CN117094222A - Neural network-based method for deducting physical and mechanical indexes from physical parameters of rock and soil body - Google Patents

Abstract

The invention discloses a neural network-based method for deducing physical and mechanical indexes of physical parameters of a rock-soil body, which comprises the following steps: s1, acquiring data; s2, preprocessing data; s3, constructing a neural network and training; s4, evaluating and adjusting parameters; and S5, model deployment. The invention provides accurate and reliable rock-soil body parameters for performance analysis, evaluation, prediction and the like of geotechnical engineering based on a rock-soil body characteristic prediction model established by geophysical exploration achievements, and further promotes the development of electromagnetic exploration technical means represented by ground penetrating radar, a semi-aviation transient electromagnetic system and a full-aviation transient electromagnetic system in the field of geotechnical engineering investigation. The invention solves the problems of incapability of long-term memory, gradient in back propagation and the like in the neural network, can better extract data characteristics, is used for predicting the parameters of the rock-soil body, and has higher prediction precision.

Description

Neural network-based method for deducting physical and mechanical indexes from physical parameters of rock and soil body

Technical Field

The invention belongs to the field of geology, relates to a geophysical exploration method, and in particular relates to a neural network method for deducting physical and mechanical indexes of a rock-soil body by using physical parameters of the rock-soil body.

Background

In recent years, geophysical exploration methods have made great progress in basic theoretical research, instrument and equipment, signal acquisition and processing and other aspects, and electromagnetic exploration technical means represented by ground penetrating radars, semi-aviation transient electromagnetic systems and full-aviation transient electromagnetic systems are increasingly used in the geotechnical engineering exploration field. The basic principle is that a high-frequency electromagnetic wave signal is transmitted into the ground through a transmitting antenna, the propagation rule is researched, and the time-frequency characteristic and the amplitude characteristic of the high-frequency electromagnetic wave signal are analyzed to know the structural characteristic and the electrical characteristic of a near-surface medium. The resistivity and the dielectric constant of the soil are basic parameters for representing the conductivity of the soil body, and are basic physical parameters of the soil. The resistivity is actually the resistance that is exhibited when the current passes vertically through a cubic soil body with a side length of 1m, and the unit is omega-m; and dielectric constant refers to the ability of a substance to hold a charge in F/m.

The resistivity and dielectric constant of the soil are closely related to a number of parameters of the rock and soil mass, such as water content, weight, friction angle, cohesion, shear strength and stiffness. The performance analysis, evaluation, prediction and other works of stability and the like of geotechnical engineering are not separated from the accurate and reliable geotechnical body parameter prediction model. However, the natural rock-soil body has rough surface, various morphology, various variability, various factors influencing the parameter characteristics, high nonlinear relation and difficulty in forming an accurate and effective theoretical model to predict the parameter characteristics of the rock-soil body. The neural network is used as a widely interconnected computing network, has strong nonlinear mapping capability, and can accurately mine the internal relation between complex targets. In addition, the PCA cyclic neural network is a neural network, can solve the problems of incapability of long-term memory, gradient in back propagation and the like in the neural network, and is easy to train. Based on the method, the physical parameters of the rock-soil body are measured based on the characteristics of the geophysical field of the rock-soil body, an artificial intelligence method of PCA (principal component analysis) cyclic neural network is introduced, the physical and mechanical indexes of the rock-soil body are researched, and the physical essence that the geophysical characteristics of the rock-soil body and the physical and mechanical indexes of the rock-soil body are connected with each other is constructed.

Disclosure of Invention

The invention aims to provide a neural network-based method for deducing physical and mechanical indexes of physical parameters of a rock and soil body, which is based on a geophysical detection result of the rock and soil body, carries out prediction research on the physical parameters of the rock and soil body including porosity, water content, gravity, friction angle, cohesive force and modulus and the mechanical indexes of the rock and soil body including compressive strength and shear strength by using the neural network, constructs the physical essence of the connection and interconnection of the geophysical characteristics of the rock and soil body and the mechanical characteristics of the rock and soil body, and solves the problem in the prior art that the physical characteristics of the rock and soil body are directly obtained by using the geophysical characteristic parameters.

The invention aims at realizing the following technical scheme:

a method for deducting physical and mechanical indexes based on physical parameters of a rock-soil body of a neural network comprises the following steps:

step S1, data acquisition:

s11, obtaining physical parameters of a rock-soil body by using a geophysical prospecting technology, wherein: physical parameters of the rock-soil body include resistivity, elastic wave velocity and dielectric constant;

step S12, measuring physical and mechanical indexes of the rock and soil sample through an indoor test, wherein: physical indexes include porosity, water content, weight, friction angle, cohesive force and elastic modulus, and mechanical indexes include compressive strength and shear strength;

step S2, data preprocessing:

deleting abnormal values and missing values from the sample data, dividing the data set into two parts of characteristic variables and target variables, and dividing the data set into a training data set and a test data set;

step S3, constructing a neural network and training:

s31, constructing a plurality of groups of PCA-RNN circulating neural networks which are respectively used for predicting physical indexes and mechanical indexes of different rock-soil bodies;

step S32, performing dimension reduction, compression and feature extraction on the network data: calculating the characteristic values corresponding to different principal components by using the covariance matrix, and then calculating the accumulated contribution rate corresponding to each principal component so as to determine the topological structure of the PCA neural network, namely the hidden layer number; performing feature reduction and noise reduction on the data through a Principal Component Analysis (PCA), and finding out a principal component feature subset containing the maximum information;

step S33, the data after the principal component analysis is further input into a cyclic neural network (RNN) for classification training;

step S4, evaluating and adjusting parameters:

s41, evaluating the accuracy and the stability of the PCA-RNN circulating neural network by a ten-fold cross validation method to obtain a neural network topological structure with good accuracy and stability;

step S42, further adjusting super parameters, and determining the learning rate, the activation function and the optimization algorithm of the model;

step S5, model deployment:

inputting physical property parameter data of the rock and soil body without the indoor test measurement section, and deducting the physical and mechanical indexes of the rock and soil body by using the PCA-RNN circulating neural network trained in the step S4.

Compared with the prior art, the invention has the following advantages:

1. the invention provides accurate and reliable rock-soil body parameters for performance analysis, evaluation, prediction and the like of geotechnical engineering based on a rock-soil body physical mechanical index prediction model established by geophysical exploration results, and further promotes the development of electromagnetic exploration technical means represented by ground penetrating radar, a semi-aviation transient electromagnetic system and a full-aviation transient electromagnetic system in the field of geotechnical engineering investigation.

2. The invention solves the problems of incapability of long-term memory, gradient in back propagation and the like in the neural network, can better extract data characteristics, is used for predicting the parameters of the rock-soil body, and has higher prediction precision.

Drawings

FIG. 1 is a geophysical model of a rock and soil mass;

FIG. 2 is a graph showing the prediction of the modulus of a rock-soil body in an example;

FIG. 3 is a diagram of training sample and test sample errors in an embodiment;

FIG. 4 is a diagram of test set correlation coefficients in an embodiment.

Detailed Description

The following description of the present invention is provided with reference to the accompanying drawings, but is not limited to the following description, and any modifications or equivalent substitutions of the present invention should be included in the scope of the present invention without departing from the spirit and scope of the present invention.

The invention provides a neural network-based method for deducing physical and mechanical indexes of physical parameters of a rock-soil body, which comprises the following steps:

step S1, data acquisition: in the process of damage and movement in the rock-soil body, the physical parameters of the rock-soil body mainly comprise parameters such as resistivity, elastic wave velocity, dielectric constant and the like. The resistivity is related to mineral components, underground water and the like in a rock-soil body, and is closely related to granularity, porosity, water content, rock mineralization degree and the like; in geotechnical engineering application, for a rock mass, the consolidation degree of a bedrock can be judged and engineering grade division can be carried out by testing the longitudinal and transverse wave speed ratio of the rock mass, and a certain correlation exists between the wave speed value and the cohesive force and internal friction angle of the rock mass; the characteristics of the ground penetrating radar are mainly that the weak structural surface of the rock and soil mass and the condition of groundwater are probed. Therefore, the step mainly obtains the physical parameters of the rock-soil mass with resistivity, elastic wave velocity and dielectric constant of 3 types. Physical indexes of the rock and soil sample, including porosity, water content, weight, friction angle, cohesive force and elastic modulus, are measured through an indoor test, and mechanical indexes include compressive strength and shear strength.

Step S2, data preprocessing: and selecting a plurality of groups of parameters representing the geophysical characteristics of the rock and soil body, and the parameters of the physical indexes and the mechanical indexes of the rock and soil body measured indoors as data sets. The selected geophysical characteristic parameters are preferably the dielectric constant, resistivity and elastic wave velocity of the rock-soil mass; the rock-soil body mechanical indexes measured indoors comprise compressive strength and shear strength, and the physical indexes comprise porosity, water content, weight, friction angle, cohesive force and elastic modulus. Deleting abnormal values and missing values in the sample, dividing the data set into a training data set and a test data set, dividing the data set into two parts of a characteristic variable and a target variable, and dividing the data set into the training data set and the test data set, wherein: the characteristics comprise attributes and input variables, and the target variables comprise categories and output variables; the training data set is used for establishing a model, the test data set is used for testing the model effect, and the training set and the test set use the same segmentation proportion for quantitatively evaluating the influence of the original input parameters and the reconstruction input parameters on the model precision. The data set is discretized by adopting a z-score standardization method to improve convergence speed and generalization performance. The data is input into the PCA-RNN cyclic neural network, and the PCA-RNN cyclic neural network model is trained.

Step S3, constructing a neural network and training:

s31, constructing eight groups of PCA-RNN circulating neural networks, which are respectively used for predicting the 6 physical indexes and the 2 mechanical indexes;

step S32, performing dimension reduction, compression and feature extraction on the network data: calculating the characteristic values corresponding to different principal components by using the covariance matrix, and then calculating the accumulated contribution rate corresponding to each principal component so as to determine the topological structure of the PCA neural network, namely the hidden layer number; performing feature reduction and noise reduction on the data through a Principal Component Analysis (PCA), and finding out a principal component feature subset containing the maximum information;

in step S33, the data after the principal component analysis is further input into a Recurrent Neural Network (RNN) for classification training, where the number of rounds is determined by the length of the input sequence, that is, specifically, if the length of the input sequence is T, the RNN will be cycled T times between time steps t=1 to t=t.

Step S4, evaluating and adjusting parameters:

and S41, evaluating the accuracy and the stability of the PCA-RNN circulating neural network through a ten-fold cross validation algorithm, dividing a data set into 10 parts, alternately taking 9 parts of the data set as training data, taking 1 part of the data set as test data for test, taking an average value of the accuracy (or error rate) obtained by 10 tests as an estimate of the algorithm accuracy, and acquiring a neural network topological structure with good accuracy and stability, namely the hidden layer number.

And S42, further adjusting the super parameters, and determining the learning rate, the activation function and the optimization algorithm of the model. Firstly, setting a smaller learning rate, then increasing the learning rate and calculating a corresponding loss value every time the network is updated, and determining the optimal learning rate by drawing a change curve of the learning rate and the loss value, wherein the functions can be realized by the existing pyrach code; the activation function selects a leak ReLU function, the function is simple and effective, the calculation speed is high, and the problem of gradient disappearance does not exist; the optimization algorithm may be determined by a gradient descent method. And retraining and evaluating by using the test set data in the hyper-parameter adjustment process until the result is satisfied.

Step S5, model deployment:

and (3) inputting geophysical characteristic data of the section which is not subjected to the indoor test, and deducing the characteristics of the rock-soil body by using the PCA-RNN cyclic neural network which is trained in the step (S4).

In the step, physical parameters of the rock and soil body obtained through means such as ground penetrating radar, electromagnetic exploration and the like are used as input data to be imported into the eight groups of PCA-RNN circulating network models after parameter adjustment in the step S4, and predicted physical indexes and mechanical indexes of the rock and soil body can be output.

In this step, R is used for the neural network trained in step S4 ² And evaluating the prediction precision of the model. R is R ² The closer to 1, the higher the model accuracy, R ² The definition is as follows:

wherein y is _i ' is the predicted outcome; y is _i The test result is an indoor test result;the average value of the indoor test results is given, and n is the capacity of the test data set.

Examples:

according to the embodiment, a physical and mechanical index prediction model of a rock and soil body is built by combining an indoor test, so that the physical and mechanical properties of the rock and soil body which are not measured by the indoor test are predicted. The model designed in this embodiment obtains the topology structure by ten-fold cross-validation algorithm, and the data is derived from southwest China, and is shown in fig. 1.

The embodiment establishes a rock-soil body characteristic prediction model (figure 1) in combination with the indoor test to predict the rock-soil body characteristics which are not measured by the indoor test. The model designed by the embodiment obtains the topological structure by a ten-fold cross-validation algorithm, and the data is derived from southwest China.

In this embodiment, when the PCA-RNN cyclic neural network structure has two hidden layers, the hidden neurons are respectively 12 and 6, and the optimization is achieved, namely an ANN:4-12-6-1, the test set MSE is 0.631, the correlation coefficient is 0.978, and the test set correlation coefficient is shown in FIG. 4. Fig. 2 shows the prediction results of the model, and fig. 3 shows the prediction error map, from which the prediction trend of the modulus of the rock-soil mass is known to be highly reliable.

Claims (6)

1. The method for deducing the physical and mechanical index of the physical parameters of the rock and soil body based on the neural network is characterized by comprising the following steps:

step S1, data acquisition:

s11, obtaining physical parameters of a rock-soil body;

step S12, determining physical indexes and mechanical indexes of the rock and soil sample through an indoor test;

step S2, data preprocessing:

deleting abnormal values and missing values from the sample data, dividing the data set into two parts of characteristic variables and target variables, and dividing the data set into a training data set and a test data set;

step S3, constructing a neural network and training:

s31, constructing a plurality of groups of PCA-RNN circulating neural networks which are respectively used for predicting physical indexes and mechanical indexes of different rock-soil bodies;

step S32, performing dimension reduction, compression and feature extraction on the network data: calculating the characteristic values corresponding to different principal components by using the covariance matrix, and then calculating the accumulated contribution rate corresponding to each principal component so as to determine the topological structure of the PCA neural network, namely the hidden layer number; performing feature reduction and noise reduction on the data by a principal component analysis method, and finding out a principal component feature subset containing the maximum information;

step S33, the data after the principal component analysis is further input into a cyclic neural network for classification training;

step S4, evaluating and adjusting parameters:

s41, evaluating the accuracy and the stability of the PCA-RNN circulating neural network by a ten-fold cross validation method to obtain a neural network topological structure with good accuracy and stability;

step S42, further adjusting super parameters, and determining the learning rate, the activation function and the optimization algorithm of the model;

step S5, model deployment:

inputting physical property parameter data of the rock and soil body of a section which is not subjected to indoor test measurement, and deducting the physical and mechanical properties of the rock and soil body by using the PCA-RNN circulating neural network which is trained in the step S4.

2. The neural network-based rock-soil body physical property parameter deduction physical mechanical index method according to claim 1, wherein the physical property parameters comprise resistivity, elastic wave velocity and dielectric constant.

3. The method for deducing physical and mechanical indexes based on physical parameters of a rock and soil body of a neural network according to claim 1, wherein the physical indexes comprise porosity, water content, weight, friction angle, cohesive force and elastic modulus, and the mechanical indexes comprise compressive strength and shear strength.

4. The neural network-based method for deriving physical and mechanical indicators based on physical parameters of a rock and soil body according to claim 1, wherein said activation function selects a leak ReLU function.

5. The neural network-based method for deducing physical and mechanical indexes from physical parameters of a rock and soil body, according to claim 1, wherein the optimization algorithm is determined by a gradient descent method.

6. The method for deriving physical and mechanical indexes based on physical parameters of a rock and soil body of a neural network according to claim 1, wherein in said step S5, R is used for the PCA-RNN cyclic neural network trained in step S4 ² Evaluating model prediction accuracy, R ² The definition is as follows:

wherein y is _i ' is the predicted outcome; y is _i The test result is an indoor test result;the average value of the indoor test results is given, and n is the capacity of the test data set.