1. A lithology identification method based on an improved radial basis function neural network is characterized by comprising the following steps:

preprocessing data obtained by geophysical joint inversion by combining samples collected in a mining area;

performing feature extraction on the preprocessed data by adopting K-L transformation to realize dimension reduction processing to obtain a compressed data set;

processing a new data set by adopting a K-fold cross verification method, dividing K into equal parts after the data set is disturbed, wherein one part is used as a test set, and the rest K-1 parts are used as training sets;

completing clustering of the training set by adopting a fuzzy C clustering algorithm to obtain the center of the hidden layer;

building a radial basis function neural network, and solving parameters of the radial basis function neural network according to the center of the hidden layer;

verifying the radial basis function neural network by using a test set, and recording the identification accuracy of various types;

and repeating the model training and testing for K times, solving the overall recognition accuracy rate, and storing the radial basis function neural network with the optimal parameters.

2. The improved lithology identification method based on the radial basis function neural network as claimed in claim 1, wherein the parameters of the radial basis function neural network comprise: center, width, weight.

3. The improved radial basis function neural network-based lithology identification method of claim 1, wherein preprocessing comprises: checking the consistency of the data, processing invalid values and missing values, and centralizing and normalizing the data.

4. The lithology recognition method based on the improved radial basis function neural network as claimed in claim 1, wherein the step of performing feature extraction on the preprocessed data by using K-L transformation to realize dimension reduction processing to obtain a compressed data set comprises:

inputting a characteristic value;

calculating an average value as a new coordinate axis origin;

solving a covariance matrix;

obtaining an eigenvalue and an eigenvector of the covariance matrix;

and sorting the eigenvalues from big to small, and taking the first m corresponding eigenvectors as new data.

5. The lithology recognition method based on the improved radial basis function neural network of claim 1, wherein the clustering of the training set is completed by adopting a fuzzy C clustering algorithm, and the center of the hidden layer is obtained, comprising:

giving the number C of clustering centers to be divided and related parameters;

initializing a membership matrix U;

c clustering centers are calculated;

calculating a distance matrix from each sample point to a clustering center to obtain a new membership matrix;

calculating an objective function value J;

judging whether the difference is smaller than a given threshold or smaller than a threshold with the target function generated in the last cycle;

if not, the above steps are repeatedly executed, and if the above steps are already less, the iteration is ended.

6. The improved radial basis function neural network-based lithology recognition system of claim 1, wherein the radial basis function neural network learning method comprises: random selection, self-organizing learning, and gradient descent.

7. A lithology recognition system based on an improved radial basis function neural network, comprising:

one or more processors;

storage means for storing one or more programs;

when executed by the one or more processors, cause the one or more processors to implement the improved radial basis function neural network-based lithology identification method of any one of claims 1-6.

Background

Lithology refers to some property that reflects the characteristics of rock, including color, composition, structure, cement and type of cement, specific minerals, etc. Lithology recognition is a fundamental task in the process of stratum understanding and reservoir parameter solution. With the continuous exploitation of mineral resources, surface resources are increasingly exhausted, and exploration and excavation of deep resources become important in current geological work. The lithology and the structural relation of the underground reservoir are finely described, and important basic information can be provided for deep resource exploration and reservoir structure recognition. Therefore, lithology identification for deep formations has become a focus of current geological work.

The lithology recognition research comprises methods such as gravity magnetism, well logging, earthquake, remote sensing, electromagnetism, geochemistry and slice analysis. The remote sensing identification is only limited to the earth surface and cannot meet the requirement on depth; seismic identification has great advantages in depth, but the high operation cost hinders the large-scale development of the seismic identification; the logging identification technology is developed most mature, and the disadvantage is that extensive lithology identification cannot be carried out. The lithology is identified by carrying out geophysical joint inversion, underground physical structure characteristics such as density, magnetism and electrical property of a research area can be obtained through geophysical data inversion, and lithology distribution in the stratum is judged according to the logical relationship between the physical structure characteristics and the lithology. The large-scale gravity and magnetic data in China are quite complete, the coverage area is wide, the sampling density is high, the inversion algorithm is developed well, and the large-area stratum lithology identification result can be conveniently obtained. Therefore, the gravity magnetic based lithology identification is considered to be the most likely successful and suitable method for popularization of the three-dimensional lithology identification at the present stage. The current heavy magnetic lithology recognition work is less developed, mainly the relation between lithology and physical properties is analyzed, the logical topology operation is carried out on the density and magnetic susceptibility data obtained by inversion with prior information constraint, and a lithology graph is drawn (strictly, perpetuation and the like, 2014; diminution and the like, 2017).

However, under the condition of no prior information constraint, the lithology identification based on the gravity and the magnetism has the problems of poor vertical resolution, strong multi-solution and the like. How to convert the physical structure obtained by the heavy magnetic inversion into lithology is essentially a pattern recognition problem. Pattern recognition is a method for realizing feature extraction and class classification by using a calculation method. From the learning of the sample space features, the belonging patterns can be mapped to the appropriate feature class space. In recent years, with the rapid development of artificial intelligence and machine learning, pattern recognition based on an artificial neural network has been widely studied and successfully applied to the field of lithology recognition. The Zhangye et al (2018) establishes a migration model of rock image set analysis based on the inclusion-v 3 deep convolutional neural network, has good performance under the condition of sufficient data, and has test probability values all reaching more than 85%; pantolo and the like (2020) optimize a BP neural network model based on a principal component analysis method, and effectively solve the problem of lithologic identification of well logging in a research area; shaohong et al (2008) transform seismic attribute data using probabilistic neural network methods, and have significant advantages in formations with strong heterogeneity. The artificial neural network has strong adaptability and learning ability, and the prediction result of the neural network is superior to that of the traditional statistical method, so that the establishment of the gravity magnetic lithology recognition based on the artificial neural network has higher research value.

A Feedforward Neural Network (FNN) is an artificial neural network model widely used. The BP (Back propagation) neural network is applied more in lithology identification, but the BP neural network essentially adopts a gradient descent method, has slow convergence rate and low training efficiency, is greatly influenced by an initial value, and has low fault tolerance rate. Therefore, the BP neural network has many difficulties in practical application. A Radial Basis Function (RBF) neural network is a feedforward neural network based on the local response of human brain neurons to the outside world, and has the advantages of concise model training, fast convergence and high calculation speed. Particularly, the method has strong nonlinear mapping capability, can carry out global approximation on a nonlinear function with any precision, and has preliminary application in the field of lithology identification. According to the old tide and the like (2008), a lithology recognition model based on a radial basis function neural network is established by combining actual logging data and lithology profile data of a certain well in the pseudo-songorian basin, and the convergence speed is high and the recognition accuracy is high; junk mail and others (2013) realize lithology recognition application based on genetic optimization radial basis function neural network, and improve the interpretation efficiency and precision of well logging data.

The self-adaption and self-learning capabilities of the radial basis function neural network are realized through a learning algorithm. The learning process involves determining the center vector of the hidden layer, the width coefficient and the weight calculation from the hidden layer to the output layer. The common radial basis function neural network learning methods mainly include three types: selecting RBF centers at random (direct calculation method), selecting RBF centers by self-organizing learning (K-means clustering), and selecting RBF centers by supervised learning (gradient descent). The random selection method is more suitable when the sample data with representative distribution is processed, and is not suitable for general sample data; the self-organizing learning selection method generally uses a traditional K-means clustering algorithm, the clustering result is easily influenced by the initial random selection of the clustering center, and the stability is poor; when supervised learning selection is adopted, the objective function to be optimized by the gradient descent method is very complex, so that the network convergence speed is slow. And comprehensively comparing the efficiency and the precision of the three RBF center selection methods, and selecting a self-organizing learning method based on K-means clustering for lithology identification of deep strata.

Physical property sample characteristic spaces obtained by different types of rocks through geophysical joint inversion are likely to have more overlap, physical property parameters of various types of rocks have no obvious boundary, the relationship between lithology and physical properties is often fuzzy, objects in a data set cannot be divided into obviously separated clusters, a sample is divided into designated clusters by adopting a K-means clustering algorithm, the assigned clusters are likely to be harder and even make mistakes, and the lithology with small sample capacity is likely to be wrongly divided; meanwhile, a large amount of redundant information exists among the physical property (speed, density, magnetic susceptibility, resistivity and the like) values obtained by inversion, so that the physical property values have high correlation with each other, and the precision and the efficiency are influenced. Therefore, for physical property data obtained by geophysical joint inversion, comprehensive, accurate and efficient lithology identification is difficult to perform only by adopting a radial basis function neural network based on a K-means clustering algorithm.

Disclosure of Invention

The invention aims to provide a lithology identification method and a lithology identification system based on an improved radial basis function neural network, which can carry out comprehensive, accurate and efficient lithology identification.

In order to solve the technical problem, the invention provides a lithology identification method based on an improved radial basis function neural network, which comprises the following steps: preprocessing data obtained by geophysical joint inversion by combining samples collected in a mining area; performing feature extraction on the preprocessed data by adopting K-L transformation to realize dimension reduction processing to obtain a compressed data set; processing a new data set by adopting a K-fold cross verification method, dividing K into equal parts after the data set is disturbed, wherein one part is used as a test set, and the rest K-1 parts are used as training sets; completing clustering of the training set by adopting a fuzzy C clustering algorithm to obtain the center of the hidden layer; building a radial basis function neural network, and solving parameters of the radial basis function neural network according to the center of the hidden layer; verifying the radial basis function neural network by using a test set, and recording the identification accuracy of various types; and repeating the model training and testing for K times, solving the overall recognition accuracy rate, and storing the radial basis function neural network with the optimal parameters.

In some embodiments, the parameters of the radial basis function neural network include: center, width, weight.

In some embodiments, the pre-treating comprises: checking the consistency of the data, processing invalid values and missing values, and centralizing and normalizing the data.

In some embodiments, performing feature extraction on the preprocessed data by using K-L transformation to implement dimension reduction processing, and obtaining a compressed data set, includes: inputting a characteristic value; calculating an average value as a new coordinate axis origin; solving a covariance matrix; sorting the eigenvalues from big to small, and taking the first m corresponding eigenvectors to perform K-L conversion; and (5) calculating an eigenvalue and an eigenvector of the covariance matrix.

In some embodiments, the clustering of the training set is completed by using a fuzzy C-clustering algorithm, and the obtaining of the center of the hidden layer includes: giving the number C of clustering centers to be divided and related parameters; initializing a membership matrix U; c clustering centers are calculated; calculating a distance matrix from each sample point to a clustering center to obtain a new membership matrix; calculating an objective function value J; judging whether the difference is smaller than a given threshold or smaller than a threshold with the target function generated in the last cycle; if not, the above steps are repeatedly executed, and if the above steps are already less, the iteration is ended.

In some embodiments, a radial basis function neural network learning method includes: random selection, self-organizing learning, and gradient descent.

In addition, the invention also provides a lithology identification system based on the improved radial basis function neural network, which comprises the following components: one or more processors; storage means for storing one or more programs; when executed by the one or more processors, cause the one or more processors to implement a method for improved radial basis function neural network based lithology identification in accordance with the foregoing.

After adopting such design, the invention has at least the following advantages:

the lithology identification method is based on physical property data obtained by geophysical joint inversion, a radial basis function neural network (FCM-RBFNN) lithology identification model based on K-L transformation and fuzzy clustering optimization is constructed, learning of neural network parameters is improved by means of dimension reduction processing of sample data, determination of hidden layer centers by means of a fuzzy clustering algorithm and the like, lithology identification efficiency and accuracy are effectively improved, and the overall average accuracy rate obtained by means of a K-fold cross verification method is up to 94.5% and is higher than 83.2% of the average accuracy rate of the RBFNN model. The model can effectively complete the lithology recognition task in geological interpretation.

Drawings

The foregoing is only an overview of the technical solutions of the present invention, and in order to make the technical solutions of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and the detailed description.

FIG. 1 is a view of a RBFNN model architecture;

FIG. 2 is a flow chart of K-L transformation;

FIG. 3 is a flow chart of a fuzzy C clustering algorithm;

FIG. 4 is a schematic flow chart of a K-fold cross-validation method;

FIG. 5 is a diagram of an FCM-RBFNN lithology identification model;

FIG. 6 shows the result of K-fold cross-validation of the RBFNN model;

FIG. 7 shows the results of cross-validation of the FCM-RBFNN model by K-turn;

FIG. 8 is a graph comparing FCM-RBFNN and RBFNN.

Detailed Description

The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation.

A radial basis function neural network (FCM-RBFNN) lithology identification model based on K-L transformation and fuzzy clustering optimization is provided. The K-L transformation belongs to mathematical transformation of data statistical characteristics, can eliminate correlation among data and plays a role in data compression. The K-L transformation can be adopted to perform dimension reduction processing on the feature space, so that not only can the time and space complexity of the model be reduced, but also the lithology identification result can be more accurate. The Fuzzy C-Means clustering algorithm (FCM) is an uncertain description of sample classes, and can obtain the uncertainty degree of the samples belonging to each lithology and express the intermediacy of the sample classes. The result of obtaining the RBFNN base center by using the fuzzy C mean value clustering analysis method is more accurate.

1. Radial Basis Function Neural Network (RBFNN)

1.1 radial basis function

The radial basis function value is a real-valued function that depends only on the distance from the origin, i.e., a function that satisfies the property of Φ (x) ═ Φ (| x |). The radial basis function is mainly applied to the fields of scattered data interpolation and approximation, numerical solution of partial differential equations, construction of neural networks and the like.

1.2 radial basis function neural network architecture

The radial basis function neural network is a three-layer static forward network and is composed of an input layer, a hidden layer and an output layer, and the topological structure of the radial basis function neural network is shown in figure 1.

Wherein, the input layer is an m-dimensional vector I ═ (I)₁，I₂，...，I_m) And consists of signal source nodes.

The hidden layer is an m-dimensional vector D ═ D₁，D₂，...，D_P) And the hidden layer node generates nonlinear change according to the basis function, maps the input space to a new space, and obtains the output layer node after the new space is subjected to linear weighting. In radial basis function neural networks, the most commonly used basis functions are gaussian functions, which are defined as:

wherein x is an input vector;is the output of the ith unit of the hidden layer; c. C_iIs the central point of the ith unit Gaussian function of the hidden layer; sigma_iIs the width of the ith hidden layer node; p is the number of hidden layer nodes. The setting of hidden layer neuron parameters (center, width) is a major factor that affects classifier performance.

The third layer is an output layer and represents the classification mode corresponding to the input layer, and the node number is the same as the sample category number.

1.3 parameter learning method of radial basis function neural network

Suppose that a radial basis function neural network with m nodes in an input layer, P nodes in a hidden layer and n nodes in an output layer is established. Wherein, the connection weight between the input layer and the hidden layer is 1; p Gaussian functions exist in the network, namely P centers (marked as c) and P widths (marked as b) need to be determined; the connection between each node of the hidden layer and the output layer is nonlinear, and a weight matrix w is obtained through calculation. Therefore, we need to learn a proper set of c, b, w to determine the parameters of the whole model.

The common radial basis function neural network learning methods mainly include three types: random selection, self-organizing learning, and gradient descent.

(1) Random selection method. And for the parameters c and b, random initialization can be carried out, and then the pseudo-inverse matrix is directly calculated and solved according to the output vector of the hidden layer, so that the parameter w can be obtained. This method is only suitable for processing sample data with representative distribution, and has a large limitation.

(2) Self-organizing learning method. And dividing the sample data into P classes by using a clustering algorithm to obtain P clustering centers, determining a parameter c, calculating according to the clustering centers to obtain P variances, determining a parameter b, and finally solving a pseudo-inverse matrix according to the output vector of the hidden layer to obtain a parameter w. The method has a greater improvement in accuracy than the direct calculation method.

(3) Gradient descent method. The gradient descent method belongs to supervised learning, firstly, parameters c, b and w are obtained through random initialization, and then, the minimum value of a loss function is solved, and the updated weight value is gradually solved in an iterative mode. The objective function to be optimized by the method is very complex, and the network convergence speed is slow.

K-L transformation

The K-L (Karhunen-Loeve) transformation is a mathematical transformation established on the basis of statistical characteristics, can be regarded as a process of feature selection and feature extraction, and extracts main features in a large input space. The mathematical principle is to transform an original matrix X by an orthogonal matrix Q formed by normalized orthogonal feature vectors Q of an autocorrelation matrix R of the matrix X, that is, Y is QX, and the method includes:

m_Y＝E(Y)＝E(Q^TX)＝Q^TE(X)＝Q^Tm_X

∑_Y＝E(YY^T)-m_Ym_Y ^T＝E[(Q^TX)(Q^TX)T]＝E[Q^T(XX^T)Q]-Q^Tm_Xm_X ^TQ＝Q^T[E(XX^T)-m_Xm_X ^T]Q＝Q^T∑_YQ＝Λ＝diag[λ₁，λ₂，...，λ_N，]

wherein m is_YIs the average value of Y, ∑_YIs a covariance matrix, and the matrix form is:

K-L transformation completes covariance matrix sigma_YDiagonalization is carried out, so that the components of Y are not related to each other, the correlation among data can be eliminated, and the purpose of data compression, namely dimension reduction processing, is achieved. The K-L has the outstanding advantage of good decorrelation, is the optimal transformation in the Mean Square Error (MSE) meaning, and plays an important role in the data compression technology.

The steps of the K-L transformation are shown in fig. 2.

3. Fuzzy clustering

Fuzzy C-Means clustering algorithm (FCM) is the most widely used Fuzzy clustering method. The FCM algorithm divides clusters according to different membership degrees of different sample points to a cluster center, and the membership degree value of the FCM algorithm is expanded to [0, 1] from {0, 1} of a K-means clustering algorithm, namely the category membership degree of each sample is a real number interval.

Note X_i(i-1, 2, …, n) indicates that each vector has i-dimensional attributes, and the division into c cluster centers is called cluster V according to the selected similarity metric function_kWherein K ═ 1, 2, …, c.n samples belong to c classes of membership matrix, and are denoted as U ═ U [, U [ ], respectively_ik]_c×n(fuzzy partition matrix) in which U_ik(1. ltoreq. i.ltoreq.n, 1. ltoreq. k.ltoreq.c) represents the ith sample X_iThe membership degree belonging to the kth category satisfies the following constraint conditions:

U_ik∈[0，1]， 1≤i≤n，1≤k≤c

the objective function of the FCM algorithm is defined as follows:

the iterative formula for the cluster center is as follows:

the steps of the FCM clustering algorithm are shown in fig. 3.

K-fold cross validation

Equally dividing the K-Fold Cross Validation method (K-Fold Cross Validation) into K parts, taking 1 group as a test set and the other K-L groups as training sets, and training and testing the model. The experiment was repeated K times, and the average was taken as the model error evaluation result of the population, as shown in fig. 4.

The K-fold cross validation can improve the generalization capability of the model, effectively avoids under-fitting, over-fitting and other phenomena, and is widely applied to tasks such as model performance evaluation, model parameter determination, over-fitting inspection and the like.

5. Radial basis function neural network lithology recognition model based on k-L transformation and fuzzy clustering optimization

According to the algorithm, an RBFNN model (FCM-RBFNN) based on k-L transformation and fuzzy clustering optimization is provided, and the implementation process is shown in FIG. 5.

1) Combining samples collected in a mining area, preprocessing data obtained by geophysical joint inversion, wherein the preprocessing comprises the steps of checking the consistency of the data, processing invalid values and missing values, centralizing and normalizing the data and the like;

2) performing feature extraction on the preprocessed data by adopting K-L transformation to realize dimension reduction processing to obtain a compressed data set;

3) processing a new data set by adopting a K-fold cross verification method, dividing K into equal parts after the data set is disturbed, wherein one part is used as a test set, and the rest K-1 parts are used as training sets;

4) completing clustering of the training set by adopting a fuzzy C clustering algorithm to obtain the center of the hidden layer;

5) building a radial basis function neural network, and solving parameters c, b and w according to the center of the hidden layer;

6) verifying the model by using a test set, and recording the identification accuracy of various types;

7) and repeating the model training and testing for K times, solving the overall recognition accuracy rate, and storing the optimal network parameters.

In order to verify the optimization effect of the model, the model and the test result of the RBFNN center model obtained by the self-adaptive center selection method based on the K-means clustering algorithm are compared and analyzed, as shown in FIGS. 6 and 7 and Table 1.

TABLE 1 summary of accuracy of various lithology predictions for RBFNN and FCM-RBFNN models

The test results of the two models are compared to obtain the FCM-RBFNN lithology identification model, and the identification accuracy is improved to a certain extent compared with the RBFNN algorithm based on K-means selection of the base center.

(1) For the identification of granite, amphibole, pyroxene and marble, the accuracy of the two models is high, and the accuracy of the FCM-RBFNN model is slightly improved compared with that of the RBFNN model.

(2) For the identification of sedimentary rocks and phyllites, the accuracy of the RBFNN model is low, namely 54.7% and 77.0%, the sedimentary rocks are easy to predict to be the phyllites, and the phyllites are easy to predict to be the pyroxene, and the reason is that the densities and the magnetic susceptibilities of the sedimentary rocks and the marbles are close (the average sedimentary rock density is 2.719g/cm3, the average marbles density is 2.726g/cm3, the average sedimentary rock magnetic susceptibilities are 0.686 multiplied by 10 < -5 > SI, and the average marbles magnetic susceptibilities are 0.285 multiplied by 10 < -5 > SI). The accuracy rate predicted on the FCM-RBFNN model is obviously improved, and the accuracy rates are 84.0% and 100% respectively. It is effective to apply the method to actual lithology identification engineering.

Referring to fig. 8, the method is based on physical property data obtained by geophysical joint inversion, a radial basis function neural network (FCM-RBFNN) lithology identification model based on K-L transformation and fuzzy clustering optimization is constructed, neural network parameter learning is improved by means of dimensionality reduction processing on sample data, a fuzzy clustering algorithm is adopted to determine a hidden layer center and the like, lithology identification efficiency and precision are effectively improved, and a K-fold cross verification method is adopted to obtain the overall average accuracy rate which is 94.5% and is 83.2% higher than that of the RBFNN model. The model can effectively complete the lithology recognition task in geological interpretation.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the present invention in any way, and it will be apparent to those skilled in the art that the above description of the present invention can be applied to various modifications, equivalent variations or modifications without departing from the spirit and scope of the present invention.