A tunnel surrounding rock spectrum interpretation method and system based on large language model

Abstract

The present invention belongs to the technical field of large language models, and provides a tunnel surrounding rock spectrum interpretation method and system based on a large language model, including: obtaining the visible light-near infrared band spectrum and thermal infrared band spectrum of the tunnel surrounding rock, and using a data fusion method to fuse the spectra of the two bands to obtain a full-band spectrum; for multiple minerals contained in the rock, determining the characteristic position of each mineral, structuring the mineral name and characteristic position of each mineral, and obtaining the content of each mineral in the tunnel surrounding rock through a tunnel surrounding rock spectrum intelligent interpretation model based on a large language model; wherein, the method for determining the characteristic positions of multiple minerals is: calculating the correlation between the full-band spectrum in the data set and the content of each mineral, and selecting the characteristic band with the largest correlation in the full-band spectrum as the characteristic position of the mineral. Rapid intelligent interpretation of large-volume spectra in tunnels is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of large language models, and in particular to a tunnel surrounding rock spectrum interpretation method and system based on a large language model. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] During tunnel construction, adverse geological conditions such as faults and karst are often encountered, which can easily lead to geological disasters such as sudden water and mud, landslides, and large deformation of surrounding rocks. Existing geological analysis technologies mostly rely on qualitative analysis based on manual experience. With the continuous excavation of tunnels, the geological environment in the tunnel has become increasingly severe. Geological experts have to spend more than five hours in and out of the tunnel, making it impossible to conduct real-time quantitative analysis of the tunnel face, which can easily lead to misjudgment and omission of the geological conditions of the face, seriously affecting on-site construction safety and construction progress.

[0004] Clay and altered minerals are often found in the tunnel's adverse geological zone, and as the tunnel is excavated toward the core of the adverse geological zone, the content of clay and altered minerals gradually increases. Therefore, mineral information can be used as a quantitative standard for quantitative analysis of tunnel adverse geology.

[0005] Existing mineral identification technologies (thin section identification under a microscope, X-ray diffraction analysis, etc.) mostly need to be tested in the laboratory. The sample preparation process is complicated and the testing time for a single sample is long. They cannot meet the requirements of in-situ and rapid testing of tunnel surrounding rocks, and it is difficult to provide real-time guidance for tunnel disaster prevention and control and construction plan improvements.

[0006] Infrared spectroscopy testing technology can quickly obtain tunnel surrounding rock spectra in situ, but the spectrum of the rock surface is affected by the combined influence of multiple end-member mineral spectra, and characteristic variations often occur, resulting in the inability of traditional mineral spectral knowledge to interpret the mixed rock spectrum. As the tunnel is excavated, the number of spectra per kilometer of tunnel reaches tens of thousands. Existing data processing methods are difficult to combine with artificial intelligence technology, and cannot accurately and timely process the huge amount of data in the tunnel, and cannot provide timely and reliable mineral analysis results for on-site geological analysis. Summary of the invention

[0007] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a tunnel surrounding rock spectrum interpretation method and system based on a large language model, which determines the characteristic positions of different minerals through correlation, structures the characteristic positions and mineral names into data, and establishes a tunnel surrounding rock spectrum intelligent interpretation model based on a large language model through natural language prompts, which can quickly process tens of thousands of engineering data per kilometer of tunnel, and realize the rapid intelligent interpretation of large-volume spectra in the tunnel.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A first aspect of the present invention provides a tunnel surrounding rock spectrum interpretation method based on a large language model.

[0010] A tunnel surrounding rock spectrum interpretation method based on a large language model, comprising:

[0011] Obtain the visible-near infrared band spectrum and thermal infrared band spectrum of the tunnel surrounding rock, and fuse the spectra of the two bands using the data fusion method to obtain the full-band spectrum;

[0012] For the various minerals contained in the rock, the characteristic position of each mineral is determined, and the mineral name and characteristic position of each mineral are structured to obtain structured data. The content of each mineral in the tunnel surrounding rock is obtained through the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model;

[0013] The method for determining the characteristic position is as follows: calculating the correlation between the full-band spectrum and the content of each mineral, and selecting the characteristic band with the largest correlation in the full-band spectrum as the characteristic position of each mineral.

[0014] Furthermore, it also includes: before obtaining the characteristic position, smoothing the full-band spectrum, removing the data noise, and performing envelope removal, baseline correction and multivariate scattering correction.

[0015] Furthermore, the data fusion method adopts cumulative fusion.

[0016] Furthermore, the mineral names of the minerals include potassium feldspar, sodium feldspar, illite and montmorillonite.

[0017] Furthermore, the training process of the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model includes:

[0018] The mineral identification and analysis methods in the laboratory are used to obtain the mineral name and content of each mineral in the rock sample, and then the visible light-near infrared band spectrum and thermal infrared band spectrum of the rock sample with known mineral composition and content are obtained. The spectra of the two bands are fused using the data fusion method to obtain the full-band spectrum, determine the characteristic position of each mineral, and use the full-band spectra of all rock samples as the data set;

[0019] The data set is divided into a training set and a test set. The mineral name, characteristic position and content of each mineral in each rock sample in the training set are structured to obtain structured data, and the large language model is commanded to learn the training set data information through natural language description. The mineral name and characteristic position coding of each mineral in each rock sample in the test set are structured to obtain structured data, and the content of the mineral is calculated by the large language model through natural language description. When the error between the calculated mineral content and the known content is less than a threshold, an intelligent interpretation model of tunnel surrounding rock spectra based on the large language model is obtained.

[0020] The second aspect of the present invention provides a tunnel surrounding rock spectrum interpretation system based on a large language model.

[0021] A tunnel surrounding rock spectrum interpretation system based on a large language model, comprising:

[0022] A spectrum fusion module is configured to: obtain a visible light-near infrared band spectrum and a thermal infrared band spectrum of the tunnel surrounding rock, and fuse the spectra of the two bands using a data fusion method to obtain a full-band spectrum;

[0023] The spectrum interpretation module is configured to: determine the characteristic position of each mineral contained in the rock, structure the mineral name and characteristic position of each mineral to obtain structured data, and calculate the content of each mineral in the tunnel surrounding rock through the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model;

[0024] The method for determining the characteristic position is as follows: calculating the correlation between the full-band spectrum in the data set and the content of each mineral, and selecting the characteristic band with the largest correlation in the full-band spectrum as the characteristic position of each mineral.

[0025] Furthermore, it also includes a preprocessing module, which is configured to: before obtaining the characteristic position, smooth the full-band spectrum, remove the data noise, and then perform envelope removal, baseline correction and multivariate scattering correction.

[0026] Furthermore, the data fusion method adopts cumulative fusion.

[0027] Furthermore, the mineral names of the minerals include potassium feldspar, sodium feldspar, illite and montmorillonite.

[0028] Furthermore, a training module is included, which is configured to:

[0029] The mineral identification and analysis methods in the laboratory are used to obtain the mineral name and content of each mineral in the rock sample, and then the visible light-near infrared band spectrum and thermal infrared band spectrum of the rock sample with known mineral composition and content are obtained. The spectra of the two bands are fused using the data fusion method to obtain the full-band spectrum, determine the characteristic position of each mineral, and use the full-band spectra of all rock samples as the data set;

[0030] The data set is divided into a training set and a test set. The mineral name, characteristic position and content of each mineral in each rock sample in the training set are structured to obtain structured data, and the large language model is commanded to learn the training set data information through natural language description; the mineral name and characteristic position of each mineral in each rock sample in the test set are structured to obtain structured data, and the content of the mineral is calculated by the large language model through natural language description. When the error between the calculated mineral content and the known content is less than a threshold, an intelligent interpretation model of tunnel surrounding rock spectra based on the large language model is obtained.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention determines the characteristic positions of different minerals through correlation, structures the characteristic positions and mineral names, and establishes a tunnel surrounding rock spectrum intelligent interpretation model based on a large language model through natural language prompts. It can quickly process tens of thousands of engineering data per kilometer of tunnel, and can solve the complex abstract relationship between spectra and minerals with excellent performance. It solves the shortcomings of existing spectral interpretation methods that most of them rely on traditional manual processing and complex parameter tuning and cannot guarantee timeliness and accuracy, and provides timely guidance for tunnel disaster prevention and control.

[0033] The present invention reduces the dimension of data through data preprocessing, feature extraction and other means, can accurately replace the source spectrum and maintain the mineral characteristics contained therein, improves the model processing efficiency, and at the same time avoids the disadvantages of feature loss and inaccurate identification caused by multiple data transmissions due to excessive data volume.

[0034] The present invention realizes full coverage identification of the types and contents of characteristic minerals of adverse geology by fusing the visible light-near infrared band and the thermal infrared band, improves the quantitative level of adverse geology identification, and eliminates the influence of subjective factors such as expert experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings, which constitute a part of the specification of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but do not constitute limitations of the present invention.

[0036] Figure 1It is a flow chart of a tunnel surrounding rock spectrum interpretation method based on a large language model according to the first embodiment of the present invention;

[0037] Figure 2 It is a detailed schematic diagram of a large language model portion of the first embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0039] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0040] In the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other. The present invention is further described below with reference to the accompanying drawings and embodiments.

[0041] Embodiment 1

[0042] This embodiment provides a tunnel surrounding rock spectrum interpretation method based on a large language model.

[0043] The present embodiment provides a tunnel surrounding rock spectrum interpretation method based on a large language model, which realizes the rapid and intelligent interpretation of large-volume spectra in the tunnel, and is of great significance for improving the efficiency and accuracy of tunnel surrounding rock mineral identification and ensuring rapid and safe tunnel construction.

[0044] This embodiment provides a tunnel surrounding rock spectrum interpretation method based on a large language model, such as Figure 1 As shown, the following steps are included:

[0045] Step 1: Tunnel on-site sample mineral testing: Collect rock samples from unfavorable geological areas in the tunnel, and use precise mineral identification and analysis methods in the laboratory to obtain accurate sample mineral composition and content information;

[0046] Step 2: For the samples with known mineral information in step 1, the spectra of the visible light-near infrared band and the thermal infrared band are collected, and the spectra of the two bands of each sample are fused by a data fusion method, and multiple band data are combined into one, so that each sample only saves one piece of data covering the visible light-near infrared band and the thermal infrared band, and obtains full-band spectrum data;

[0047] Step 3, collecting the full-band spectral data of all tunnel field samples obtained in step 2, and establishing a full-band spectral set of tunnel field samples (also called spectral data set);

[0048] Step 4: preprocess the spectrum in the spectrum set in step 3 to improve the spectral signal-to-noise ratio, filter out cluttered signals, highlight spectral features, and weaken the differences caused by different test conditions;

[0049] Step 5, feature extraction: perform correlation calculation on the data preprocessed in step 4, and extract the single feature band or multiple feature bands with the strongest correlation with the content of different minerals as the characteristic position of the mineral, so as to achieve the purpose of data dimensionality reduction;

[0050] Step 6: Divide the data after extracting features in step 5 into a training set and a test set, which are used for training and testing the tunnel surrounding rock spectrum intelligent interpretation model respectively;

[0051] Step 7: Structure the training set and test set data in step 6. The training set data label format after structured is {'mineral name':'sample M', 'x: feature position nm', 'y: mineral content %'}, and the test set data is structured as {'mineral name':'sample N', 'x: feature position nm'};

[0052] Step 8: Use the first natural language prompt (natural language description) Prompt A to describe the structured training set data obtained in step 7, and command the large language model to learn the training set data information; the second natural language prompt (natural language description) Prompt B is used to command the prediction of the test set data;

[0053] Step 9: Finally, according to the order of the test set input samples, the y value is output to obtain the type and content of the test set's bad geological indicator mineral samples. When the error between the predicted mineral content and the known content is less than the threshold, the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model is obtained;

[0054] Step 10. In actual application, obtain the visible light-near infrared band spectrum and thermal infrared band spectrum of the tunnel surrounding rock, and use the data fusion method to fuse the spectra of the two bands to obtain the full-band spectrum; for the various minerals contained in the rock, determine the characteristic position of each mineral, structure the mineral name and characteristic position of each mineral to obtain structured data, and use the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model to obtain the content of various minerals in the tunnel surrounding rock.

[0055] In step 1, samples are collected from the unfavorable geological areas in the tunnel, and accurate mineral identification and analysis methods are used in the laboratory to obtain accurate information on mineral composition and content.

[0056] Among them, unfavorable geological areas include faults, alteration zones, joint-dense zones, and lithological contact zones, as well as areas where the color of the face has changed significantly due to geological processes such as weathering.

[0057] Among them, sample collection means collecting rock samples from the tunnel face according to the tunnel mileage, and numbering them according to the mileage and the position of the samples on the tunnel face.

[0058] Among them, the precise mineral identification and analysis methods in the laboratory include X-ray diffraction analysis (XRD), thin section identification under a microscope and other methods, which can accurately obtain the mineral type and content information of rock samples.

[0059] In this embodiment, the adverse geological area specifically refers to the faults where clay minerals and alteration minerals are enriched. The rock samples mainly contain four minerals: potassium feldspar (Kfs), albite (Ab), illite (Ilt) and montmorillonite (Mnt). Their contents are shown in Table 1 (taking 6 samples as an example).

[0060] Table 1. Mineral contents of rock samples

[0061]

[0062] In step 2, the visible light-near infrared band and thermal infrared band spectra are collected for the samples with known mineral information in step 1, and the spectra of the two bands of each sample are fused using a data fusion method, so that each sample only saves one piece of data covering the visible light-near infrared band and the thermal infrared band.

[0063] Among them, visible light-near infrared and thermal infrared both belong to infrared light, and the two bands are divided based on the different frequencies of light.

[0064] Among them, the visible light-near infrared band has advantages in detecting clays and altered minerals such as montmorillonite, chlorite, and muscovite, while the thermal infrared band has the fingerprint characteristics of dark minerals and rock-forming minerals such as feldspar, quartz, and biotite. Therefore, combining the data of the two bands can achieve full coverage identification of mineral types inside the sample.

[0065] Among them, data fusion methods include but are not limited to cumulative fusion (CF), equal-weight fusion (ERF), outer product fusion (OPF), etc.

[0066] In this embodiment, the data fusion method of cumulative fusion (CF) is adopted to retain all the information of the visible light-near infrared band and thermal infrared band data of the rock sample, so that the two data are spliced ​​into one. The reason for selecting this method is that feature extraction and other operations will be performed in step 5, which will greatly reduce the dimension of the data. Therefore, there is no need to reduce the dimension of the spectrum through other data fusion methods in step 2, so as to avoid problems such as data loss and failure of feature band extraction.

[0067] In step 3, the full-band data of all samples obtained in step 2 are collected to establish a full-band spectrum set of tunnel field samples.

[0068] The sample full-band spectrum set is to save the sample full-band spectrum generated in step 2 according to the mileage number, the position of the tunnel face, and the spectrum sequence.

[0069] In step 4, the spectra in the spectrum set in step 3 are preprocessed to achieve the purpose of improving the spectral signal-to-noise ratio, highlighting the spectral features, and weakening the differences caused by different test conditions.

[0070] The acquired full-band spectrum is preprocessed, wherein the preprocessing methods include but are not limited to multivariate scattering correction, fractional-order differentiation, smoothing, envelope removal, baseline correction and the like.

[0071] In this embodiment, convolution smoothing (SG smoothing) is performed on the full-band spectrum of the rock sample obtained in step 3 to remove data noise, and then envelope removal and baseline correction are performed to highlight the spectral features. Finally, multivariate scattering correction is performed to eliminate the influence of different test conditions.

[0072] In step 5, correlation calculation is performed on the data preprocessed in step 4, and the band with the strongest correlation with the mineral content is extracted as the characteristic position of the mineral, thereby achieving the purpose of extracting features.

[0073] The correlation calculation is to calculate the band with the strongest correlation with the target mineral content, including but not limited to the Pearson correlation coefficient calculation method.

[0074] Among them, extracting the characteristic bands of minerals can greatly reduce the dimension of the data, thereby saving time in data import and model processing and reducing the difficulty of model learning.

[0075] In this embodiment, the Pearson correlation calculation is performed on the mineral content and the spectrum to select the characteristic band.

[0076] Among them, the Pearson correlation coefficient calculation follows the following formula: .

[0077] In the formula, r represents the Pearson correlation coefficient, n represents the number of samples, and X iis the characteristic value of a characteristic band of the spectrum of the i-th sample, is the average value of the characteristic value of a certain characteristic band of all mineral spectra, Y i is the content of a certain mineral in the i-th sample, The average value of the content of a certain mineral.

[0078] In this embodiment, the characteristic bands of potassium feldspar, albite, illite and montmorillonite selected by the Pearson correlation coefficient are 9480nm, 9616nm, 2207nm and 2210nm respectively, and only one characteristic position is selected here.

[0079] As an implementation method, there may be multiple selected characteristic positions, and the number of characteristic positions is calculated based on the Pearson correlation coefficient. Specifically: assuming that the Pearson correlation coefficient between the a-th mineral and the b-th characteristic band is rab; for the a-th mineral, all characteristic bands are arranged from large to small according to the Pearson correlation coefficient with the a-th mineral to obtain a sequence (ra1, ra2, …, rak), where k is the number of characteristic bands; then for the a-th mineral, the number of characteristic positions N=Q / (ra1+ra2+ra3), where Q is a set value.

[0080] In step 6, the data after the features are extracted in (5) are divided into a training set and a test set, which are used for training and testing the tunnel surrounding rock spectrum intelligent interpretation model respectively;

[0081] Among them, the data is divided into training set and test set, that is, the data is divided according to a certain ratio, the training set is used to train the model, and the test set is used to verify the accuracy of the model.

[0082] Among them, the methods of dividing the training set and the test set include but are not limited to the holdout method, the cross-validation method and the bootstrap method.

[0083] In this embodiment, the hold-out method is adopted to divide 2 / 3 of the data into a training set and 1 / 3 of the data into a test set.

[0084] In step 7, the training set and test set data in step 6 are structured. The training set data label format after structured is {'mineral name':'sample M', 'x: feature position nm', 'y: mineral content %'}, and the test set data is structured as {'mineral name':'sample N', 'x: feature position nm'}, such as Figure 2 shown.

[0085] Among them, data structuring means simplifying the data in the data set according to certain rules. The simplification process should use specific characters and formats to facilitate machine understanding.

[0086] In this embodiment, sample 1, sample 3, sample 5 and sample 6 are used as training sets, and their structured information is shown in Table 2.

[0087] Table 2. Structural information of the training set

[0088]

[0089] Sample 2 and sample 4 are used as test sets, and their structural information is shown in Table 3.

[0090] Table 3. Structural information of the test set

[0091]

[0092] In step 8, the first natural language prompt Prompt A is used to describe the structured data obtained in step 7 and command the large language model to learn the training set data information; the second natural language prompt Prompt B is used to command the prediction of the test set data.

[0093] For example, Prompt A can be: the input data set is structured data including mineral name, x: feature position nm and y: mineral content %; Prompt B can be: calculate y: mineral content % corresponding to each sample in the test set. Instructing the large language model to learn refers to using a deep learning model based on the transformer algorithm to extract abstract relationships in the training set data. The large language model includes but is not limited to the text-davinci-003 model, etc. Instructing the large language model to calculate the mineral content through natural language description means: inputting Prompt B into the deep learning model based on the transformer algorithm, so that the deep learning model outputs the mineral content.

[0094] In this embodiment, the large language model uses the text-davinci-003 model. The total natural language prompt TotalPrompt consists of Prompt A and Prompt B. Prompt A is used to command the text-davinci-003 model to learn the structured information of samples 1, 3, 5, and 6 in the training set, and Prompt B is used to command the text-davinci-003 model to predict the mineral content information of samples 2 and 4 in the test set.

[0095] In step 9, the y value is finally output according to the order of the test set input samples to obtain the mineral content of the test set rock samples.

[0096] In this embodiment, the contents of various minerals of Sample 2 and Sample 4 are outputted through the trained large language model as shown in Table 4.

[0097] Table 4. Content of various minerals in sample 2 and sample 4

[0098]

[0099] This embodiment determines the characteristic positions of different minerals through correlation, structures the characteristic positions and mineral names, and obtains structured data. It can quickly process tens of thousands of engineering data per kilometer of tunnel through natural language prompts, and can solve the complex abstract relationship between spectra and minerals with excellent performance. It solves the shortcomings of existing spectral interpretation methods that most of them rely on traditional manual processing and complex parameter tuning and cannot guarantee timeliness and accuracy, and provides timely guidance for tunnel disaster prevention and control.

[0100] This embodiment reduces the dimension of the data through data preprocessing, feature extraction and other means, can accurately replace the source spectrum and maintain the mineral characteristics contained therein, improves the model processing efficiency, and at the same time avoids the disadvantages of feature loss and inaccurate identification caused by multiple data transmissions due to excessive data volume.

[0101] This embodiment achieves full coverage identification of the types and contents of characteristic minerals of adverse geology by fusing the visible light-near infrared band and the thermal infrared band, improves the quantitative level of adverse geology identification, and eliminates the influence of subjective factors such as expert experience.

[0102] Embodiment 2

[0103] This embodiment provides a tunnel surrounding rock spectrum interpretation system based on a large language model.

[0104] A tunnel surrounding rock spectrum interpretation system based on a large language model, comprising:

[0105] The training module is configured as follows: after obtaining the mineral name and content of each mineral in the rock sample by using the mineral identification and analysis method in the laboratory, the visible light-near infrared band spectrum and thermal infrared band spectrum of the rock sample with known mineral composition and content are obtained, and the spectra of the two bands are fused by using the data fusion method to obtain the full-band spectrum, determine the characteristic position of each mineral, and use the full-band spectrum of all rock samples as the data set; divide the data set into a training set and a training set, structure the mineral name, characteristic position and content of each mineral in each rock sample in the training set to obtain structured data, and use natural language description to command the large language model to learn the training set data information; structure the mineral name and characteristic position of each mineral in each rock sample in the test set to obtain structured data, and use natural language description to command the large language model to calculate the mineral content, and when the error between the calculated mineral content and the known content is less than a threshold, a tunnel surrounding rock spectrum intelligent interpretation model based on the large language model is obtained.

[0106] A spectrum fusion module is configured to: obtain a visible light-near infrared band spectrum and a thermal infrared band spectrum of the tunnel surrounding rock, and fuse the spectra of the two bands using a data fusion method to obtain a full-band spectrum;

[0107] The spectrum interpretation module is configured to: obtain the characteristic position of each mineral contained in the rock, structure the mineral name and characteristic position of each mineral to obtain structured data, and calculate the content of the mineral in the tunnel surrounding rock through the natural language prompt command large language model;

[0108] The method for determining the characteristic position is as follows: calculating the correlation between the full-band spectrum in the data set and the content of each mineral, and selecting the characteristic band with the largest correlation in the full-band spectrum as the characteristic position of each mineral.

[0109] It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0111] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A tunnel surrounding rock spectrum interpretation method based on a large language model, characterized in that: include: Obtain the visible-near infrared band spectrum and thermal infrared band spectrum of the tunnel surrounding rock, and fuse the spectra of the two bands using the data fusion method to obtain the full-band spectrum; For the various minerals contained in the rock, the characteristic position of each mineral is determined, and the mineral name and characteristic position of each mineral are structured to obtain structured data. The content of each mineral in the tunnel surrounding rock is obtained through the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model; The method for determining the characteristic position is as follows: calculate the correlation between the full-band spectrum in the data set and the content of each mineral, select the characteristic band with the largest correlation in the full-band spectrum as the characteristic position of each mineral, and the number of characteristic positions is calculated based on the Pearson correlation coefficient. Specifically, assume that the Pearson correlation coefficient between the a-th mineral and the b-th characteristic band is rab; for the a-th mineral, arrange all characteristic bands from large to small according to the Pearson correlation coefficient with the a-th mineral to obtain a sequence (ra1, ra2, …, rak), where k is the number of characteristic bands; then for the a-th mineral, the number of characteristic positions N=Q / (ra1+ra2+ra3), where Q is the set value; The large language model uses a deep learning model based on the transformer algorithm; The training process of the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model includes: using the laboratory mineral identification and analysis method to obtain the mineral name and content of each mineral in the rock sample, obtaining the visible light-near infrared band spectrum and thermal infrared band spectrum of the rock sample with known mineral composition and content, using the data fusion method to fuse the spectra of the two bands to obtain the full-band spectrum, determining the characteristic position of each mineral, and using the full-band spectrum of all rock samples as a data set; dividing the data set into a training set and a test set, structuring the mineral name, characteristic position and content of each mineral in each rock sample in the training set to obtain structured data {'mineral name':'sample M', 'x: characteristic position nm', 'y: mineral content %'}, and using the natural language description Prompt A to command the large language model to learn the training set data information; structuring the mineral name and characteristic position coding of each mineral in each rock sample in the test set to obtain structured data {'mineral name':'sample N', 'x: feature position nm'}, through natural language description PromptB commands the large language model to calculate the mineral content. When the error between the calculated mineral content and the known content is less than the threshold, the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model is obtained; Prompt A is: the input data set is structured data containing mineral name, x: feature position nm and y: mineral content %; Prompt B is: calculate the y: mineral content % corresponding to each sample in the test set.

2. The tunnel surrounding rock spectrum interpretation method based on a large language model according to claim 1 is characterized in that: Also includes: Before obtaining the characteristic position, the full-band spectrum is smoothed, and after removing the data noise, envelope removal, baseline correction and multivariate scattering correction are performed.

3. The tunnel surrounding rock spectrum interpretation method based on a large language model according to claim 1 is characterized in that: The data fusion method adopts cumulative fusion.

4. The tunnel surrounding rock spectrum interpretation method based on a large language model according to claim 1 is characterized in that: The mineral names of the minerals include potassium feldspar, sodium feldspar, illite and montmorillonite.

5. A tunnel surrounding rock spectrum interpretation system based on a large language model, characterized in that: include: A spectrum fusion module is configured to: obtain a visible light-near infrared band spectrum and a thermal infrared band spectrum of the tunnel surrounding rock, and fuse the spectra of the two bands using a data fusion method to obtain a full-band spectrum; The spectrum interpretation module is configured to: determine the characteristic position of each mineral contained in the rock, structure the mineral name and characteristic position of each mineral to obtain structured data, and obtain the content of each mineral in the tunnel surrounding rock through the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model; The method for determining the characteristic position is as follows: calculate the correlation between the full-band spectrum in the data set and the content of each mineral, select the characteristic band with the largest correlation in the full-band spectrum as the characteristic position of each mineral, and the number of characteristic positions is calculated based on the Pearson correlation coefficient. Specifically, assume that the Pearson correlation coefficient between the a-th mineral and the b-th characteristic band is rab; for the a-th mineral, arrange all characteristic bands from large to small according to the Pearson correlation coefficient with the a-th mineral to obtain a sequence (ra1, ra2, …, rak), where k is the number of characteristic bands; then for the a-th mineral, the number of characteristic positions N=Q / (ra1+ra2+ra3), where Q is the set value; The large language model uses a deep learning model based on the transformer algorithm; The training process of the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model includes: using the laboratory mineral identification and analysis method to obtain the mineral name and content of each mineral in the rock sample, obtaining the visible light-near infrared band spectrum and thermal infrared band spectrum of the rock sample with known mineral composition and content, using the data fusion method to fuse the spectra of the two bands to obtain the full-band spectrum, determining the characteristic position of each mineral, and using the full-band spectrum of all rock samples as a data set; dividing the data set into a training set and a test set, structuring the mineral name, characteristic position and content of each mineral in each rock sample in the training set to obtain structured data {'mineral name':'sample M', 'x: characteristic position nm', 'y: mineral content %'}, and using the natural language description Prompt A to command the large language model to learn the training set data information; structuring the mineral name and characteristic position coding of each mineral in each rock sample in the test set to obtain structured data {'mineral name':'sample N', 'x: feature position nm'}, through natural language description PromptB commands the large language model to calculate the mineral content. When the error between the calculated mineral content and the known content is less than the threshold, the tunnel surrounding rock spectrum intelligent interpretation model based on the large language model is obtained; Prompt A is: the input data set is structured data containing mineral name, x: feature position nm and y: mineral content %; Prompt B is: calculate the y: mineral content % corresponding to each sample in the test set.

6. The tunnel surrounding rock spectrum interpretation system based on a large language model according to claim 5, characterized in that: It also includes a preprocessing module, which is configured to: smooth the full-band spectrum before acquiring the characteristic position, remove the data noise, and then perform envelope removal, baseline correction and multivariate scattering correction.

7. The tunnel surrounding rock spectrum interpretation system based on a large language model according to claim 5, characterized in that: The data fusion method adopts cumulative fusion.

8. The tunnel surrounding rock spectrum interpretation system based on a large language model according to claim 5, characterized in that: The mineral names of the minerals include potassium feldspar, sodium feldspar, illite and montmorillonite.

Images