Tunnel surrounding rock lithology identification method and system based on multi-modal data knowledge distillation

Abstract

The application provides a tunnel surrounding rock lithology identification method and system based on multi-modal data knowledge distillation, belongs to the technical field of geotechnical and tunnel engineering and geological exploration, and comprises the following steps: acquiring LIBS spectrum data, major element composition data and mineral composition data of a rock sample; taking the LIBS spectrum data as a main modal input, taking the element composition data and the mineral composition data as auxiliary knowledge modes, constructing a multi-modal data knowledge distillation lithology identification model and training the model; stripping an element encoder and a mineral encoder from the trained multi-modal data knowledge distillation lithology identification model to obtain a lithology reasoning model; and inputting LIBS spectrum data of a sample to be detected into the lithology reasoning model to obtain a lithology identification result. The application improves the lithology identification accuracy and generalization ability based on the LIBS spectrum.

Description

Technical Field

[0001] This invention belongs to the field of geotechnical and tunnel engineering and geological exploration technology, and particularly relates to a method and system for identifying the lithology of tunnel surrounding rock based on multimodal data knowledge distillation. Background Technology

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

[0003] Lithological identification of tunnel surrounding rock is a fundamental task in tunnel construction geological logging and surrounding rock classification, directly affecting the selection of support schemes and construction safety assessments. Traditional lithological identification methods mainly rely on geologists' visual observation and microscopic identification, classifying rocks by observing characteristics such as color, structure, texture, and mineral composition. This method is limited by the professional experience level of the identification personnel and has a long identification cycle, making it difficult to meet the needs of tunnel construction for rapid acquisition of geological information.

[0004] In recent years, with the development of spectral analysis technology, lithological identification methods based on visible-near-infrared (VNIR) and short-wave infrared (SWIR) reflectance spectra have gradually attracted attention. Different rocks exhibit different spectral absorption characteristics in specific wavelength bands due to differences in their mineral composition and chemical composition, which provides a physical basis for lithological classification using spectral data. However, existing spectral-based lithological identification methods generally suffer from the following problems: (1) There is a gap between spectral features and geological semantics. Traditional machine learning methods (such as SVM, random forest, etc.) and deep learning methods directly learn the mapping relationship between spectral curves and lithological labels. Although the models can achieve good results on the training set, the learned features often lack clear geological interpretability, resulting in insufficient generalization ability and difficulty in transferring to different geological conditions.

[0005] (2) Geological knowledge is not being effectively utilized. Existing spectral lithology identification methods generally rely only on a single spectral data mode, while ignoring the geological knowledge contained in the elemental and mineral composition data of rocks.

[0006] (3) The harsh environment at the tunnel construction site places special demands on lithological identification methods. After the tunnel face is excavated, there are often unfavorable conditions such as dust, insufficient light, limited space, and dripping water. Moreover, the construction pace is tight, requiring geological information to be acquired within minutes or even less to match the pace of construction cycle. Currently, commonly used on-site rapid lithological identification methods mainly include: simple methods such as tapping with a geological hammer to listen to the sound, dripping with dilute hydrochloric acid, and observing fracture surfaces and mineral crystals with a magnifying glass. Although these methods are quick, they are highly dependent on the professional experience and subjective judgment of the operators, making it difficult to form standardized and quantifiable identification results. Furthermore, they are not sensitive to changes in mineral assemblage that are highly concealed, which can easily lead to misjudgments. Although laboratory analysis methods such as XRD, XRF, and thin section identification are highly accurate, they require samples to be transported back to the laboratory for sample preparation, testing, and expert interpretation, which usually takes several hours to several days. This cannot meet the timeliness requirement of "evaluation as excavation progresses" in tunnel construction. Summary of the Invention

[0007] To overcome the shortcomings of the existing technologies, this invention proposes a method and system for identifying the lithology of tunnel surrounding rock based on multimodal data knowledge distillation. By fusing data from three modalities—LIBS spectral, elemental composition, and mineral composition—and using the major elemental composition data of the rock as a knowledge intermediary, a cascaded feature alignment distillation strategy is employed to transfer geological knowledge from the elemental and mineral data to the LIBS spectral encoder. This allows the LIBS spectral encoder to learn feature representations with more geological semantics. After model deployment, only LIBS spectral data needs to be input to achieve high-precision lithology identification, meeting the need for rapid in-situ lithology identification. This method is suitable for scenarios such as advanced geological prediction of tunnels and real-time in-situ determination of lithology at tunnel excavation faces.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention discloses a method for identifying the lithology of tunnel surrounding rock based on multimodal data knowledge distillation, comprising: Acquire LIBS spectral data, major elemental composition data, and mineral composition data of rock samples; Using the LIBS spectral data as the primary modal input and elemental composition data and mineral composition data as auxiliary knowledge modalities, a multimodal data knowledge distillation lithology identification model is constructed and trained. This model includes a LIBS spectral encoder, an elemental encoder, a mineral encoder, and a classification head. During training, the model employs a cascaded feature alignment distillation strategy: first, mineral features are aligned to elemental features, using elemental features as knowledge anchors to constrain the mineral encoder's feature learning, enabling it to learn mineral assemblage representations that conform to elemental stoichiometry constraints; then, spectral projection features are distilled onto the aligned mineral features, using mineral features as geological semantic bridges to constrain the spectral encoder's feature learning, enabling it to learn geological genetic chain feature representations. The element encoder and mineral encoder are stripped from the knowledge distillation lithology identification model of the trained multimodal data to obtain the lithology inference model. The LIBS spectral data of the sample to be detected is then input into the lithology inference model to obtain the lithology identification result.

[0009] Secondly, this invention discloses a tunnel surrounding rock lithology identification system based on multimodal data knowledge distillation, comprising: The data acquisition module is configured to acquire LIBS spectral data, major elemental composition data, and mineral composition data of rock samples. The model building module is configured to: use the LIBS spectral data as the primary modal input and elemental composition data and mineral composition data as auxiliary knowledge modalities to construct and train a multimodal data knowledge distillation lithology identification model; the multimodal data knowledge distillation lithology identification model includes a LIBS spectral encoder, an elemental encoder, a mineral encoder, and a classification head; during training, the multimodal data knowledge distillation lithology identification model adopts a cascaded feature alignment distillation strategy: first, mineral features are aligned to elemental features, using elemental features as knowledge anchors to constrain the feature learning of the mineral encoder, enabling the mineral encoder to learn mineral assemblage representations that conform to elemental stoichiometry constraints; then, spectral projection features are distilled to the aligned mineral features, using mineral features as geological semantic bridges to constrain the feature learning of the spectral encoder, enabling the spectral encoder to learn geological genetic chain feature representations; The lithology identification module is configured to: extract the element encoder and mineral encoder from the lithology identification model by knowledge distillation of the trained multimodal data to obtain a lithology inference model, and input the LIBS spectral data of the sample to be detected into the lithology inference model to obtain the lithology identification result.

[0010] Thirdly, the present invention discloses an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when run by the processor, complete the steps of the above-mentioned method for identifying the lithology of tunnel surrounding rock based on multimodal data knowledge distillation.

[0011] Fourthly, the present invention discloses a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-described method for identifying the lithology of tunnel surrounding rock based on multimodal data knowledge distillation.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a cascaded feature alignment distillation strategy, transferring knowledge according to the geological genetic chain of "element → mineral → lithology." This ensures that geological knowledge is injected into the model step-by-step according to its natural genetic logic, enabling the spectral encoder to ultimately learn feature representations with a complete geological semantic chain. This significantly improves the accuracy and generalization ability of lithology identification based on LIBS spectroscopy. The invention uses a three-modal distillation training framework, using elemental and mineral data as auxiliary knowledge sources only during the training phase. After training, the auxiliary encoder is removed, and lithology identification can be completed by inputting only LIBS spectral data during the inference phase. This approach balances high identification accuracy with low inference costs, making it suitable for the rapid in-situ detection needs of tunnel engineering sites.

[0013] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0015] Figure 1 This is a flowchart of the tunnel surrounding rock lithology identification method based on multimodal data knowledge distillation as described in Embodiment 1 of the present invention.

[0016] Figure 2 This is a schematic diagram of the training and inference stages of the multimodal data knowledge distillation lithology identification model described in Embodiment 1 of the present invention.

[0017] Figure 3 This is a schematic diagram of the tunnel surrounding rock lithology identification system based on multimodal data knowledge distillation as described in Embodiment 1 of the present invention. Detailed Implementation

[0018] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0020] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0021] Example 1 In one or more embodiments, a method for identifying the lithology of tunnel surrounding rock based on multimodal data knowledge distillation is disclosed. This provides a rapid in-situ lithology identification technique that can be directly deployed at the tunnel face, is easy to operate, and can output standardized identification results within seconds to minutes. Figure 1 As shown, it includes the following steps: Step S1: Obtain LIBS spectral data, major elemental composition data, and mineral composition data of the rock sample, and preprocess the LIBS spectral data.

[0022] In this embodiment, the rock samples can be derived from rock blocks collected at the tunnel face, borehole cores, or cutaway samples. For each sample, data from three modalities are collected in parallel: (1) LIBS spectral data acquisition: Laser-induced breakdown (LIBS) spectra of rock samples are acquired using a portable or laboratory-scale LIBS spectrometer under pulsed laser excitation. LIBS utilizes a high-energy pulsed laser focused onto the sample surface to generate a transient high-temperature plasma. By analyzing the characteristic spectral lines of each element in the plasma emission spectrum, rapid analysis of the sample's elemental composition is achieved. Typical operating parameters are: laser wavelength 1064 nm, pulse energy 50-200 mJ, repetition frequency 1-10 Hz, and laser spot diameter 50-200 μm. The pulse energy can be adjusted according to the sample type; for example, a higher energy of 150-200 mJ is used for hard rocks such as granite, while a lower energy of 50-100 mJ is used for loose or fragile samples. The spectral acquisition range covers the 200-1000 nm band, which includes the main characteristic emission lines of most rock-forming elements, such as Si, Al, Fe, Ca, Mg, K, Na, Ti, Mn, and P. The spectrometer resolution is better than 0.1 nm, ensuring the ability to resolve adjacent characteristic spectral lines.

[0023] During the actual data collection, LIBS measurements were performed at 3-5 different locations on each rock sample. At each location, 10-30 laser pulse signals were collected and averaged to eliminate random errors caused by laser energy fluctuations and sample surface inhomogeneities. Each sample ultimately yielded an average LIBS spectral curve containing 3994 wavelength channels.

[0024] After LIBS spectral acquisition is completed, the following preprocessing operations are performed: (a) Noise reduction: A Savitzky-Golay smoothing filter is used, preferably with a window size of 15 and a polynomial order of 3, which effectively removes high-frequency noise while preserving the sharpness of the characteristic spectral lines. (b) Baseline correction: The asymmetric least squares baseline correction method (asLS) is used to subtract the continuous background radiation generated by bremsstrahlung and recombination radiation in the LIBS spectrum, highlighting the characteristic emission lines; (c) Spectral line normalization: A full-spectrum area normalization method is used, dividing each spectrum by the sum of the integrated intensities across the entire spectrum to eliminate the influence of laser pulse energy fluctuations and measurement distance variations on the absolute spectral intensity. The formula is:

[0025] in, The normalized spectral intensity of the k-th wavelength channel; This represents the original spectral intensity of the k-th wavelength channel; is the original spectral intensity of the i-th wavelength channel.

[0026] Preferably, when argon is used as the protective gas, the internal standard normalization method can also be used, selecting an inert element spectral line (such as Ar or He) that is not affected by changes in sample composition as the internal standard for normalization.

[0027] After the above preprocessing, a one-dimensional LIBS spectral vector S∈R is formed. m， (m=2048), which serves as the input for subsequent models.

[0028] (2) Data collection of principal element composition: X-ray fluorescence spectrometry (XRF) was used to obtain major elemental composition data of rock samples. XRF technology excites the characteristic X-ray fluorescence of elements in the sample by primary X-rays, and performs qualitative and quantitative analysis of elements based on the wavelength and intensity of the characteristic X-rays. It has the advantages of simple sample preparation, fast analysis speed, and high precision, and is a classic method for major element analysis of rocks.

[0029] In practice, the rock sample is crushed and ground to a fineness below 200 mesh, then pressed into a smooth, flat disc. Measurements are performed using a wavelength dispersive X-ray fluorescence spectrometer (WD-XRF) or an energy-dispersive X-ray fluorescence spectrometer (ED-XRF) in a helium atmosphere. Measurement conditions: X-ray tube voltage 40-50 kV, current 40-60 mA, scanning range covering the Kα or Lα characteristic spectral lines of Na-U. The X-ray fluorescence intensity of each element in the sample is quantitatively calculated using the national standard rock material calibration curve.

[0030] Finally, the percentage content of oxides of each major element was obtained. In this example, 10 major rock-forming element oxides were selected: SiO2, Al2O3, Fe2O3 (or FeO), CaO, MgO, K2O, Na2O, TiO2, MnO, and P2O5. The content of each component was normalized to 100% after correction for loss on ignition.

[0031] It should be noted that elemental composition data plays the role of a "knowledge anchor" in the distillation framework of this invention—it is constrained by stoichiometric rules and provides a reference for the accuracy of mineral characteristics and the adaptability to LIBS spectral features.

[0032] (3) Mineral composition data collection: X-ray diffraction (XRD) is used to obtain the volume percentage of major minerals in rock samples. XRD technology utilizes the diffraction effect of X-rays in crystalline materials to perform qualitative and quantitative phase analysis based on the position and intensity of characteristic peaks in the diffraction pattern. It is an authoritative method for analyzing the composition of rock minerals.

[0033] In practice, the prepared rock powder sample was used, and an X-ray diffractometer (Cu target, Kα radiation, λ=1.5406Å) was employed with an operating voltage of 40kV, an operating current of 40mA, a scanning range of 3°–70° 2θ, a scanning step size of 0.02°, and a scanning speed of 2° / min to obtain the XRD diffraction pattern of the sample.

[0034] After obtaining the XRD diffraction patterns of the samples, mineral phase identification and quantitative analysis were performed using Jade software. Jade software uses a PDF standard diffraction database to match and identify each diffraction peak in the patterns, determining the mineral species present in the samples. The relative content of each mineral is calculated by combining the intensity of its characteristic diffraction peaks with either full-spectrum fitting or reference intensity ratio (RIR) methods. Finally, the volume percentage content of each major mineral is obtained.

[0035] In this embodiment, 15 common rock-forming minerals were selected: quartz, potassium feldspar, plagioclase, amphibole, pyroxene, olivine, biotite, muscovite, calcite, dolomite, chlorite, kaolinite, illite, montmorillonite, and epidote, with the total content of each component being 100%.

[0036] The three types of data (LIBS spectra, elemental composition, and mineral composition) are paired and stored one by one according to the sample to form a labeled training sample set, with each sample labeled with the corresponding lithology category.

[0037] In this embodiment, the lithological categories cover 15 types of common igneous rocks (such as granite, diorite, gabbro, peridotite, rhyolite, andesite, basalt), sedimentary rocks (such as sandstone, mudstone, limestone, dolomite), and metamorphic rocks (such as gneiss, marble, quartzite, slate).

[0038] Step S2: Using the LIBS spectral data as the primary modal input and elemental composition data and mineral composition data as auxiliary knowledge modalities, a multimodal data knowledge distillation lithology identification model is constructed and trained. The multimodal data knowledge distillation lithology identification model includes a LIBS spectral encoder, an elemental encoder, a mineral encoder, and a classification head. During training, the multimodal data knowledge distillation lithology identification model adopts a cascaded feature alignment distillation strategy: first, mineral features are aligned to elemental features, using elemental features as knowledge anchors to constrain the feature learning of the mineral encoder, enabling the mineral encoder to learn mineral assemblage representations that conform to elemental stoichiometry constraints; then, spectral projection features are distilled to the aligned mineral features, using mineral features as geological semantic bridges to constrain the feature learning of the spectral encoder, enabling the spectral encoder to learn a complete "element → mineral → lithology" geological genetic chain feature representation.

[0039] like Figure 2 As shown, the multimodal data knowledge distillation lithology identification model integrates three types of data: LIBS spectral data, XRF elemental composition data, and XRD mineral composition data. The three types of data are encoded in their respective encoders to form spectral features, elemental features, and mineral features. The mineral features are reorganized and updated using the elemental features, and then the spectral features are aligned with the updated mineral features. Finally, the aligned spectral features are used for lithology identification.

[0040] Specifically, the multimodal data knowledge distillation lithology identification model includes a LIBS spectral encoder for extracting spectral features, an element encoder for encoding major elemental composition data to generate elemental features, a mineral encoder for encoding mineral composition data to generate mineral features, and a classification head for performing classification and outputting lithology identification results.

[0041] First, LIBS spectral data is input into the LIBS spectral encoder to obtain spectral projection feature vectors.

[0042] The LIBS spectral encoder is constructed using a one-dimensional convolutional neural network (1D-CNN) structure, which runs once for each input spectral vector to extract local to global features from the spectrum.

[0043] The encoder comprises three convolutional blocks: the first convolutional block uses a large kernel (size 11, stride 2) to capture broad spectral line features and baseline trends; the second convolutional block uses a medium kernel (size 7, stride 1) to extract combination patterns between adjacent feature lines; and the third convolutional block uses a small kernel (size 3, stride 1) to finely locate the peak position and peak height of individual feature lines, outputting a 256-dimensional global feature vector after adaptive average pooling. This feature vector is then mapped to a 128-dimensional feature space through a fully connected projection layer to obtain a spectral projection feature vector, used for subsequent classification and distillation alignment. The design of the spectral encoder follows a multi-scale feature extraction strategy from coarse to fine, fully considering the characteristics of LIBS spectra: dense spectral lines, sharp feature peaks, and complex baselines.

[0044] Secondly, the elemental composition data vector is input into the element encoder to obtain the elemental feature vector. The element encoder adopts a multilayer perceptron (MLP) structure, consisting of three fully connected layers. The input data dimension is 10, and the output feature dimension of each fully connected layer is 128. Each fully connected layer is followed by a ReLU activation function and a Dropout layer. The elemental composition data includes the percentage content of oxides of 10 major rock-forming elements: SiO2, Al2O3, Fe2O3, CaO, MgO, K2O, Na2O, TiO2, MnO, and P2O5. There are fixed ratio relationships between the contents of each element, determined by the mineral chemical formula. For example, quartz requires Si:O = 1:2, and plagioclase requires (Na+Ca):Al:Si to follow a specific stoichiometric ratio. The element encoder encodes these element contents into a 128-dimensional elemental feature vector, which serves as a "knowledge anchor" to provide stoichiometric constraints for mineral features.

[0045] Next, the mineral composition data vector is input into the mineral encoder to obtain the mineral feature vector. The mineral encoder adopts a multilayer perceptron (MLP) structure symmetrical to the element encoder, consisting of three fully connected layers. The input data dimension is 15, and the output feature dimension of each fully connected layer is 128. Each fully connected layer is followed by a ReLU activation function and a Dropout layer. The mineral composition data includes the volume percentage of 15 major rock-forming minerals: quartz, potassium feldspar, plagioclase, amphibole, pyroxene, olivine, biotite, muscovite, calcite, dolomite, chlorite, kaolinite, illite, montmorillonite, and epidote. The mineral encoder encodes these mineral contents into a 128-dimensional mineral feature vector, which carries the lithologically decisive feature of the mineral assemblage.

[0046] Furthermore, after obtaining spectral projection features, elemental features, and mineral features, a cascaded feature-aligned distillation strategy is implemented. The distillation process is divided into two stages, following the geological genetic chain of "element → mineral → lithology": The first stage is mineral feature reconstruction. The mean squared error loss between mineral features and elemental features is calculated, and the mineral features are aligned to the elemental feature space, as shown in the following formula:

[0047] in, This is due to mineral-element alignment loss; This is the mineral feature vector; This alignment operation uses elemental feature vectors. The physical basis for this alignment is that each rock-forming mineral has a fixed chemical formula and elemental ratio requirements. For example, quartz has the chemical formula SiO2, requiring a Si to O atomic ratio of 1:2; plagioclase has the chemical formula (Na,Ca)Al(Al,Si)Si2O8, and its Na / Ca and Al / Si ratios are limited by the solid solution components. Therefore, there is a deterministic mapping relationship between mineral composition and elemental composition, determined by the mineral's chemical formula. Aligning mineral features to elemental features essentially involves reorganizing mineral features using objective stoichiometry as a constraint, forcing the mineral encoder to learn mineral assemblage representations that conform to the objective limitations of elemental ratios. After this stage, mineral features are "reorganized" into feature expressions that are highly correlated with elemental features.

[0048] The second stage involves distilling spectral features into reconstructed mineral features. The mean square error loss between the projected spectral features and the reconstructed mineral features is calculated, and the spectral features are spatially aligned to the reconstructed mineral features. The formula is as follows:

[0049] in, For spectral-mineral distillation losses; It is the spectral projection eigenvector; This is the recombined mineral feature vector.

[0050] This stage uses reconstructed mineral features as the distillation target instead of directly using the original mineral features. The rationale is that while LIBS spectra directly reflect the elemental composition of rocks, there is no direct physical correspondence between LIBS spectra and mineral composition, and the data correlation is very low. Aligning LIBS spectra directly to the original mineral features essentially forces the model to learn a physically non-existent "spectrum → mineral" mapping, increasing learning difficulty and causing the spectral encoder to learn feature representations with unclear physical meaning. However, through a cascaded strategy, the spectral encoder indirectly inherits mineral semantics through the path of "spectrum → element (implicit) → reconstructed mineral." The spectral features learn chemical composition information consistent with the essence of physical measurements. This allows the spectral features to inherit the decisive characteristic of lithology—mineral composition. When using such spectral features for lithology identification, both physical plausibility and identification accuracy are simultaneously improved.

[0051] The distillation loss and classification loss from the two stages described above together constitute the overall optimization objective for model training. The overall loss function is expressed as:

[0052] in, Total loss; For classification loss; , , The weighting coefficients represent the loss. Through the aforementioned cascade distillation, the LIBS spectral encoder, driven by the classification task and simultaneously constrained by elemental anchor points and guided by mineral semantics, learns spectral projection features that contain a complete geological genetic chain of "element → mineral → lithology".

[0053] Finally, after completing the distillation constraint, the spectral projection features are input into the classification head. The classification head maps the 128-dimensional spectral projection features to the lithology category space through a 64-dimensional hidden layer. The specific structure of the classification head consists of a first fully connected layer, a ReLU activation function layer, a Dropout layer, and a second fully connected layer connected in sequence. The first fully connected layer maps the 128-dimensional input features to 64 dimensions, and the second fully connected layer maps the 64-dimensional input features to C dimensions, where C is the total number of lithology categories. The classification head outputs a C-dimensional logits vector, which is normalized using the Softmax function to obtain the predicted probability distribution of each lithology category, thus achieving the final lithology classification.

[0054] In this embodiment, the elemental encoder encodes the principal elemental composition as knowledge anchor features, providing stoichiometric constraints for mineral features. This ensures that the mineral assemblage representation learned by the mineral encoder must conform to the objective limitations of elemental ratios, significantly improving the physical rationality and consistency of mineral features. The mineral encoder encodes mineral composition data as mineral features and performs subsequent reorganization and updates, providing alignment targets for the LIBS spectral encoder to train and update spectral features. Ultimately, the entire lithology identification process based on LIBS spectral features is constrained by elemental and mineral knowledge. The training framework, composed of LIBS spectral, elemental composition, and mineral composition data through a cascaded feature alignment distillation strategy, not only integrates multi-level geological information such as spectral response, elemental stoichiometry, and mineral assemblage, overcoming the incompleteness of information from a single data source, but also achieves a step-by-step transfer of knowledge from the auxiliary mode to the main mode through the natural genetic chain of "element → mineral → lithology." This enables the spectral encoder to learn feature representations with a complete geological semantic chain, achieving a balance between high model accuracy and high generalization ability.

[0055] Step S3: Extract the element encoder and mineral encoder from the knowledge distillation lithology identification model of the trained multimodal data to obtain the lithology inference model, and input the LIBS spectral data of the sample to be detected into the lithology inference model to obtain the lithology identification result.

[0056] After the multimodal data knowledge distillation lithology identification model is trained, the element encoder and mineral encoder are completely removed and do not participate in subsequent inference deployment. The retained inference model consists of a LIBS spectral encoder and a classification head. The classification head used in this stage has the same network structure and parameters as the classification head described in the training stage, that is, the parameters of this module saved after optimization by the loss function during training. During inference, the 128-dimensional feature vector output by the LIBS spectral encoder is directly input into the classification head, mapped to 64 dimensions through the first fully connected layer, and then mapped to C dimensions through the second fully connected layer, outputting a lithology category logits vector. After normalization by the Softmax function, the predicted probability distribution of the lithology category is output.

[0057] During inference, only a preprocessed LIBS spectral curve needs to be input. The spectral features are extracted by the LIBS spectral encoder, mapped through the projection layer and the classification head, and the predicted lithology category is output. The entire inference process involves only one forward propagation, resulting in fast computation and a small model size. It can be deployed on the built-in processor of portable LIBS devices, edge computing terminals, or cloud inference platforms.

[0058] During the inference phase, the two-stage distillation constraint is no longer needed because, during training, elemental and mineral knowledge has been internalized step-by-step into the network parameters of the LIBS spectral encoder in a cascaded manner. The constraint of mineral features converging towards elemental anchors enables the mineral encoder to learn mineral assemblage representations that conform to stoichiometry, while the constraint of spectral features distilling towards mineral features allows the spectral encoder to indirectly inherit this geological semantic chain. The model no longer needs to explicitly calculate feature distances but directly uses the already learned features that contain the complete geological genetic chain of "element → mineral → lithology" for classification.

[0059] Preferably, it may also include step S4: outputting and storing the lithology identification results.

[0060] The lithology identification results output by the model include: (1) Lithology name: The name of the lithology category with the highest prediction probability in the model, such as "granite", "diorite", "sandstone", etc.; (2) Category probability distribution: The predicted probability vector of the model for all C lithological categories is used to evaluate the prediction confidence. The more concentrated the probability, the higher the confidence. (3) Confidence score: The confidence index (between 0 and 1) is calculated based on the maximum probability value and probability entropy of the predicted probability distribution. When the confidence score is lower than the set threshold (e.g., 0.7), the system prompts "The result is uncertain and manual review is recommended".

[0061] The identification results are linked and stored with information such as sample number, sampling location, tunnel mileage marker, and collection time to form a lithology identification database, which can be used for tunnel surrounding rock classification and dynamic adjustment of construction support parameters. In actual tunnel construction, for each cycle of advance, representative rock blocks are collected after muck removal or drilling and blasting at the working face for LIBS measurement and identification. By repeatedly executing the steps of acquiring spectral data, inputting model identification, and outputting and saving the identification results, identification can be achieved on-the-spot, providing real-time geological information support for construction decisions.

[0062] This embodiment considers that the mineral composition of a rock directly determines its lithological classification, and the elemental composition is the material basis of the mineral composition. There is a clear physical genetic chain among the three: element-mineral-lithology. The essence of LIBS spectroscopy is the optical characterization of elemental composition. By combining knowledge at the elemental and mineral levels to guide model learning, the recognition accuracy and interpretability will be significantly improved.

[0063] The inference model trained by this invention has a small number of parameters and a fast calculation speed. It can be deployed in the built-in processor or edge computing terminal of a portable LIBS device to meet the engineering needs of tunnel advanced geological prediction, rapid identification of tunnel face, and identification of borehole cores.

[0064] Example 2 In one or more embodiments, a tunnel surrounding rock lithology identification system based on multimodal data knowledge distillation is disclosed, such as... Figure 3 As shown, it specifically includes: The data acquisition module is configured to acquire LIBS spectral data, major elemental composition data, and mineral composition data of rock samples. The model building module is configured to: use the LIBS spectral data as the primary modal input and elemental composition data and mineral composition data as auxiliary knowledge modalities to construct and train a multimodal data knowledge distillation lithology identification model; the multimodal data knowledge distillation lithology identification model includes a LIBS spectral encoder, an elemental encoder, a mineral encoder, and a classification head; during training, the multimodal data knowledge distillation lithology identification model adopts a cascaded feature alignment distillation strategy: first, mineral features are aligned to elemental features, using elemental features as knowledge anchors to constrain the feature learning of the mineral encoder, enabling the mineral encoder to learn mineral assemblage representations that conform to elemental stoichiometry constraints; then, spectral projection features are distilled to the aligned mineral features, using mineral features as geological semantic bridges to constrain the feature learning of the spectral encoder, enabling the spectral encoder to learn geological genetic chain feature representations; The lithology identification module is configured to: extract the element encoder and mineral encoder from the lithology identification model by knowledge distillation of the trained multimodal data to obtain a lithology inference model, and input the LIBS spectral data of the sample to be detected into the lithology inference model to obtain the lithology identification result.

[0065] When deployed on-site, the data acquisition module is integrated into a portable LIBS measurement unit. The LIBS spectrometer uses a handheld or small benchtop LIBS system, equipped with a built-in laser and miniature spectrometer, and a built-in touchscreen display for operation control and result display. During sample measurement, the spectrometer probe is aimed at a fresh rock section, the trigger is pulled to activate the laser, and the spectrometer automatically completes spectral acquisition and preprocessing. For elemental analysis, a handheld XRF analyzer can be used for on-site supplementary determination, or representative samples can be collected and sent to the laboratory for precise analysis using wavelength dispersive XRF. For mineral analysis, representative samples must be collected, sealed, and sent to the laboratory for XRD analysis to construct the training set.

[0066] The model building module typically runs in the cloud or on a high-performance workstation, utilizing GPUs to accelerate the training process. Once training is complete, the lightweight inference model is deployed to the field terminal.

[0067] The lithology identification module can be deployed on the built-in processor of portable LIBS devices, supporting offline inference without a network connection.

[0068] Preferably, it may also include a data recording module for associating and storing the identification results with information such as sample number, sampling location, tunnel mileage marker, and collection time, forming a lithology identification database that supports the querying, statistics, and visualization of historical data. This module is provided in the form of a mobile app or web platform and supports data integration with the tunnel engineering management system.

[0069] In this embodiment, the data flow between modules is as follows: After acquiring LIBS spectral data, the multimodal data acquisition module transmits the data to the lithology identification and inference module via Bluetooth or USB for real-time inference; the inference results are automatically pushed to the data recording module for archiving and visualization. For updating the model training module, once a sufficient number of new samples have been accumulated, the model is retrained in the cloud, and the updated lightweight model is pushed to the on-site identification terminal via OTA (Over-The-Air) updates, enabling continuous iterative optimization of the model.

[0070] Example 3 This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps of the above-described method for identifying the lithology of tunnel surrounding rock based on multimodal data knowledge distillation.

[0071] The electronic device can be a built-in embedded system of a portable LIBS spectrometer, an industrial tablet PC, a laptop computer, or a cloud server. For field deployment scenarios, an industrial-grade ruggedized tablet PC is preferred, powered by a built-in battery, connecting to the LIBS spectrometer via USB or Bluetooth to run lithology identification inference software.

[0072] The memory can be a solid-state drive (SSD), an embedded multimedia card (eMMC), or a flash memory chip, used to store trained model parameters and a lithology identification database. The processor can be an ARM Cortex-A series processor or an embedded processor with equivalent computing power. When the processor has a neural network acceleration unit, it can further accelerate the inference calculation of the LIBS spectral encoder, achieving millisecond-level response.

[0073] Example 4 This embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the tunnel surrounding rock lithology identification method based on multimodal data knowledge distillation described above.

[0074] The computer-readable storage medium can be any medium capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), or random access memory (RAM). The program code stored in the storage medium includes: a LIBS spectral preprocessing module, which implements Savitzky-Golay smoothing, asLS baseline correction, and normalization; a LIBS spectral encoder and classification head inference module, which implements spectral feature extraction and lithological classification; and a data recording interface module, which enables automatic storage and synchronization of recognition results.

[0075] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0076] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0077] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0078] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying the lithology of tunnel surrounding rock based on multimodal data knowledge distillation, characterized in that, include: Acquire LIBS spectral data, major elemental composition data, and mineral composition data of rock samples; Using the LIBS spectral data as the primary modal input and elemental composition data and mineral composition data as auxiliary knowledge modalities, a multimodal data knowledge distillation lithology identification model is constructed and trained. This model includes a LIBS spectral encoder, an elemental encoder, a mineral encoder, and a classification head. During training, the model employs a cascaded feature alignment distillation strategy: first, mineral features are aligned to elemental features, using elemental features as knowledge anchors to constrain the mineral encoder's feature learning, enabling it to learn mineral assemblage representations that conform to elemental stoichiometry constraints; then, spectral projection features are distilled onto the aligned mineral features, using mineral features as geological semantic bridges to constrain the spectral encoder's feature learning, enabling it to learn geological genetic chain feature representations. The element encoder and mineral encoder are stripped from the knowledge distillation lithology identification model of the trained multimodal data to obtain the lithology inference model. The LIBS spectral data of the sample to be detected is then input into the lithology inference model to obtain the lithology identification result.

2. The method for identifying tunnel surrounding rock lithology based on multimodal data knowledge distillation as described in claim 1, characterized in that, The LIBS spectral data is input into the LIBS spectral encoder to obtain the spectral projection feature vector, specifically: The LIBS spectral encoder is constructed using a one-dimensional convolutional neural network structure, which runs once for each input spectral vector to extract local to global features from the spectrum. The spectral encoder includes three convolutional blocks with sequentially decreasing kernel size, and a fully connected projection layer that maps to a 128-dimensional feature space to obtain a spectral projection feature vector.

3. The method for identifying tunnel surrounding rock lithology based on multimodal data knowledge distillation as described in claim 1, characterized in that, The element composition data is input into the element encoder to obtain element features, specifically: The element encoder adopts a multilayer perceptron structure, consisting of three fully connected layers, each followed by a ReLU activation function and a Dropout layer; The elemental composition data includes the percentage content of several rock-forming element oxides.

4. The method for identifying tunnel surrounding rock lithology based on multimodal data knowledge distillation as described in claim 1, characterized in that, The mineral composition data is input into the mineral encoder to obtain the mineral feature vector, specifically: The mineral encoder adopts a multilayer perceptron structure that is symmetrical to that of the element encoder. It consists of three fully connected layers, each followed by a ReLU activation function and a Dropout layer. The mineral composition data includes the volume percentage of several rock-forming minerals.

5. The method for identifying tunnel surrounding rock lithology based on multimodal data knowledge distillation as described in claim 1, characterized in that, The process of aligning mineral features with elemental features, using elemental features as knowledge anchors to constrain the feature learning of the mineral encoder, enables the mineral encoder to learn mineral assemblage representations that conform to elemental stoichiometry constraints. Specifically, the mean squared error loss between mineral features and elemental features is calculated, and the mineral features are aligned to the elemental feature space, as follows: in, This is due to mineral-element alignment loss; This is the mineral feature vector; is the eigenvector of the element.

6. The method for identifying tunnel surrounding rock lithology based on multimodal data knowledge distillation as described in claim 1, characterized in that, The process involves distilling the spectral projection features towards the aligned mineral features, using the mineral features as a geological semantic bridge to constrain the feature learning of the spectral encoder, enabling the spectral encoder to learn a geological genetic chain feature representation. Specifically, the mean squared error loss between the spectral projection features and the reconstructed mineral features is calculated, and the spectral features are aligned to the reconstructed mineral feature space, as follows: in, For spectral-mineral distillation losses; It is the spectral projection eigenvector; This is the recombined mineral feature vector.

7. The method for identifying tunnel surrounding rock lithology based on multimodal data knowledge distillation as described in claim 1, characterized in that, The lithology reasoning model includes a LIBS spectral encoder and a classification head; it acquires preprocessed LIBS spectral data of the sample to be tested, extracts spectral features through the LIBS spectral encoder, and outputs lithology category identification results through the classification head.

8. A tunnel surrounding rock lithology identification system based on multimodal data knowledge distillation, characterized in that, include: The data acquisition module is configured to acquire LIBS spectral data, major elemental composition data, and mineral composition data of rock samples. The model building module is configured to: use the LIBS spectral data as the primary modal input and elemental composition data and mineral composition data as auxiliary knowledge modalities to construct and train a multimodal data knowledge distillation lithology identification model; the multimodal data knowledge distillation lithology identification model includes a LIBS spectral encoder, an elemental encoder, a mineral encoder, and a classification head; during training, the multimodal data knowledge distillation lithology identification model adopts a cascaded feature alignment distillation strategy: first, mineral features are aligned to elemental features, using elemental features as knowledge anchors to constrain the feature learning of the mineral encoder, enabling the mineral encoder to learn mineral assemblage representations that conform to elemental stoichiometry constraints; then, spectral projection features are distilled to the aligned mineral features, using mineral features as geological semantic bridges to constrain the feature learning of the spectral encoder, enabling the spectral encoder to learn geological genetic chain feature representations; The lithology identification module is configured to: extract the element encoder and mineral encoder from the lithology identification model by knowledge distillation of the trained multimodal data to obtain a lithology inference model, and input the LIBS spectral data of the sample to be detected into the lithology inference model to obtain the lithology identification result.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the tunnel surrounding rock lithology identification method based on multimodal data knowledge distillation as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the tunnel surrounding rock lithology identification method based on multimodal data knowledge distillation as described in any one of claims 1-7.

Images