Underground soil layer identification method based on static cone penetration test device

Abstract

The invention discloses an underground soil layer identification method based on a static cone penetration test device. The method comprises the steps that the static cone penetration test device collects an original spectral data curve; preprocessing the original spectral data curve to obtain a standardized spectral reflectivity curve and forming a data set; manually annotating the data set and dividing the annotated data set into a training set, a verification set and a test set; training and verifying the constructed mineral recognition model by adopting a training set and a verification set to obtain a trained mineral recognition model; inputting the test set into the trained mineral identification model, and outputting prediction probabilities of n minerals corresponding to a spectral reflectivity curve; and according to the prediction probability of the n minerals, performing soil layer category judgment on the collected soil body to obtain a depth-soil layer layering information table. According to the underground soil layer identification method, on the premise that a soil body structure is not damaged, high-precision identification of fine soil layer changes is achieved through the trained mineral identification model, and manual intervention and subjective judgment errors are greatly reduced.

Description

Technical Field

[0001] This invention relates to the fields of geological exploration and soil analysis technology, and specifically to a method for identifying underground soil layers based on a static cone penetration test device. Background Technology

[0002] With the continuous expansion of land and marine engineering construction, the importance of geological exploration in the early stages of project construction is becoming increasingly prominent. Soil layer identification, as a fundamental step in geological exploration, directly affects the scientific nature of subsequent engineering design and the safety of construction. Traditional soil layer identification methods mainly rely on manual sampling, laboratory physical property analysis, and engineers' experience-based judgment. Although existing methods have high accuracy, they generally suffer from low work efficiency, long response cycles, cumbersome operation procedures, and significant interference from subjective factors, making it difficult to meet the current engineering demands for high efficiency and high accuracy.

[0003] In recent years, with the rapid development of sensor and remote sensing technologies, hyperspectral imaging and non-imaging techniques have been introduced into the field of soil research. Hyperspectral non-imaging techniques, in particular, can sensitively reflect the physical, chemical, and mineral composition characteristics of materials by acquiring spectral reflectance information across multiple continuous bands. Compared to traditional RGB imagery or limited-band sensors, hyperspectral imaging offers greater information dimensionality and demonstrates broad application potential in soil mineral identification, moisture analysis, and organic matter content estimation.

[0004] However, most current hyperspectral studies focus on surface soil analysis, commonly used in agriculture and ecological environment monitoring. For information on subsurface soil layers needed in engineering geology, especially the fine division between multiple layers with different physical structures (such as clay, sand, and saturated layers) within a profile, mature and reliable methods are still lacking. Furthermore, existing technologies largely rely on professionals manually interpreting spectral images, resulting in lagging information processing and the absence of automated and intelligent analytical pathways.

[0005] Meanwhile, the rise of machine learning technology has provided new approaches for the rapid processing and pattern recognition of complex spectral data. By modeling and training on the spectral characteristics of known soil samples, machine learning models can learn the spectral characteristics of different soil layers, thereby achieving efficient identification of unknown samples. Combined with hyperspectral non-imaging techniques, machine learning can achieve real-time classification and dynamic updating of profile soil layers, greatly improving the efficiency and objectivity of soil layer identification.

[0006] In summary, there is an urgent need for a novel soil layering method that integrates hyperspectral non-imaging static cone penetration testing technology and machine learning algorithms. This method should enable real-time, multi-level soil layer identification without damaging the soil structure and should also output clear analytical results to assist in engineering decision-making and soil quality assessment. Summary of the Invention

[0007] This invention provides a method for identifying underground soil layers based on a static cone penetration test device. The method employs a one-dimensional convolutional neural network model based on spectral reflectance curves to identify the specific mineral composition of spectral data, and then classifies the soil layers based on mineral distribution, thus completing the automatic classification and identification of typical underground soil layer categories.

[0008] Therefore, the present invention adopts the following technical solution:

[0009] A method for identifying underground soil layers based on a static cone penetration test device, the specific steps of which are as follows:

[0010] S1, the static cone penetration test device is inserted into the soil, and the original spectral data curve is collected once for every hcm increase in the depth of the probe into the soil.

[0011] S2, preprocess the original spectral data curves to obtain standardized spectral reflectance curves, and combine the standardized spectral reflectance curves corresponding to different depths into a dataset.

[0012] S3. Manually label the spectral reflectance curves in the dataset, and divide the labeled spectral reflectance curves into training set, validation set and test set in a 6:2:2 ratio.

[0013] S4 uses a one-dimensional convolutional neural network to construct a mineral identification model;

[0014] S5. The mineral identification model is trained using the training set. The output of the mineral identification model is a spectral reflectance curve corresponding to the predicted probability of n minerals. The mineral identification model is validated using the validation set to obtain the trained mineral identification model.

[0015] S6. Input the spectral reflectance curves of the test set into the trained mineral recognition model and output a spectral reflectance curve corresponding to the predicted probability of n minerals.

[0016] S7. Based on the predicted probabilities of n minerals, determine the soil layer category of the collected soil and obtain a depth-soil layer information table.

[0017] In step S2, the preprocessing specifically involves: using a whiteboard and dark current to correct the original spectral data curve to obtain a corrected spectral data curve; using a reflectance conversion formula to convert the corrected spectral data curve into a spectral reflectance curve; and using Savitzky-Golay filtering to smooth and denoise the spectral reflectance curve to obtain a standardized spectral reflectance curve.

[0018] In step S3, the labels of the annotated spectral reflectance curves include: kaolinite, montmorillonite, illite, and mica.

[0019] In step S4, the one-dimensional convolutional neural network includes: an input layer, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a first dropout layer, a Flatten layer, a fully connected layer, a second dropout layer, and an output layer connected in sequence.

[0020] The input layer is used to input the spectral reflectance curve; the first convolutional layer is used to extract features from the output of the input layer, capturing features such as absorption valleys within multiple spectral bands; the first max-pooling layer is used to downsample the output of the first convolutional layer, which reduces the feature dimension, retains important information, and improves the robustness of the mineral identification model; the second convolutional layer is used to extract deep spectral features from the output of the first max-pooling layer, enabling the identification of finer-grained or composite features; the second max-pooling layer is used to perform dimensionality reduction on the output of the second convolutional layer to reduce the computational load of the mineral identification model. This improves the robustness of the mineral identification model. The first dropout layer is used to prevent overfitting by randomly discarding some neurons during training. The Flatten layer flattens the multidimensional feature tensor output by the first dropout layer into a one-dimensional vector and uses it as the input to the fully connected layer. The first fully connected layer is used to fuse multiple local features and construct a global representation. The second dropout layer is used to further suppress overfitting. The output layer is a fully connected layer with the number of nodes equal to the total number of mineral categories, n, and the activation function is Softmax. The output of the output layer is a spectral reflectance curve corresponding to the predicted probability of each mineral.

[0021] The first convolutional layer has a kernel size of 9, a number of kernels of 64, and an activation function of ReLU; the first max pooling layer has a pooling window size of 2; the second convolutional layer has a kernel size of 5, a number of kernels of 128, and an activation function of ReLU; the first dropout layer has a dropout rate of 0.3; the first fully connected layer contains 128 neurons and has an activation function of ReLU; the second dropout layer has a dropout rate of 0.3; and the output layer has an activation function of Softmax.

[0022] The method for determining the soil layer type in step S7 is as follows:

[0023] If, among the predicted probabilities of n minerals corresponding to the spectral reflectance curve, the probability of kaolinite is >0.7 and the probability of illite is <0.2, then the soil type at the penetration depth corresponding to the spectral reflectance curve in the collected soil is: clay soil.

[0024] If, among the predicted probabilities of n minerals corresponding to the spectral reflectance curve, the probability of illite mineral + the probability of mica mineral > 0.6 and the probability of montmorillonite mineral < 0.3, then the soil quality at the penetration depth corresponding to the spectral reflectance curve in the collected soil is: well-graded sand.

[0025] If the probability of montmorillonite is greater than 0.5 among the predicted probabilities of n minerals corresponding to the spectral reflectance curve, then the soil type at the penetration depth corresponding to the spectral reflectance curve in the collected soil is silty soil.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. Improved accuracy: The mineral identification model of this invention can achieve higher accuracy in identifying subtle changes in soil layers.

[0028] 2. Automation and intelligence: The underground soil layer identification method of the present invention is fully automated from collection to classification, which greatly reduces human intervention and subjective judgment errors.

[0029] 3. Non-destructive detection: The underground soil layer identification method of the present invention does not require damage to the soil structure throughout the entire process, making it suitable for high-risk scenarios such as urban foundations, subway projects, and slope monitoring.

[0030] 4. Strong model transferability: Under the premise of sufficient training sample coverage, the mineral identification model of the present invention can be applied to different regions and different types of engineering projects, and is suitable for geological assessment engineering scenarios that require mineral composition interpretation or are in a geographic processing context.

[0031] 5. Extensibility: The mineral identification model of this invention can also be extended to the analysis of features such as moisture content, organic matter, and pollutant elements, to achieve multi-objective and multi-dimensional comprehensive soil analysis. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the static cone penetration test apparatus according to an embodiment of the present invention;

[0033] Figure 2 for Figure 1 A schematic diagram of the internal structure of the optical fiber assembly in a static penetration test apparatus;

[0034] Figure 3 This is a diagram showing the distribution of the transmitting and receiving fiber arrays in the static penetration test apparatus of the present invention.

[0035] Figure 4 This is a flowchart of the underground soil layer identification method of the present invention;

[0036] Figure 5 This is a graph of the original spectral data from Example 2;

[0037] Figure 6 This is a typical hyperspectral absorption band characteristic diagram from Example 2;

[0038] Figure 7 This is a columnar layered diagram from Example 2.

[0039] In the picture:

[0040] 1. Hyperspectral component, 2. Signal transmission cable, 3. Friction sleeve sensor, 4. Friction sleeve, 5. Inclinometer, 6. Cone tip, 7. Pressure sensor, 8. Cone tip resistance sensor, 9. Sealing ring, 10. Probe, 11. Optical window, 12. Integrated lens, 13. Fiber optic assembly, 131. Transmitting fiber, 132. Receiving fiber array, 133. Coupling lens assembly, 134. Light source, 135. Heat dissipation assembly, 14. Fixing ring, 15. Hollow inner tube, 16. First annular groove, 17. Second annular groove. Detailed Implementation

[0041] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0042] Example 1

[0043] like Figure 1 As shown, the static cone penetration test apparatus used in this invention includes: a probe 10, a hyperspectral component 1, and a static cone penetration component. Wherein:

[0044] The main body of the probe rod 10 is a hollow cylinder without a bottom surface. Four equally spaced openings of equal size are provided on the circumference of the probe rod near the bottom for installing the optical window 11.

[0045] The hyperspectral component 1 is installed inside the probe 10 and near the bottom of the probe. There are four hyperspectral components 1, each including: one fiber optic group 13, one integrated lens 12, and one optical window 11. Wherein:

[0046] See Figure 2 Each fiber optic group 13 includes: a light source 134, a heat dissipation assembly 135, a coupling lens group 133, a transmitting fiber 131, and a receiving fiber optic array 132. Wherein:

[0047] The light source 134 is used to emit spatial light;

[0048] The heat dissipation component 135 includes heat dissipation fins and a heat insulation layer. The heat dissipation component 135 is used to block the heat generated by the light source 134 and prevent thermal deformation of the coupling lens group 133.

[0049] The coupling lens group 133 is used to perform spatial light-to-fiber mode conversion on the spatial light emitted by the light source 134; the base of the coupling lens group is made of a material with a low coefficient of thermal expansion (CTE) to suppress mechanical deformation caused by temperature changes, thereby ensuring the accuracy of spectral measurement.

[0050] The transmitting optical fiber 131 is used to receive the optical fiber mode beam output by the coupling lens group 133, and transmit the received beam to the integrated lens 12, thereby illuminating the soil.

[0051] The receiving fiber array 132 is used to receive the reflected light beam from the soil, such as Figure 3 As shown, the receiving fiber array 132 consists of six receiving fibers.

[0052] In each fiber optic group 13, a heat dissipation component 135 is attached to the surface of a light source 134. The light source 134 is kept at a certain distance from the coupling lens group 133. The coupling lens group 133 is located at the center of the inner sheath of the fiber optic group. The coupling lens group 133 is connected to the transmitting fiber 131, which is located at the center of the inner sheath of the fiber optic group. The receiving fiber array 132 is located inside the outer sheath. The six receiving fibers of the receiving fiber array 132 are evenly distributed in a circumferential direction with the transmitting fiber as the center.

[0053] The integrated lens 12 is a plano-convex lens. The integrated lens 12 is used to collimate the beam output from the fiber optic group 13 into a parallel beam and focus the parallel beam through the optical window 11 to illuminate the soil.

[0054] The optical window 11 is used to protect the integrated lens 12 and the fiber optic assembly 13 from the influence of the external environment. The optical window 11 has high light transmittance and high mechanical strength. The optical window 11 is made of sapphire material. Anti-reflection coatings are provided on both the inner and outer surfaces of the optical window to improve transmittance and reduce the loss of reflected light beams.

[0055] In this embodiment, the integrated lens 12 is a fiber collimating lens of brand name LS-FOL or a fiber focusing lens of brand name LS-MC manufactured by LiSen Optics; the fiber optic assembly 13 is an ASD fiber manufactured by Malvern Panalytical.

[0056] In the hyperspectral component 1, four optical windows 11 are located at the four openings of the probe rod and are flush with the outer wall of the probe rod. An integrated lens 12 is provided between the optical windows 11 and the fiber optic assembly 13, and a fixed distance is maintained between the integrated lens 12 and the optical windows 11. The position of the fiber optic assembly 13 is fixed by a retaining ring 14 so that when using the static cone penetration test device, the spatial light emitted by the light source 134 travels along a preset optical path. The receiving fiber array receives the reflected beam of the soil and transmits it to the external host.

[0057] The static cone penetration test assembly includes: a signal transmission cable 2, a friction sleeve sensor 3, a friction sleeve 4, an inclinometer 5, a cone tip 6, a pressure sensor 7, a cone tip resistance sensor 8, and a sealing ring 9.

[0058] Among them, the friction sleeve sensor 3 is used to measure the lateral friction resistance experienced by the static penetration test device when penetrating the soil. The friction sleeve sensor 3 is annular; the friction sleeve 4 is a hollow circular tube, and the outer diameter of the friction sleeve 4 is consistent with the maximum diameter of the cone tip 6. Figure 1 (Diameter of the cylindrical surface at the top of the cone tip); Inclinometer 5 is used to detect the verticality of the probe and provide inclination data to correct the resistance, side friction resistance and penetration depth of the probe tip 6; Cone tip 6 is a solid cone; Pressure sensor 7 is used to measure pore water pressure; Cone tip resistance sensor 8 is used to measure the cone tip resistance, and cone tip resistance sensor 8 is annular; The side friction resistance, cone tip resistance and pore water pressure measured by friction sleeve sensor 3, pressure sensor 7 and cone tip resistance sensor 8 are transmitted to the external host via signal transmission cable 2 through microcircuit signals.

[0059] In the static penetration test assembly, one end of the friction sleeve 4 is fixedly connected to the end of the probe 10 near the bottom surface, and a sealing ring 9 is installed at the interface; the other end of the friction sleeve 4 is fixedly connected to the circular surface of the cone tip 6, and a sealing ring 9 is installed at the interface; the hollow inner tube 15 is fixedly installed in the hollow position inside the friction sleeve 4, and the outer wall of the hollow inner tube 15 is not completely fitted with the inner wall of the friction sleeve 4, leaving a first annular groove 16 and a second annular groove 17. The friction sleeve sensor 3 is located in the second annular groove and is fixed on the outer wall of the hollow inner tube; the cone tip resistance sensor 8 is located in the first annular groove and is fixed on the outer wall of the hollow inner tube; the inclinometer 5 is located in the middle of the cavity of the hollow inner tube, and the inclinometer 5 is connected to the end of the signal transmission cable 2, the other end of the signal transmission cable 2 is connected to the external host; the pressure sensor 7 is located at the bottom of the cavity of the hollow inner tube and near the cone tip 6; the friction sleeve sensor 3, the pressure sensor 7 and the cone tip resistance sensor 8 are all connected to the signal transmission cable 2. The sealing rings 9 installed in multiple locations are used to prevent liquid from entering the static cone penetration test device.

[0060] In the static cone penetration test apparatus, the optical fiber group 13 transmits spectral data to the external host, and the signal transmission cable 2 transmits the side friction resistance, cone tip resistance and pore water pressure to the external host, which then performs data analysis and processing.

[0061] The cone angle, cone base cross-sectional area, and surface area of ​​the friction sleeve 4 of cone tip 6 all meet the requirements of GB / T 50267-2018 Static cone penetration test technical standard, ISO 22476-1 Geotechnical Engineering Investigation Specification and ASTM D5775-20.

[0062] Example 2

[0063] like Figure 4 As shown, a method for identifying underground soil layers based on a static cone penetration test device includes the following specific steps:

[0064] S1, Acquire raw spectral data curve:

[0065] The static cone penetration test device described in Example 1 is inserted into the soil. Raw spectral data is collected once for every h cm increase in the depth of the probe (each penetration). In this example, raw spectral data is collected every 10 cm.

[0066] S2, preprocessing the raw spectral data curves, includes the following steps:

[0067] The original spectral data curve was corrected using a whiteboard and dark current to obtain the corrected spectral data curve;

[0068] The corrected spectral data curve is converted into a spectral reflectance curve using a reflectance conversion formula, such as... Figure 5 As shown;

[0069] Savitzky-Golay filtering was used to smooth and denoise the spectral reflectance curves, resulting in standardized spectral reflectance curves. The standardized spectral reflectance curves corresponding to different depths were then combined into a dataset.

[0070] The reflectance conversion formula is as follows:

[0071]

[0072] In the formula, R(λ) is the reflectance of the tested mineral at wavelength λ in the corrected spectral data curve; S(λ) is the original spectral value of the tested mineral in the corrected spectral data curve; W(λ) is the spectral value of the white board (high reflectance standard reference board); and D(λ) is the dark current, which is the spectral value of the tested mineral at wavelength λ under completely dark conditions in the corrected spectral data curve (sensor noise baseline).

[0073] S3, label and partition the dataset sequentially:

[0074] Based on the typical hyperspectral absorption band feature table, the spectral reflectance curves in the dataset were manually labeled. The labels included: kaolinite, montmorillonite, illite, and muscovite. The labeled spectral reflectance curves were then divided into training, validation, and test sets in a 6:2:2 ratio. The typical hyperspectral absorption band feature table was obtained based on the USGS hyperspectral library, the JHU spectral library, and field sampling comparison experience, as detailed in Table 1. Figure 6 As shown.

[0075] Table 1

[0076]

[0077] In Table 1, the absorption bands represent the locations of absorption peaks (especially in the near-infrared and short-wave infrared regions) of common minerals in the soil.

[0078] S4, Construct a mineral identification model:

[0079] The mineral identification model employs a one-dimensional convolutional neural network (1D-CNN), which comprises: an input layer, a first convolutional layer, a first max-pooling layer (Pool1), a second convolutional layer, a second max-pooling layer, a first dropout layer, a Flatten layer, a fully connected layer, a second dropout layer, and an output layer, connected sequentially. Wherein:

[0080] The input layer is used to input spectral reflectance curves.

[0081] The first convolutional layer is used to extract features from the output of the input layer, capturing features such as absorption valleys within a range of 9 bands. The kernel size of the first convolutional layer is 9, the number of kernels is 64, and the activation function is ReLU.

[0082] The first max pooling layer is used to downsample the output of the first convolutional layer. Downsampling can reduce the feature dimension, retain important information, and improve the robustness of building the mineral recognition model. The pooling window size of the first max pooling layer is 2.

[0083] The second convolutional layer (Conv2) is used to extract deep spectral features from the output of the first max pooling layer. The second convolutional layer can identify finer-grained or complex features. The kernel size of the second convolutional layer is 5, the number of kernels is 128, and the activation function is ReLU.

[0084] The second max pooling layer (Pool2) is used to perform dimensionality reduction on the output of the second convolutional layer to reduce the computational cost of the mineral identification model and improve its robustness.

[0085] The first dropout layer (Dropout1) is used to prevent overfitting, with a dropout rate of 0.3. A dropout rate of 0.3 means that some neurons are randomly dropped with a 30% probability during training.

[0086] The Flatten layer is used to flatten the multidimensional feature tensor output by the first dropout layer into a one-dimensional vector, which is then used as the input to the fully connected layer.

[0087] The first fully connected layer (Dense1) contains 128 neurons with the ReLU activation function, and is used to fuse multiple local features and construct a global representation.

[0088] The second dropout layer (Dropout2) is used to further suppress overfitting, with a dropout rate of 0.3.

[0089] The output layer is a fully connected layer with the number of nodes equal to the total number of mineral categories, n. The activation function is Softmax, and the output is a spectral reflectance curve corresponding to the predicted probability of each mineral.

[0090] S5, Training the mineral recognition model:

[0091] The mineral recognition model is trained using a training set. The model outputs a spectral reflectance curve representing the predicted probabilities of n minerals. The model is then validated using a validation set, resulting in a fully trained model. During training, the model automatically detects discrepancies in the spectral reflectance curves.

[0092] In this embodiment, the maximum number of training rounds is 100, and the number of spectral reflectance curves input into the mineral recognition model in each batch is 64. The mineral recognition model uses the Adam optimizer with a learning rate of 0.001 and a loss function of categorical crossentropy. The early stopping strategy is as follows: the loss function value corresponding to the validation set is used as the monitoring indicator. If the loss function value of the mineral recognition model does not decrease after several consecutive rounds of training, the training is terminated early to prevent overfitting.

[0093] S6, Testing the mineral identification model:

[0094] The spectral reflectance curves of the test set are input into the trained mineral recognition model, and the trained mineral recognition model outputs a spectral reflectance curve corresponding to the predicted probability of n minerals.

[0095] S7, Determine soil layer type: Based on the predicted probability of n minerals, determine the soil layer type of the collected soil, obtain the depth-soil layer information table, and draw a columnar stratification diagram based on the depth-soil layer information table.

[0096] In one embodiment of the present invention, a method for determining the soil layer type of the collected soil is as follows:

[0097] If, among the predicted probabilities of n minerals corresponding to the spectral reflectance curve, the probability of kaolinite is >0.7 and the probability of illite is <0.2, then the soil type at the penetration depth corresponding to the spectral reflectance curve in the collected soil is: clay soil (CL).

[0098] If, among the predicted probabilities of n minerals corresponding to the spectral reflectance curve, the probability of illite mineral + the probability of mica mineral > 0.6 and the probability of montmorillonite mineral < 0.3, then the soil quality at the penetration depth corresponding to the spectral reflectance curve in the collected soil is: well-graded sand (SW).

[0099] If the probability of montmorillonite is >0.5 among the predicted probabilities of n minerals corresponding to the spectral reflectance curve, then the soil type at the penetration depth corresponding to the spectral reflectance curve in the collected soil is: silty sand (SM).

[0100] In this embodiment, a certain engineering foundation in a certain region was selected as the experimental sample area. The original spectral data curves of the soil at four locations in the experimental sample area were collected using a static cone penetration test device. The maximum penetration depth of the static cone penetration test device at each location was 4m, and h was 10cm. The data set included 160 standardized spectral reflectance curves.

[0101] In this embodiment, the prediction metrics of the trained mineral identification model are shown in Table 2, and the confusion matrix of the trained mineral identification model is shown in Table 3.

[0102] Table 2

[0103]

[0104] Table 3

[0105] Real / Predicted Kaolinite Montmorillonite illite mica total Kaolinite 73 2 3 2 80 Montmorillonite 4 69 5 2 80 illite 2 3 74 1 80 mica 3 2 2 73 80 total 82 76 84 78 320

[0106] According to Tables 2 and 3, the classification accuracy of the trained mineral identification model is basically over 90%, and the characteristic bands of various types basically match the standard absorption peaks: for example, for kaolinite, focus on the area around 2200nm, and for montmorillonite, focus on the areas around 1900nm and 1400nm.

[0107] In this embodiment, the depth-soil layer stratification information table is as follows: 0–1.6m is clayey soil (CL, mainly kaolinite); 1.6–2.7m is silty soil (SM, with very obvious montmorillonite); 2.7–4.0m is well-graded sand (SW, with a high proportion of illite / mica); the columnar stratification diagram drawn based on the depth-soil layer stratification information table is as follows. Figure 7 As shown.

Claims

1. A method for identifying underground soil layers based on a static cone penetration test device, characterized in that, The specific steps are as follows: S1, the static cone penetration test device is inserted into the soil, and the original spectral data curve is collected once for every hcm increase in the depth of the probe into the soil. S2, preprocess the original spectral data curves to obtain standardized spectral reflectance curves, and combine the standardized spectral reflectance curves corresponding to different depths into a dataset. S3, Manually label the spectral reflectance curves in the dataset, and divide the labeled spectral reflectance curves into training set, validation set and test set; S4 uses a one-dimensional convolutional neural network to construct a mineral identification model; S5. The mineral identification model is trained using the training set. The output of the mineral identification model is a spectral reflectance curve corresponding to the predicted probability of n minerals. The mineral identification model is validated using the validation set to obtain the trained mineral identification model. S6. Input the spectral reflectance curves of the test set into the trained mineral recognition model and output a spectral reflectance curve corresponding to the predicted probability of n minerals. S7. Based on the predicted probabilities of n minerals, determine the soil layer category of the collected soil and obtain a depth-soil layer stratification information table.

2. The method for identifying underground soil layers according to claim 1, characterized in that, In step S2, the preprocessing specifically involves: using a whiteboard and dark current to correct the original spectral data curve to obtain a corrected spectral data curve; using a reflectance conversion formula to convert the corrected spectral data curve into a spectral reflectance curve; and using Savitzky-Golay filtering to smooth and denoise the spectral reflectance curve to obtain a standardized spectral reflectance curve.

3. The method for identifying underground soil layers according to claim 1, characterized in that: In step S3, the labels of the annotated spectral reflectance curves include: kaolinite, montmorillonite, illite, and mica.

4. The method for identifying underground soil layers according to claim 1, characterized in that, In step S4, the one-dimensional convolutional neural network includes: an input layer, a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a first dropout layer, a Flatten layer, a fully connected layer, a second dropout layer, and an output layer connected in sequence.

5. The method for identifying underground soil layers according to claim 4, characterized in that: The input layer is used to input the spectral reflectance curve; the first convolutional layer is used to extract features from the output of the input layer, capturing features such as absorption valleys within multiple spectral bands; the first max-pooling layer is used to downsample the output of the first convolutional layer, which reduces the feature dimension, retains important information, and improves the robustness of the mineral identification model; the second convolutional layer is used to extract deep spectral features from the output of the first max-pooling layer, enabling it to identify finer-grained or composite features; the second max-pooling layer is used to perform dimensionality reduction on the output of the second convolutional layer to reduce the computational load of the mineral identification model and improve its performance. The high mineral recognition model exhibits strong robustness. The first dropout layer prevents overfitting by randomly discarding some neurons during training. The Flatten layer flattens the multidimensional feature tensor output by the first dropout layer into a one-dimensional vector, which is then used as input to the fully connected layer. The first fully connected layer fuses various local features to construct a global representation. The second dropout layer further suppresses overfitting. The output layer is a fully connected layer with the number of nodes equal to the total number of mineral categories (n), and the activation function is Softmax. The output of the output layer is a spectral reflectance curve corresponding to the predicted probability of each mineral.

6. The method for identifying underground soil layers according to claim 4, characterized in that: The first convolutional layer has a kernel size of 9, a number of kernels of 64, and an activation function of ReLU; the first max pooling layer has a pooling window size of 2; the second convolutional layer has a kernel size of 5, a number of kernels of 128, and an activation function of ReLU; the first dropout layer has a dropout rate of 0.3; the first fully connected layer contains 128 neurons and has an activation function of ReLU; the second dropout layer has a dropout rate of 0.3; and the output layer has an activation function of Softmax.

7. The method for identifying underground soil layers according to claim 1, characterized in that, The method for determining the soil layer category of the collected soil in S7: If, among the predicted probabilities of n minerals corresponding to the spectral reflectance curve, the probability of kaolinite is >0.7 and the probability of illite is <0.2, then the soil type at the penetration depth corresponding to the spectral reflectance curve in the collected soil is: clay soil. If, among the predicted probabilities of n minerals corresponding to the spectral reflectance curve, the probability of illite mineral + the probability of mica mineral > 0.6 and the probability of montmorillonite mineral < 0.3, then the soil quality at the penetration depth corresponding to the spectral reflectance curve in the collected soil is: well-graded sand. If the probability of montmorillonite is greater than 0.5 among the predicted probabilities of n minerals corresponding to the spectral reflectance curve, then the soil type at the penetration depth corresponding to the spectral reflectance curve in the collected soil is silty soil.

Images