the standard sample includes a standard coal sample.

4. The adaptive calibration method according to claim 1, characterized in that: In step (3), the prediction model is trained based on a machine learning / deep learning model.

5. A device for correcting environmental impacts on coal quality testing, characterized in that: The method for adaptive calibration according to any one of claims 1 to 4 comprises: Control host (1); A spectrometer (3) connected to the control host (1); A laser generator (2) connected to the control host (1); The turntable sample stage comprises a main turntable (4), a fixed table for the sample to be tested (5) and a calibration sample stage (6), wherein the main turntable (4) is driven by a motor to rotate, and the fixed table for the sample to be tested (5) and the calibration sample stage (6) are both arranged on the main turntable (4); The laser optical path component comprises a laser reflection mirror (7) and a first plano-convex lens (8), wherein the laser reflection mirror (7) is arranged between the laser generator (2) and the turntable sample stage, and the first plano-convex lens (8) is arranged between the laser reflection mirror (7) and the turntable sample stage. The laser emitted by the laser generator (2) is reflected by the laser reflection mirror (7) to adjust the optical path, and the reflected laser is focused by the first plano-convex lens (8) and then emitted into the turntable sample stage. The detection optical path component comprises a dichroic mirror (9) and a second plano-convex lens (10), wherein the dichroic mirror (9) is arranged between a laser reflection mirror (7) and a first plano-convex lens (8), and the second plano-convex lens (10) is located between the dichroic mirror (9) and a spectrometer (3). Light generated after the laser penetrates a sample returns through the first plano-convex lens (8), and is sequentially reflected by the dichroic mirror (9) and focused by the second plano-convex lens (10), and then transmitted to the spectrometer (3) by an optical fiber.

6. The device for correcting environmental impact of coal quality detection according to claim 5, characterized in that: The calibration sample stage (6) comprises a standard sample stage and a driving mechanism, wherein the driving mechanism is connected to the standard sample stage, a plurality of standard sample positions are arranged on the top surface of the standard sample stage, and the driving mechanism drives the standard sample stage to rotate so that one of the standard samples is punctured by the laser.

7. The device for correcting environmental impact of coal quality detection according to claim 5 or 6, characterized in that: It also includes an energy monitoring optical path component, which includes a laser window lens (11) and an energy monitoring device (12). The energy monitoring device (12) is connected to the control host (1). The laser window lens (11) is arranged between the laser generator (2) and the laser reflector (7). When laser is irradiated on the laser window lens (11), part of the laser is reflected to the energy monitoring device (12), and the rest of the laser passes through the laser window lens (11).

8. An application of an adaptive calibration method for environmental impact correction of coal quality detection, characterized in that: The adaptive calibration method according to any one of claims 1 to 4 is applied to the field of collecting and analyzing spectral data, including the electric power field, metallurgy field, power machinery field and petrochemical field.

Images