Rockburst Risk Level Early Warning Method Based on Tag Consistency Features

Abstract

The present invention discloses a rockburst risk level early warning method based on label consistent features. The method of the present invention constructs a label consistent feature learning model, thereby being able to obtain the label consistent features of rockburst index data and effectively improve the accuracy of rockburst early warning. The specific implementation process is as follows: (1) With the help of consistent analysis theory and class label information, a novel label consistent feature learning model is constructed; (2) the analytical solution of the feature transformation matrix in the model is optimized; (3) based on the feature transformation matrix, the label consistent features of rockburst index data are directly obtained by matrix multiplication, and the rockburst risk level early warning result is obtained by using the nearest neighbor classifier. Compared with the prior art, the rockburst early warning method of the present invention is more effective and robust.

Description

Technical Field

[0001] The present invention relates to the technical field of rockburst early warning, and in particular to a rockburst risk level early warning method based on label consistent features, which can be applied to the fields of rockburst risk classification, rockburst dynamic early warning, etc. Background Art

[0002] In the field of rockburst prediction, due to the complexity of the original data, the research process will not only consume a lot of storage and computing costs, but also cause serious distortion of the warning results. Therefore, how to effectively warn of rockburst risks is still a topic worth studying. Feature learning is an effective method among all solutions. The existing consistency analysis method is a bimodal problem. By learning the direction of feature transformation, the consistency between the two groups of variables can be maximized. However, it not only does not consider the class label information between samples, resulting in a waste of resources, but also does not consider the consistency and geometric structure between modes. With the help of consistency analysis theory, the present invention takes the class label information as one of the modes in the bimodal consistency analysis theory. In the framework of feature learning, the class label information is directly used to constrain the learning of class label consistent features, so that the learned class label consistent features have better class separation in the rockburst risk level warning, improve the robustness of the algorithm, and form a label consistent feature learning model, so as to find more effective class labels and geometric structures between samples, obtain more comprehensive identification information of rockburst indicator data, and effectively improve the accuracy of rockburst warning. Summary of the invention

[0003] In order to better grasp the geometric structure and identification information between rockburst index data, the present invention constructs a label consistent feature learning model based on the rockburst index data, and theoretically derives the analytical solution of the model, thereby obtaining good early warning accuracy. The specific implementation steps of the present invention are as follows:

[0004] 1. Convert the training rockburst index data into column vectors to form a rockburst training sample matrix where d x represents the sample dimension of X, N represents the number of samples of X, and x i is the i-th (i=1,2,…,N) sample in X; use one-hot encoding to convert x i The class label is converted into a class label indicator vector y i ,y i The definition of y i =[y i1 ,y i2 ,…,y ic ] T ∈R c×1 , if and only if x i When y belongs to the vth class (v=1,2,…,c)iv = 1, otherwise y iv = 0, where c is the number of categories of the rockburst risk level; Using all the class label indicator vectors to form the class label indicator matrix Y = [y 1 , y 2 , …, y N ∈ R c×N ;

[0005] 2. Construct a label-consistent feature learning model.

[0006] The consistency analysis theory aims to evaluate the consistency between two sets of modalities. The method of the present invention uses a feature transformation matrix to transform the rockburst index samples and the corresponding class label indicator vectors into class label-consistent features, and then constructs a consistency coefficient between the class label-consistent features in the rockburst index samples and the corresponding class label indicator vectors, thereby forming a label-consistent feature learning model. This model solves for the feature transformation matrix that maximizes this consistency coefficient. The objective optimization function of this model is:

[0007]

[0008] where α T is the transpose vector of α, α is the feature transformation vector in A, A is the feature transformation matrix of X, β is the feature transformation vector in B, and B is the feature transformation matrix of Y; In this model, R is the covariance matrix of X and Y, and the specific calculation formula of R is:

[0009]

[0010] where is the sample mean of X, is the sample mean of Y; H is the variance of X, and the specific calculation formula of H is:

[0011]

[0012] F is the variance of Y, and the specific calculation formula of F is:

[0013]

[0014] 3. Optimize and solve the analytical solution of the feature transformation matrix A.

[0015] The implementation steps for optimizing and solving the feature transformation matrix A are as follows:

[0016] Since the objective optimization function of the label-consistent feature learning model is scale-invariant with respect to α and β, this objective optimization function can be transformed into:

[0017]

[0018] s.t. αT Hα = 1, β T Fβ = 1

[0019] Construct the Lagrangian multiplier function L(α, β) of α and β for the objective function:

[0020]

[0021] Among them, and are Lagrange multipliers;

[0022] Set the partial derivatives of L(α, β) with respect to α and β to 0, and we can get:

[0023] Rβ = λ 1 Hα

[0024] R T α = λ 2 Fβ

[0025] Multiply both sides on the left by α T and β T We can get: Then the above formula can be transformed into:

[0026]

[0027]

[0028] Assume that F is invertible and From We can get:

[0029]

[0030] Substitute into We can get the generalized eigenvalue decomposition problem:

[0031]

[0032] Let The above eigenvalue decomposition problem can be optimized to:

[0033] H -1 RF -1 R T α = λα

[0034] In this eigenvalue decomposition problem, λ is the eigenvalue, H -1 is the inverse matrix of H, F -1 is the inverse matrix of F, R T is the transpose matrix of R, and the eigenvector transformation matrix A is composed of H -1 RF -1 R TIt is composed of the eigenvectors corresponding to the first d largest eigenvalues, that is where d is the projection parameter of the projection direction, and α p is the eigenvector corresponding to the p-th largest eigenvalue, p = 1, 2, …, d x ;

[0035] 4. Using the feature transformation matrix, the label-consistent training feature z of the rockburst index data can be directly obtained through matrix multiplication γ and the label-consistent test feature of the rockburst index data z γ The calculation formula of is z γ = A T x γ , The calculation formula of is Finally, use the nearest neighbor classifier to classify the label-consistent training feature and the label-consistent test feature, and then obtain the risk level of the rockburst index test data.

[0036] The method of the present invention has the following advantages:

[0037] (1) The present invention uses the consistent analysis theory and class label information, takes the class label information as one modality in the two-modal consistent analysis theory, and directly uses the class label information in the feature learning framework to constrain the learning of class label-consistent features, so that the learned class label-consistent features have better class separability in the rockburst risk level warning, thus effectively improving the accuracy of the rockburst warning;

[0038] (2) Through theoretical derivation, the present invention obtains the analytical solution of the label-consistent feature learning model, and can directly obtain the label-consistent features of the rockburst index data through matrix multiplication, so as to realize the effective warning of the rockburst risk level. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is the flow chart of the present invention

[0040] Figure 2 is the warning accuracy rate of the rockburst risk level in ten random experiments DETAILED DESCRIPTION OF THE INVENTION

[0041] The specific implementation steps of the present invention are as follows:

[0042] 1. Convert the training rockburst index data into column vectors to form a rockburst training sample matrix where d x represents the sample dimension of X, N represents the number of samples of X, and x i is the i-th (i = 1, 2, …, N) sample in X; use one-hot encoding to encode x iThe class label is converted into a class label indicator vector y i ,y i The definition of y i =[y i1 ,y i2 ,…,y ic ] T ∈R c×1 , if and only if x i When y belongs to the vth class (v=1,2,…,c) iv =1, otherwise y iv =0, where c is the number of rockburst risk level categories; all class label indicator vectors are used to form the class label indicator matrix Y of the rockburst training sample = [y 1 ,y 2 ,…,y N ]∈R c×N ;

[0043] 2. Based on the label consistent feature learning model, H -1 RF -1 R T By performing eigenvalue decomposition with α=λα, the analytical solution of the eigenvalue transformation matrix A can be obtained.

[0044] 3. Using the feature transformation matrix, the label consistent training feature z of the rockburst index data can be directly obtained through matrix multiplication γ Test features consistent with rockburst indicator data labels z γ The calculation formula is z γ =A T x γ , The calculation formula is Finally, the nearest neighbor classifier is used to classify the label-consistent training features and label-consistent test features to obtain the risk level of the rockburst indicator test data.

[0045] The effect of the present invention is further verified by the following experiments:

[0046] The rockburst indicator dataset consists of 220 data samples. The sample labels of this dataset are divided into four categories: 1, 2, 3, and 4 correspond to the four levels of rockburst: none, weak, medium, and strong. In this experiment, 50 training data are randomly selected from each category of the 220 rockburst indicator data for training, and the rest are used as test data. This random selection is performed ten times. Figure 2 The rockburst risk warning accuracy of each random experiment is intuitively demonstrated, and the average accuracy of ten random experiments reaches 74%, which reveals that the method of the present invention is an effective rockburst risk level warning method.

Claims

1. A method for warning of rockburst risk level based on label-consistent features, characterized in that the method comprises the following steps: (1) Convert the training rockburst index data into column vectors to form a rockburst training sample matrix where d x represents the sample dimension of X, N represents the number of samples of X, and x i is the i-th (i = 1, 2, …, N) sample in X; use one-hot encoding to convert the class label of x i into a class label indicator vector y i , and the definition of y i is y i = [y i1 , y i2 , …, y ic T ∈R c×1 , and y i = 1 if and only if x iv belongs to the v-th (v = 1, 2, …, c) class, otherwise y iv = 0, where c is the number of categories of rockburst risk levels; use all the class label indicator vectors to form the class label indicator matrix Y = [y 1 , y 2 , …, y N ∈ R c×N ; (2) Construct a label-consistent feature learning model, and the specific implementation method is as follows: With the help of the correlation theory, and taking the class label information as one of the modalities, the constructed label-consistent feature learning model is as follows: where α T is the transposed vector of α, α is the eigen transformation vector in A, A is the eigen transformation matrix of X, β is the eigen transformation vector in B, and B is the eigen transformation matrix of Y; in this model, R is the covariance matrix of X and Y, and the specific calculation formula of R is: where is the sample mean of X, is the sample mean of Y; H is the variance of X, and the specific calculation formula of H is: F is the variance of Y, and the specific calculation formula of F is: (3) Optimize and solve the analytical solution of the feature transformation matrix A, and the specific implementation method is as follows: The optimization problem of constructing the label-consistent feature learning model is transformed into the following eigenvalue decomposition problem: H -1 RF -1 R T α = λ α In this eigenvalue decomposition problem, λ is the eigenvalue, and H -1 is the inverse matrix of H, F -1 is the inverse matrix of F, R T is the transpose matrix of R. The eigen-transformation matrix A is composed of the eigenvectors corresponding to the first d largest eigenvalues of H -1 RF -1 R T , that is where d is the projection parameter of the projection direction, and α p is the eigenvector corresponding to the p-th largest eigenvalue, p = 1, 2,..., d x ; (4) With the help of the feature transformation matrix, the label-consistent training features and label-consistent test features of the rockburst index data can be directly obtained through matrix multiplication, and the label-consistent test features are classified by means of a nearest neighbor classifier, so as to obtain the risk level of the rockburst index test data.

Images