Soil compact state grade classification method based on deep learning network

Abstract

The invention discloses a soil compact state grade classification method based on a deep learning network. The method comprises the following steps: obtaining a soil compact state grade classification condition; a detector is arranged in the target area, a frequency dispersion curve is picked up based on a surface wave exploration method, the picked frequency dispersion curve is input into the prediction model of soil body dense state grade classification, and the soil body dense state grade classification condition of the target road section is obtained; the method for establishing the prediction model for grade classification of the compact state of the soil body comprises the following steps of: acquiring a training sample, wherein the training data is used for constructing a frequency dispersion curve, soil texture and a compact state of the soil body corresponding to a stratum structure; and inputting the training sample into the deep learning network model, and training to obtain a soil compact state grade classification model of the target area. According to the method provided by the invention, the underground structure does not need to be inversed, the efficiency is high, and errors caused by individual differences can be effectively avoided, so that rapid grade division of the compact state of the soil body can be accurately and reliably carried out on the target section.

Description

Technical Field

[0001] The invention relates to the technical field of geophysical detection, and in particular to a method for determining soil compaction state classification. Background Art

[0002] Existing surface wave exploration technology is widely used in the fields of geophysical exploration and engineering investigation. By analyzing the propagation characteristics of surface waves, the shear wave velocity structure of the underground medium can be obtained, thereby inferring the physical and mechanical properties of the soil.

[0003] As a cutting-edge technology in the field of artificial intelligence, deep learning network technology has powerful data processing and pattern recognition capabilities. Deep learning networks are applied to the analysis of geophysical exploration data, especially for the evaluation of soil compaction status.

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for determining the classification of soil compaction state. The method combines surface wave exploration technology and a deep learning network model, and uses the surface wave dispersion characteristics and stratum velocity and density information to construct an assessment model for the soil compaction state. Through this method, the soil compaction state can be efficiently and accurately classified, providing a scientific basis for engineering design and geological disaster prevention. Summary of the invention

[0005] The present invention overcomes the deficiency of traditional methods that cannot accurately evaluate the compaction state of soil under complex geological conditions by integrating the efficient data acquisition capability of surface wave exploration technology and the powerful learning capability of deep learning network. The method realizes the classification of the compaction state of soil.

[0006] A method for classifying soil compaction levels is as follows:

[0007] Step 1: Through preliminary investigation of the soil types in the target area and its surroundings and the density standards of different soils, Soild = [Soild1, Soild2, …, SoildN], the corresponding soil density state standard at the corresponding Vs speed is [Com1, Com2, …, Com3].

[0008] Furthermore, the construction standard in step 1 is as follows: Soild1: The soil compactness state (Compactness) is divided into slightly loose (Com1) and the speed Vs range is [v 11 ,v 12 ]; Medium density (Com2) speed Vs range is [v 13 ,v 14 ]; the speed Vs of compaction (Com3) ranges from [v 15 ,v 16 ]; Soil N:The soil compaction state is divided into slightly loose (Com1) and the velocity Vs range is [v N1 ,v N2 ]; Medium density (Com2) speed Vs range is [v N3 ,v N4 ]; the speed Vs of compaction (Com3) ranges from [v N5 ,v N6 ];

[0009] Step 2: According to the underground geological structure of the exploration area, determine the physical parameters of each stratum and obtain the detection depth H of the target area; divide into layers by D (m), N = H / D, and obtain N layers; for the velocity and soil quality of each layer, N underground structure models can be randomly generated by the Monte Carlo method; including the longitudinal wave velocity Vp, shear wave velocity Vs, density ρ and thickness h, and establish a layered model.

[0010] Step 3: Obtain the theoretical dispersion curve f n , where in each layer, it is assumed that the surface wave propagates horizontally, and its displacement is in the form of a plane wave, satisfying the elastic dynamics equation: Among them, u represents the particle displacement vector, ρ is the density, and λ and μ are the Lame constants.

[0011] Furthermore, in step 3, for Rayleigh waves, the following dispersion equation is derived by combining the boundary condition (free surface stress is zero) and the interlayer continuity condition (displacement and stress are continuous): D(ω,k)=0 Where ω is the angular frequency, k is the wave number, c=ω / k, and c is the phase velocity.

[0012] Furthermore, in step 3, for the Love wave, when only the horizontal displacement component is considered, the dispersion equation can be expressed as: in:

[0013] Furthermore, in step 3, the dispersion equation is numerically solved to obtain f n ,; (1) Select the appropriate frequency range ω (2) According to the formation model parameters, the initial phase velocity c is set and the corresponding wave number k = ω / c is calculated. (3) Use the numerical iteration method to solve the dispersion equation D(ω,k)=0 to obtain the phase velocity c at a specific frequency. (4) Repeat (1), (2), and (3) to obtain the corresponding relationship between frequency ω and phase velocity c, that is, the corresponding theoretical dispersion curve fn =[(ω1,c1),(ω2,c2),…,(ωn,cn)], where the phase velocity c is Vs.

[0014] Step 4: Repeat step 3 according to the different geological model structures constructed to obtain N theoretical dispersion curves F = [f1, f2, f3, ..., f n ].

[0015] Step 5: Determine Soild = [Soild1, Soild2, ..., SoildN] and Vs = [[v 11 ,v 12 ],[v 21 ,v ＝22 ],…,[v N1 ,v N2 ]], the corresponding soil compaction state Com=[Com1, Com2,…, Com3] is determined by Soild and Vs;

[0016] Step 6: Construct the training set [F, Soild], the corresponding label is [Com].

[0017] Step 7: Establish a soil compaction level prediction model, which includes a shared feature layer and a soil embedding layer. The corresponding K layer establishes K branches; each branch contains a classifier; and the classifier is composed of a fully connected layer.

[0018] Furthermore, in step 7, the shared feature layer is constructed as follows: each layer includes a one-dimensional convolution layer; the convolution layer is followed by a batch normalization layer; the batch normalization layer is fed into an activation function; the output after the activation function is fed into a regularization layer; the regularization layer is fed into a pooling layer;

[0019] Furthermore, in step seven, the embedding layer and the fully connected classification layer in the model are constructed as follows: each layer of soil enters the embedding layer; the shared feature layer and the embedding layer data of the corresponding layer are spliced; the spliced ​​data is sent to the fully connected layer of the corresponding branch for output to obtain the corresponding soil density state classification;

[0020] Step 8: Send F to the feature extraction layer accordingly; Soild of different thicknesses are sent to the embedding layer respectively; after the shared feature extraction layer and the embedding layer are spliced, branches of different soil depths will be obtained; its branch is a fully connected layer, and the corresponding output is the soil compaction state [Com] of the corresponding soil layer;

[0021] Step 9: Build the above model structure and use the training set model to train to obtain a soil density level prediction model that meets the requirements.

[0022] Furthermore, in the step nine, the judging criteria are determined by the loss function Loss and the accuracy Accuracy;

[0023] Furthermore, in step nine, the training process includes a training set and a validation set; the validation set is a part split from the training set; the loss function sizes of the training set and the validation set are Loss1 and Loss2; the accuracy rates are Acc1 and Acc2;

[0024] Furthermore, during the training process of step nine, the Loss function and Acc need to reach certain thresholds ∈1, ∈2, depending on the training set; at the same time, the values ​​of (Loss1-Loss2) and (Acc1-Acc2) also need to reach certain thresholds ∈3, ∈4, depending on the training situation;

[0025] Step 10: Arrange detectors in the target area, analyze and process the detector data to obtain the dispersion energy map E(ω,c);

[0026] Furthermore, in step 10, the processing and analysis of the detector data includes the following calculation method: A surface wave seismic record u(x, t), that is, the collected detector data; its corresponding Fourier transform can be expressed as u(x, ω), which consists of the amplitude term A(x, ω) and the phase term P(x, ω), that is: u(x,ω)=A(x,ω)P(x,ω)=A(x,ω)exp(-iωt) Among them, ω is the circular frequency, i is the imaginary unit, the amplitude term A(x,ω) mainly reflects the influence of spherical diffusion on the amplitude and the attenuation characteristics of the medium, and the phase term P(x,ω) preserves the travel time information or phase information of the wave. The negative sign in the phase term indicates that the source is excited on the left side.

[0027] Furthermore, the processing and analysis of the detector data in step 10 includes the following calculation method: under the far-field approximation condition, the wave diffusion mode can be regarded as plane diffusion, that is: P(x,w)=exp(-iwt)=exp(-iwx / c w ) Where x is the distance from the source to the receiver, c ( is the phase velocity corresponding to the frequency. The surface wave dispersion energy E(ω,c) can be extracted by scanning the spectrum u(x,ω) of the seismic record with another e index. For a surface wave seismic record with N channels, the extraction formula can be expressed as:

[0028] Step 11: Obtain the dispersion curve of the target area. The dispersion curve is obtained as follows: Pick the maximum value of the velocity at the corresponding frequency in the dispersion energy diagram E(ω,c) to obtain the dispersion curve f actual ;

[0029] Step 12: Send the dispersion curve of the target area to the trained soil compaction state level prediction model to obtain the corresponding prediction of the soil compaction state level of the target area. predict =[Com p1 ,Com p2 ,…,Com pn ]. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 This is the soil compaction level prediction model diagram; Figure 2 is the shared feature layer graph; Figure 3 Graphs for the embedding layer and the fully connected classification layer; Figure 4 The loss function and accuracy graph of the training set and validation set; Figure 5 The soil compaction level distribution map of the target area and its actual underground structure; Figure 6 It is the soil compaction standard for different soil types; Figure 7 This is a flow chart of the soil compaction classification method; Figure 8 Flowchart for establishing a soil compaction classification prediction model DETAILED DESCRIPTION The present invention is further described in detail below in conjunction with the specific method implementation process. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention. For a road section with a depth of 30m and a length of 30m, the implementation method is as follows: 1) According to the national and local standards, the soil types and density standards of different soils in the target area and its surrounding areas are established; Soild1: The soil state is divided into slightly loose velocity range [v 11 ,v 12 ]; Medium compaction speed range [v 13 ,v 14 ]; the speed range of density is [v 15 ,v 16 ]; SoildN: The soil state is divided into slightly loose velocity range [v N1 ,v N2 ]; Medium compaction speed range [v N3 ,vN4 ]; the speed range of density is [v N5 ,v N6 ], see Figure 6 ; 2) Construct a training set for the soil compaction state prediction model, 100,000 theoretical dispersion curves F = [f1, f2, f3, …, f n ]. 3) According to the generated underground structure model, determine Soild = [Soild1, Soild2, ..., SoildN] and Vs = [[v 11 ,v 12 ],[v 21 ,v ＝22 ],…,[v N1 ,v N2 ]], the corresponding soil compaction state Com=[Com1, Com2,…, Com3] is determined by the corresponding standard; 4) Construct a training set (F, Soild), the corresponding label is (Com); 5) Construct a soil compaction level prediction model, see Figure 1 , Figure 2 , Figure 3 ; Description of the model as a whole and its shared feature layer, embedding layer and fully connected classification layer respectively. 6) Train the model to obtain a soil compaction state prediction model that meets the requirements; the loss function size of the training set and the validation set is Loss1, Loss2; the accuracy is Acc1, Acc2; the respective Loss function and Acc need to reach a certain threshold ∈1, ∈2, depending on the training set; at the same time, the values ​​of (Loss1-Loss2) and (Acc1-Acc2) also need to reach a certain threshold ∈3, ∈4, depending on the training situation, see for details Figure 4 . 7) The surface wave exploration method is used to collect data in the target area and analyze it to obtain the dispersion curve; for the surface wave seismic record with N channels, the extraction formula can be expressed as: For the above dispersion energy diagram, the maximum value of the velocity at the corresponding frequency is picked up to obtain the dispersion curve. 8) Using the dispersion curve obtained in the target area, the trained soil compaction level prediction model is input to obtain the soil compaction level classification of the target area; In summary, the present invention achieves prediction and classification of the density at different depths underground in the target area by establishing a soil density level prediction model through a neural network. See Figure 5 ; The above description is only a preferred example for implementing the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for classifying soil compaction levels, used for classifying soil compaction levels, characterized in that: A soil compaction level prediction model is established through a neural network to predict and classify the compactness at different depths underground in the target area, including the following steps: Step 1: Through preliminary investigation of the soil types and density standards of different soils in the target area and its surroundings, Soild = [Soild1, Soild2, …, SoildN], the corresponding soil density state standard at the corresponding Vs speed is [Com1, Com2, …, Com3]; Step 2: Obtain the detection depth H of the target area; divide the layers into layers according to D (m), N = H / D, and obtain N layers; for the velocity and soil quality of each layer, N underground structure models can be randomly generated by the Monte Carlo method; Step 3: Based on the generated underground structure model, obtain the theoretical dispersion curve f of the underground structure model n ; Step 4: According to the generated underground structure model, repeatedly calculate step 3 to obtain N theoretical dispersion curves F = [f1, f2, f3, ..., f n ]; Step 5: Determine Soild = [Soild1, Soild2, ..., SoildN] and Vs = [[v 11 ,v 12 ],[v 21 ,v ＝22 ],…,[v N1 ,v N2 ]], the corresponding soil compaction state Com=[Com1, Com2,…, Com3] is determined by Soild and Vs; Step 6: Construct the training set (F, Soild), the corresponding label is (Com); Step 7: Establish a soil compaction level prediction model, which includes a shared feature layer and a soil embedding layer. The corresponding N layers establish N branches; each branch contains a classifier; and the classifier is composed of a fully connected layer; Step 8: Send F to the feature extraction layer accordingly; Soild of different thicknesses are sent to the embedding layer respectively; after splicing the shared feature extraction layer and embedding layer, branches with different soil depths will be obtained; Its branches are fully connected layers, and the corresponding outputs are the soil compaction states Com corresponding to the soil layers; Step 9: Use the training set to train the constructed model to obtain a soil compaction level prediction model that meets the requirements; Step 10: Arrange detectors in the target area, analyze and process the detector data to obtain the dispersion energy map E(ω, c); Step 11: Pick up the maximum value of the velocity at the corresponding frequency in the dispersion energy diagram E(ω, c) to obtain the dispersion curve factual; Step 12: Send the dispersion curve of the target area to the trained model to obtain the prediction of the underground soil compaction state classification of the target area Compredict = [Comp1, Comp2, ..., Compn].

2. The method according to claim 1, characterized in that In step 1, the density standards for different soil types are constructed as follows: Soild1: The soil compactness state is divided into slightly loose (Com1) and slightly loose (Com2). The velocity Vs range is [v 11 ,v 12 ]; Medium density (Com2) speed Vs range is [v 13 ,v 14 ]; the speed Vs of compaction (Com3) ranges from [v 15 ,v 16 ]; Soil N :The soil compaction state is divided into slightly loose (Com1) and the velocity Vs range is [v N1 ,v N2 ]; Medium density (Com2) speed Vs range is [v N3 ,v N4 ]; the speed Vs of compaction (Com3) ranges from [v N5 ,v N6 ].

3. The method according to claim 1, characterized in that In step 3, for a theoretical dispersion curve f n The method is calculated as follows: According to the underground geological structure of the exploration area, determine the physical parameters of each layer, including the longitudinal wave velocity Vp, the shear wave velocity Vs, the density ρ and the thickness h, and establish a layered model. Where Vp = 1.732*Vs; In each layer, it is assumed that the surface wave propagates horizontally, and its displacement is in the form of a plane wave, satisfying the elastic dynamics equation: Where u represents the particle displacement vector, ρ is the density, λ and μ are the Lame constants; For Rayleigh waves, the following dispersion equation is derived by combining the boundary conditions (free surface stress is zero) and the interlayer continuity conditions (displacement and stress are continuous): D(ω, k)=0 Where ω is the angular frequency, k is the wave number, c = ω / k, and c is the phase velocity; For Love waves, when only the horizontal displacement component is considered, the dispersion equation can be expressed as: in: (1) Select a suitable frequency range ω; (2) According to the formation model parameters, the initial phase velocity c is set and the corresponding wave number k = ω / c is calculated; (3) Use numerical iteration method to solve the dispersion equation D(ω,k)=0 to obtain the phase velocity c at a specific frequency; (4) Repeat (1), (2), and (3) to obtain the corresponding relationship between frequency ω and phase velocity c, that is, the corresponding theoretical dispersion curve f n＝ [(ω1,c1),(ω2,c2),…,(ω n ,c n )].

4. The method according to claim 1, characterized in that In step 5, according to the constructed different layer Vs speed ranges [V1, V2, ..., V n ], the soil type corresponding to [Soild1, Sold2,…, Soild N ]; By querying the standard of step 1 based on the soil type and the velocity range of Vs, the compaction state of soil in different layers [Com1, Com2,…, Com3] can be obtained.

5. The method according to claim 1, characterized in that In the step seven, the shared feature layer is constructed as follows: each layer includes a one-dimensional convolution layer; the convolution layer is followed by a batch normalization layer; the batch normalization is input into an activation function; the output after the activation function is input into a regularization layer; and the regularization layer is input into a pooling layer.

6. The method according to claim 1, characterized in that In the step seven, the embedding layer and the fully connected classification layer in the model are constructed as follows: each layer of soil enters the embedding layer; the shared feature layer and the embedding layer data of the corresponding layer are spliced; the spliced ​​data is sent to the fully connected layer of the corresponding branch for output to obtain the corresponding soil density state classification.

7. The method according to claim 1, characterized in that In the step nine, the evaluation criteria for meeting the requirements are composed of the loss function Loss and the accuracy Accuracy, wherein the training process includes a training set and a validation set; the loss function sizes of the training set and the validation set are Loss1 and Loss2; the accuracy is Acc1 and Acc2, wherein the respective Loss functions and Acc need to reach certain thresholds ∈1 and ∈2, depending on the specific situation of the training set; at the same time, the values ​​of (Loss1-Loss2) and (Acc1-Acc2) also need to reach certain thresholds ∈3 and ∈4, depending on the specific situation of the training.

8. The method according to claim 1, characterized in that In the step 10, the processing and analysis of the detector data includes the following calculation method: A surface wave seismic record u(x, t), that is, the collected detector data; its corresponding Fourier transform can be expressed as u(x, ω), which consists of the amplitude term A(x, ω) and the phase term P(x, ω), that is: u(x,ω)=A(x,ω)P(x,ω)=A(x,ω)exp(-iωt) Among them, ω is the circular frequency, i is the imaginary unit, the amplitude term A(x, ω) mainly reflects the influence of spherical diffusion on the amplitude and the attenuation characteristics of the medium, and the phase term P(x, ω) preserves the travel time information or phase information of the wave. The negative sign in the phase term indicates that the source is excited on the left side; Under the far-field approximation, the wave diffusion pattern can be regarded as plane diffusion, that is: P(x,w)=exp(-iwt)=exp(-iwx / c w ) Where x is the distance from the source to the receiver, c w is the phase velocity corresponding to the frequency; The surface wave dispersion energy E(ω, c) can be extracted by scanning the spectrum u(x, ω) of the seismic record through another e-index. The extraction formula can be expressed as: Where N is the number of surface wave seismic records.

Images