Small sample coal quality characteristics analysis system and method based on deep transfer learning

Abstract

The present invention discloses a small sample coal quality characteristic analysis system and method based on deep transfer learning, including taking a large number of coal samples with known coal quality characteristics and different sources from the samples to be tested as source model calibration samples, detecting the calibration samples of the source model to obtain the characteristic spectrum and characteristic spectral line intensity of the source model calibration samples; taking the coal quality characteristics as the target characteristics, establishing a neural network source model; taking a small number of coal samples with known coal quality characteristics and the same source as the samples to be tested as the calibration samples of the target model, detecting the calibration samples of the target model to obtain the characteristic spectrum and characteristic spectral line intensity of the target model calibration samples; establishing a neural network target model; detecting the samples to be tested with unknown coal quality characteristics to obtain the characteristic spectrum of the samples to be tested, and substituting the characteristic spectrum into the trained neural network source model to obtain the numerical value of the target characteristic in the samples to be tested. The present invention has high generalization performance and can effectively improve the analysis accuracy in small sample scenarios.

Description

Technical Field

[0001] The present invention relates to a coal quality analysis system and method, and in particular to a small sample coal quality characteristic analysis system and method based on deep transfer learning. Background Art

[0002] At present, the commonly used coal quality detection method in coal mines, coal plants and power plants is offline sampling and laboratory analysis. The existing coal quality detection method is complicated, time-consuming, and has poor sampling representativeness, making it difficult to timely feedback the various components of coal. However, coal users need to timely control the composition of coal on the belt conveyor in order to guide production and conduct regulation. Therefore, traditional offline measurement is difficult to meet the needs of industrial production.

[0003] At present, the technologies used in online coal quality detection are X-ray fluorescence technology, neutron-induced prompt gamma-ray analysis technology and dual-energy gamma-ray transmission technology. However, X-ray fluorescence technology is only suitable for measuring elements with atomic numbers greater than 11, and its measurement accuracy and sensitivity are not high. Neutron-induced prompt gamma-ray analysis technology has the disadvantages of high cost, radiation hazard and short half-life of the radioactive source. The biggest disadvantage of dual-energy gamma-ray projection technology is that it cannot analyze all elements, has high cost and safety risks.

[0004] In recent years, laser induced plasma spectroscopy (LIBS) has become a new laser analysis technology due to its advantages such as high sensitivity, no need for sample pretreatment and multi-element measurement. It has great application potential in online coal quality detection. However, due to the obvious matrix effect of this technology, its application in online coal quality detection is also limited. Accurate quantitative measurement is the premise and basis for the LIBS system to play a role in online coal quality detection. Since quantitative models, especially complex models such as neural networks, require a large number of calibration samples for modeling and training, and the acquisition cost and time of the calibration value of the sample are high, and the process of the national standard coal quality analysis test method is as long as dozens of hours, it is impossible to quickly accumulate a large number of calibration samples for quantitative model modeling. In addition, different types of samples, different measurement environments, system drift and other problems will have a certain impact on the spectrum measured by the LIBS system, so that the sample to be tested cannot fall into the applicable domain of the model established with previous sample data.

[0005] Therefore, it is urgent to solve the above problems. Summary of the invention

[0006] Purpose of the invention: The first purpose of the present invention is to provide a small sample coal quality characteristics analysis method based on deep transfer learning that has high generalization performance and effectively improves the analysis accuracy in small sample scenarios.

[0007] The second object of the present invention is to provide a small sample coal quality characteristic analysis system based on deep transfer learning with small target sample size and high accuracy.

[0008] Technical solution: To achieve the above objectives, the present invention discloses a small sample coal quality characteristic analysis method based on deep transfer learning, comprising the following steps:

[0009] A large number of coal samples with known coal quality characteristics and different sources from the samples to be tested are used as source model calibration samples, and the source model calibration samples are tested to obtain characteristic spectra of the source model calibration samples and characteristic spectral line intensities of various elements;

[0010] Taking a certain coal quality characteristic as the target characteristic, a neural network source model M is established for the target characteristic of the source model calibration sample. s The source model is used to calibrate the characteristic spectrum of the sample and the characteristic spectral line intensity of various elements to obtain the trained neural network source model M s ;

[0011] A small amount of coal samples with known coal quality characteristics and the same source as the sample to be tested are used as calibration samples of the target model, and the calibration samples of the target model are tested to obtain the characteristic spectrum of the calibration samples of the target model and the characteristic spectral line intensities of various elements;

[0012] The neural network target model M is established based on the target characteristics of the target model calibration sample. t , using the trained neural network source model M s The characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements are trained to obtain the trained neural network source model M t ;

[0013] The sample with unknown coal quality characteristics is tested to obtain the characteristic spectrum of the sample, that is, the spectral line intensity I of the sample is obtained. f , the spectral line intensity I f Substitute the trained neural network source model M t , and obtain the numerical value of the target characteristic in the sample to be tested.

[0014] Among them, the characteristic spectrum of the source model calibration sample and the characteristic spectral line intensity of various elements are used for training to obtain the trained neural network source model M s The specific method is: the characteristic spectrum of the source model calibration sample is used as the independent variable matrix E s , with the target characteristic values ​​of the source model calibration samples as the dependent variable matrix F s , train the neural network source model M s Until the prediction accuracy reaches the first preset value, the trained neural network source model M is obtained. s .

[0015] Preferably, the trained neural network source model M is used sThe characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements are trained to obtain the trained neural network source model M t The specific method is: the trained neural network source model M s The convolution layer structure, normalization layer structure and pooling layer structure of the neural network are copied to the neural network target model M t The corresponding position of the neural network target model M t The rest of the matrix is ​​randomly initialized; the characteristic spectrum of the target model calibration sample is used as the independent variable matrix E t , with the target characteristic values ​​of the target model calibration samples as the dependent variable matrix F t , in the neural network target model M t Based on this, fine-tune the parameters, reduce the initial learning rate, increase the regularization coefficient, and train the neural network target model M t When the prediction accuracy reaches the second preset value, the trained neural network source model M is obtained. t .

[0016] Furthermore, coal properties include carbon content, ash, volatile matter, calorific value and moisture.

[0017] Furthermore, the number of calibration samples of the target model is within 100.

[0018] The invention discloses a small sample coal quality characteristic analysis system based on deep transfer learning, comprising a laser induced breakdown spectroscopy system for coal quality detection, a neural network source model module, a neural network target model module and a sample analysis module to be tested, wherein

[0019] The laser-induced breakdown spectroscopy system detects a large number of coal samples with known coal quality characteristics and different sources from the samples to be tested to obtain the characteristic spectrum of the source model calibration sample and the characteristic spectral line intensity of various elements, and transmits them to the neural network source model module;

[0020] The laser-induced breakdown spectroscopy system detects a small amount of coal samples with known coal quality characteristics and the same source as the sample to be tested to obtain the characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements, and transmits them to the neural network target model module;

[0021] The laser induced breakdown spectroscopy system detects the unknown coal quality characteristics of the sample to be tested to obtain the characteristic spectrum of the sample to be tested, that is, the spectral line intensity I of the sample to be tested is obtained. f , and transmit it to the sample analysis module to be tested;

[0022] The neural network source model module is used to receive the characteristic spectrum of the source model calibration sample and the characteristic spectral line intensity of various elements, take a certain coal quality characteristic as the target characteristic, and establish a neural network source model M for the target characteristic of the source model calibration sample. sThe source model is used to calibrate the characteristic spectrum of the sample and the characteristic spectral line intensity of various elements to obtain the trained neural network source model M s ;

[0023] The neural network target model module is used to receive the characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements, and establish a neural network target model M for the target characteristics of the target model calibration sample. t , using the trained neural network source model M s The characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements are trained to obtain the trained neural network source model M t ;

[0024] The sample analysis module is used to convert the spectral line intensity I f Substitute the trained neural network source model M t , and obtain the numerical value of the target characteristic in the sample to be tested.

[0025] Among them, coal quality characteristics include carbon content, ash content, volatile matter, calorific value and moisture.

[0026] Preferably, the neural network source model module uses the characteristic spectrum of the source model calibration sample as the independent variable matrix E s , with the target characteristic values ​​of the source model calibration samples as the dependent variable matrix F s , train the neural network source model M s Until the prediction accuracy reaches the first preset value, the trained neural network source model M is obtained. s .

[0027] Furthermore, the neural network target model module converts the trained neural network source model M s The convolution layer structure, normalization layer structure and pooling layer structure of the neural network are copied to the neural network target model M t The corresponding position of the neural network target model M t The rest of the matrix is ​​randomly initialized; the characteristic spectrum of the target model calibration sample is used as the independent variable matrix E t , with the target characteristic values ​​of the target model calibration samples as the dependent variable matrix F t , in the neural network target model M t Based on this, fine-tune the parameters, reduce the initial learning rate, increase the regularization coefficient, and train the neural network target model M t When the prediction accuracy reaches the second preset value, the trained neural network source model M is obtained. t .

[0028] Furthermore, the laser induced breakdown spectroscopy system includes a pulsed laser, a focusing lens located on the light transmission path of the laser emitted by the pulsed laser and used to focus the laser to act on the surface of the calibration sample to generate luminescent plasma, a collection lens for sensing the light information radiated by the luminescent plasma, and a spectrometer that receives light information through optical fiber and generates electrical signals to transmit to a computer for coal quality characteristic analysis.

[0029] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0030] (1) Spectral data from sources different from the samples to be tested are used to improve the modeling effect in small sample scenarios. First, the neural network uses a large amount of data from different sources to train the source model. In the second step, the source model is transferred to the target model and fine-tuned with small sample data of the target task. To a certain extent, this improves the disadvantage that neural network training requires a large amount of calibration data and solves the problem that traditional quantitative models require a large number of calibration samples. Compared with traditional quantitative models, this improves the generalization ability of the model and its accuracy in small sample scenarios, and reduces the number of samples required for the calibration model.

[0031] (2) It can be used on the laser induced plasma spectroscopy system to solve the problem of rapid online analysis of coal quality. Based on similar measurement conditions and the intrinsic connection of plasma spectral signals, the accumulated data in the past that cannot be directly used for new data modeling is utilized, and the generalization performance of the neural network is improved through transfer learning. The correlation between past data and the data to be measured is fully utilized to improve the generalization performance of the model and the accuracy of the model when the number of calibration samples is extremely small. This method fully utilizes the similarity of laser induced plasma spectra and the transferability of neural networks, and is easy to implement quickly on computers. It can not only perform online full-element analysis of coal quality, but also further improve measurement accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the structure of the laser induced breakdown spectroscopy system in the present invention;

[0033] Figure 2 It is a schematic diagram of the process of the present invention;

[0034] Figure 3 The analysis result of the carbon content calibration model in the present invention;

[0035] Figure 4 This is the analysis result of the volatile matter calibration model in the present invention. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0037] Example 1

[0038] like Figure 2 As shown, the present invention provides a small sample coal quality characteristic analysis method based on deep transfer learning, comprising the following steps:

[0039] A large number of coal samples with known coal quality characteristics and different sources from the samples to be tested are used as source model calibration samples, and the source model calibration samples are tested to obtain the characteristic spectrum of the source model calibration samples and the characteristic spectral line intensity of various elements; the coal quality characteristics include carbon content, ash content, volatile matter, calorific value and moisture;

[0040] Taking a certain coal quality characteristic as the target characteristic, a neural network source model M is established for the target characteristic of the source model calibration sample. s The source model is used to calibrate the characteristic spectrum of the sample and the characteristic spectral line intensity of various elements to obtain the trained neural network source model M s ; That is, the characteristic spectrum of the source model calibration sample is used as the independent variable matrix F s , with the target characteristic values ​​of the source model calibration samples as the dependent variable matrix F s , train the neural network source model M s Until the prediction accuracy reaches the first preset value, the trained neural network source model M is obtained. s ;

[0041] A small sample of coal with known coal quality characteristics and the same source as the sample to be tested is used as a calibration sample of the target model, and the calibration sample of the target model is tested to obtain the characteristic spectrum of the calibration sample of the target model and the characteristic spectral line intensity of various elements;

[0042] The neural network target model M is established based on the target characteristics of the target model calibration sample. t , using the trained neural network source model M s The characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements are trained to obtain the trained neural network source model M t ; The neural network source model M to be trained s The convolution layer structure, normalization layer structure and pooling layer structure of the neural network are copied to the neural network target model M t The corresponding position of the neural network target model M t The rest of the matrix is ​​randomly initialized; the characteristic spectrum of the target model calibration sample is used as the independent variable matrix E t , with the target characteristic values ​​of the target model calibration samples as the dependent variable matrix F t , in the neural network target model M t Based on this, fine-tune the parameters, reduce the initial learning rate, increase the regularization coefficient, and train the neural network target model M t When the prediction accuracy reaches the second preset value, the trained neural network source model M is obtained. t;

[0043] The sample with unknown coal quality characteristics is tested to obtain the characteristic spectrum of the sample, that is, the spectral line intensity I of the sample is obtained. f , the spectral line intensity I f Substitute the trained neural network source model M t , and obtain the numerical value of the target characteristic in the sample to be tested.

[0044] The present invention is applicable to the situation where the number of target model calibration samples is very small and insufficient to complete the training of existing conventional neural networks; when the number of target model calibration samples with calibration values ​​is less than 100, the method of the present invention can be used for training to obtain better results. In the present invention, it is required that the source model calibration samples and the target model calibration samples have a strong correlation, such as using a laser induced breakdown spectroscopy system of the same model and configuration and sampling under the same experimental conditions, or the sample category, elemental composition and matrix of the source model calibration samples and the target model calibration samples are relatively similar, and the cumulative data volume of the source model calibration samples is large enough, generally not less than hundreds of calibration samples.

[0045] The principle of the present invention is: the shallow structure of the neural network after training has the function of a feature extractor, which can extract general features from the input spectrum; the source model calibration sample has a strong similarity with the target model calibration sample, so the shallow source model trained on more data is migrated to the target model as a feature extractor, and the effect of strengthening the target model with the source model data can be achieved, thereby improving the generalization performance of the target model and the prediction accuracy in a small sample situation. The present invention utilizes the features learned by the shallow structure of the neural network, and applies it to the modeling process of the data to be tested based on the correlation between previous data and the data to be tested; compared with the existing methods, the present invention makes full use of previous data that cannot be directly used by conventional methods, improves the generalization performance of the neural network in a small sample situation, thereby improving the accuracy of the neural network on the sample to be tested, and has better measurement accuracy than traditional methods in a wider range of coal quality characteristics.

[0046] Example 2

[0047] like Figure 1As shown, a small sample coal quality characteristic analysis system based on deep transfer learning of the present invention includes a laser induced breakdown spectroscopy system and a computer 7, wherein the laser induced breakdown spectroscopy system is used for online detection of coal quality, and the computer is used for coal quality characteristic analysis. The laser induced breakdown spectroscopy system includes a pulse laser 1, a focusing lens 2, a collection lens 4, an optical fiber 5 and a spectrometer 6. First, a coal sample with various known coal quality characteristics is used as a calibration sample, and each calibration sample is detected separately using a laser induced breakdown spectroscopy system installed above the coal conveyor belt. The coal quality characteristics include carbon content, ash content, volatile matter, calorific value and moisture. With a pulse laser 1 as an excitation light source, the laser emitted from the laser is focused by a focusing lens 2 and acts on the surface of the calibration sample 3, generating plasma at the focus point, and the plasma is cooled in an atmosphere of protective gas. The generated radiation light signal enters the optical fiber 5 through the collection lens 4, and is converted into an electrical signal after being processed by the spectrometer 6 and collected by the computer 7. Computer 7 obtains characteristic spectra of two groups of calibration samples. The first group is a large number of samples accumulated in past measurements and having a different source from the external measurement samples, called source model calibration samples. The second group is calibration samples having the same source as the samples to be measured, called target model calibration samples. The characteristic spectral line intensities of various elements are obtained from the characteristic spectra of the source model calibration samples and the target model calibration samples, respectively.

[0048] Computer 7 includes a neural network source model module, a neural network target model module and a sample analysis module; wherein the laser induced breakdown spectroscopy system detects a large number of coal samples with known coal quality characteristics and different sources from the sample to be tested to obtain the characteristic spectrum of the source model calibration sample and the characteristic spectral line intensity of various elements, and transmits them to the neural network source model module;

[0049] The laser-induced breakdown spectroscopy system detects a small amount of coal samples with known coal quality characteristics and the same source as the sample to be tested to obtain the characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements, and transmits them to the neural network target model module;

[0050] The laser induced breakdown spectroscopy system detects the unknown coal quality characteristics of the sample to be tested to obtain the characteristic spectrum of the sample to be tested, that is, the spectral line intensity I of the sample to be tested is obtained. f , and transmit it to the sample analysis module to be tested;

[0051] The neural network source model module is used to receive the characteristic spectrum of the source model calibration sample and the characteristic spectral line intensity of various elements, take a certain coal quality characteristic as the target characteristic, and establish a neural network source model M for the target characteristic of the source model calibration sample. s The source model is used to calibrate the characteristic spectrum of the sample and the characteristic spectral line intensity of various elements to obtain the trained neural network source model M sThe neural network source model module uses the characteristic spectrum of the source model calibration sample as the independent variable matrix E s , with the target characteristic values ​​of the source model calibration samples as the dependent variable matrix F s , train the neural network source model M s Until the prediction accuracy reaches the first preset value, the trained neural network source model M is obtained. s ;

[0052] The neural network target model module is used to receive the characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements, and establish a neural network target model M for the target characteristics of the target model calibration sample. t , using the trained neural network source model M s The characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements are trained to obtain the trained neural network source model M t ; The neural network target model module converts the trained neural network source model M s The convolution layer structure, normalization layer structure and pooling layer structure of the neural network are copied to the neural network target model M t The corresponding position of the neural network target model M t The rest of the matrix is ​​randomly initialized; the characteristic spectrum of the target model calibration sample is used as the independent variable matrix E t , with the target characteristic values ​​of the target model calibration samples as the dependent variable matrix F t , in the neural network target model M t Based on this, fine-tune the parameters, reduce the initial learning rate, increase the regularization coefficient, and train the neural network target model M t When the prediction accuracy reaches the second preset value, the trained neural network source model M is obtained. t ;

[0053] The sample analysis module is used to convert the spectral line intensity I f Substitute the trained neural network source model M t , and obtain the numerical value of the target characteristic in the sample to be tested.

[0054] Example 3

[0055] Analyzing coal quality characteristics of a group of coal samples at a power plant

[0056] 1) In this example, 339 coal samples are used as source data set training samples, and 10 coal samples are used as target data set calibration samples. The coal quality characteristics of the target data set calibration samples are analyzed by traditional offline analysis as shown in Table 1:

[0057] Table 1: Coal quality characteristics

[0058]

[0059]

[0060] The coal samples are ground and pressed into cakes, and then tested using a laser induced plasma spectroscopy system. A pulsed laser 1 is used as the excitation light source. The laser emitted from the laser is focused by a focusing lens 2 and acts on the surface of the coal sample 3. Plasma is generated at the focus point, and the plasma is cooled in an air atmosphere. The generated radiation light signal is collected in real time by a collection lens 4, and is converted into an electrical signal after being processed by an optical fiber 5 and a spectrometer 6 and collected by a computer 7. The characteristic spectra of a group of coal samples with known mass concentrations of each element are obtained, and the characteristic spectral line intensities of the laser induced breakdown spectra of various elements in the coal samples are further obtained. According to a coal quality characteristic analysis method based on deep transfer learning of the present invention, calibration models related to carbon content and sulfur content are established respectively, and the entire calculation process is implemented using a Python program. Due to too many variables, it is not convenient to list the fitting formulas here, and only the calculated average relative error is displayed. Among them, the average relative error of the carbon content is 6.78%, and the analysis results of the carbon content calibration model are as follows. Figure 3 As shown; the average relative error of volatile content is 7.31, and the analysis results of the volatile calibration model are as follows Figure 4 Show.

Claims

1. A small sample coal quality characteristics analysis method based on deep transfer learning, characterized in that: The steps include: A large number of coal samples with known coal quality characteristics and different sources from the samples to be tested are used as source model calibration samples, and the source model calibration samples are tested to obtain characteristic spectra of the source model calibration samples and characteristic spectral line intensities of various elements; Taking a certain coal quality characteristic as the target characteristic, a neural network source model is established for the target characteristic of the source model calibration sample. The source model is used to calibrate the characteristic spectrum of the sample and the characteristic spectral line intensity of various elements to obtain the trained neural network source model. ; A small amount of coal samples with known coal quality characteristics and the same source as the sample to be tested are used as calibration samples of the target model, and the calibration samples of the target model are tested to obtain the characteristic spectrum of the calibration samples of the target model and the characteristic spectral line intensities of various elements; Establish a neural network target model based on the target characteristics of the target model calibration sample , using the trained neural network source model The characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements are trained to obtain the trained neural network source model ; The sample with unknown coal quality characteristics is tested to obtain the characteristic spectrum of the sample, that is, the spectral line intensity of the sample is obtained. I f , the spectral line intensity I f Substitute the trained neural network source model , obtain the value of the target characteristic in the sample to be tested; Coal quality characteristics include carbon content, ash content, volatile matter, calorific value and moisture. A pulsed laser is used as the excitation light source. The laser emitted from the laser is focused by a focusing lens and acts on the surface of the calibration sample. Plasma is generated at the focusing point. The plasma is cooled in an atmosphere of protective gas. The generated radiation light signal enters the optical fiber through a collection lens and is converted into an electrical signal after being processed by a spectrometer and collected by a computer. The computer obtains two groups of characteristic spectra of calibration samples. The first group is a large number of samples accumulated in past measurements and having a different source from the external measurement samples, which are called source model calibration samples. The second group is calibration samples that are the same source as the samples to be measured, which are called target model calibration samples. The characteristic spectral line intensities of various elements are obtained from the characteristic spectra of the source model calibration samples and the target model calibration samples, respectively.

2. According to claim 1, a method for analyzing coal quality characteristics of a small sample based on deep transfer learning is characterized by: The method uses the source model to calibrate the characteristic spectrum of the sample and the characteristic spectral line intensity of various elements to obtain the trained neural network source model The specific method is: the characteristic spectrum of the source model calibration sample is used as the independent variable matrix , with the target characteristic values ​​of the source model calibration samples as the dependent variable matrix , train the neural network source model Until the prediction accuracy reaches the first preset value, the trained neural network source model is obtained. .

3. The method for analyzing coal quality characteristics of a small sample based on deep transfer learning according to claim 1, characterized in that: The trained neural network source model is used The characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements are trained to obtain the trained neural network source model The specific method is: the trained neural network source model The convolution layer structure, normalization layer structure and pooling layer structure are copied to the neural network target model respectively. The corresponding position of the neural network target model The rest of the matrix is ​​randomly initialized; the characteristic spectrum of the target model calibration sample is used as the independent variable matrix , with the target characteristic values ​​of the target model calibration samples as the dependent variable matrix , in the neural network target model Based on this, fine-tune the parameters, reduce the initial learning rate, increase the regularization coefficient, and train the neural network target model. Until the prediction accuracy reaches the second preset value, the trained neural network source model is obtained .

4. The method for analyzing coal quality characteristics of a small sample based on deep transfer learning according to claim 1, characterized in that: The number of calibration samples of the target model is within 100.

5. A small sample coal quality characteristics analysis system based on deep transfer learning, characterized by: It includes a laser induced breakdown spectroscopy system for online detection of coal quality and a computer for coal quality characteristic analysis, wherein the computer includes a neural network source model module, a neural network target model module and a sample analysis module to be tested; The laser-induced breakdown spectroscopy system detects a large number of coal samples with known coal quality characteristics and different sources from the samples to be tested to obtain the characteristic spectrum of the source model calibration sample and the characteristic spectral line intensity of various elements, and transmits them to the neural network source model module; The laser-induced breakdown spectroscopy system detects a small amount of coal samples with known coal quality characteristics and the same source as the sample to be tested to obtain the characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements, and transmits them to the neural network target model module; The laser induced breakdown spectroscopy system detects the unknown coal quality characteristics of the sample to be tested to obtain the characteristic spectrum of the sample to be tested, that is, the spectral line intensity of the sample to be tested is obtained. I f , and transmit it to the sample analysis module to be tested; The neural network source model module is used to receive the characteristic spectrum of the source model calibration sample and the characteristic spectral line intensity of various elements, take a certain coal quality characteristic as the target characteristic, and establish a neural network source model for the target characteristic of the source model calibration sample The source model is used to calibrate the characteristic spectrum of the sample and the characteristic spectral line intensity of various elements to obtain the trained neural network source model. ; The neural network target model module is used to receive the characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements, and establish a neural network target model for the target characteristics of the target model calibration sample , using the trained neural network source model The characteristic spectrum of the target model calibration sample and the characteristic spectral line intensity of various elements are trained to obtain the trained neural network source model ; The sample analysis module is used to convert the spectral line intensity I f Substitute the trained neural network source model , obtain the value of the target characteristic in the sample to be tested; Coal quality characteristics include carbon content, ash content, volatile matter, calorific value and moisture. A pulsed laser is used as the excitation light source. The laser emitted from the laser is focused by a focusing lens and acts on the surface of the calibration sample. Plasma is generated at the focusing point. The plasma is cooled in an atmosphere of protective gas. The generated radiation light signal enters the optical fiber through a collection lens and is converted into an electrical signal after being processed by a spectrometer and collected by a computer. The computer obtains two groups of characteristic spectra of calibration samples. The first group is a large number of samples accumulated in past measurements and having a different source from the external measurement samples, which are called source model calibration samples. The second group is calibration samples that are the same source as the samples to be measured, which are called target model calibration samples. The characteristic spectral line intensities of various elements are obtained from the characteristic spectra of the source model calibration samples and the target model calibration samples, respectively.

6. The small sample coal quality characteristic analysis system based on deep transfer learning according to claim 5 is characterized by: The neural network source model module uses the characteristic spectrum of the source model calibration sample as the independent variable matrix , with the target characteristic values ​​of the source model calibration samples as the dependent variable matrix , train the neural network source model Until the prediction accuracy reaches the first preset value, the trained neural network source model is obtained. .

7. The small sample coal quality characteristic analysis system based on deep transfer learning according to claim 5 is characterized by: The neural network target model module converts the trained neural network source model The convolution layer structure, normalization layer structure and pooling layer structure are copied to the neural network target model respectively. The corresponding position of the neural network target model The rest of the matrix is ​​randomly initialized; the characteristic spectrum of the target model calibration sample is used as the independent variable matrix , with the target characteristic values ​​of the target model calibration samples as the dependent variable matrix , in the neural network target model Based on this, fine-tune the parameters, reduce the initial learning rate, increase the regularization coefficient, and train the neural network target model. Until the prediction accuracy reaches the second preset value, the trained neural network source model is obtained .

8. The small sample coal quality characteristic analysis system based on deep transfer learning according to claim 5 is characterized by: The laser induced breakdown spectroscopy system includes a pulsed laser, a focusing lens located on the light transmission path of the laser emitted by the pulsed laser and used to focus the laser to act on the surface of the calibration sample to generate luminous plasma, a collection lens used to sense the light information radiated by the luminous plasma, and a spectrometer that receives light information through optical fiber and generates electrical signals to transmit to a computer for coal quality characteristic analysis.

Images