CN104239708A - Karst collapse column prediction method based on wavelet neural network - Google Patents

Abstract

The invention provides a karst collapse column prediction method based on a wavelet neural network, and belongs to the technical field of hydrogeololgy. The prediction method comprises the following steps that a wavelet transform method and a neural network model method are used for building a karst collapse column prediction model, further, the judging type of a karst collapse column is determined, and the karst collapse column prediction is realized. Further, the prediction method concretely comprises the following steps including S1, data collection and feature extraction; S2, neural network structure design; S3, neural network training; S4, prediction type judgment. The karst collapse column prediction method based on the wavelet neural network has the advantage that a wavelet analysis method and a BP neural network are sufficiently combined, so the position judging prediction of the karst collapse column is more scientific and more accurate.

Description

Based on the Genesis of Karst Subsided Column Forecasting Methodology of wavelet neural network

Technical field

The present invention relates to hydrogeological technical field, particularly relate to a kind of Genesis of Karst Subsided Column Forecasting Methodology based on wavelet neural network.

Background technology

At present, the position prediction of lot of domestic and international expert to karst collapse col umn rests on the level rule of thumb analogized, do not predict by other better forecast models, although Field Using Fuzzy Comprehensive Assessment makes some progress in prediction, but up to the present, even do not carry out the research of the position of Wavelet Neural Network Method prediction karst collapse col umn.

Summary of the invention

In order to solve the problems of the technologies described above, the object of this invention is to provide a kind of Genesis of Karst Subsided Column Forecasting Methodology based on wavelet neural network, this Forecasting Methodology has wavelet analysis method and the well-bound feature of BP neural network, and the differentiation of the position of Genesis of Karst Subsided Column predicts that science is with accurate more.

The object of the invention is to be achieved through the following technical solutions:

Based on a Genesis of Karst Subsided Column Forecasting Methodology for wavelet neural network, comprise the steps: to use small wave converting method and neural net model method construction Genesis of Karst Subsided Column forecast model, and then determine that Genesis of Karst Subsided Column differentiates type, realize Genesis of Karst Subsided Column prediction.

Further, the described Genesis of Karst Subsided Column Forecasting Methodology based on wavelet neural network specifically comprises the steps:

S1, data acquisition and feature extraction, according to overall target signal wave feature, collection can reflect the fluctuation of overall target characteristic, after its discretize, carries out the wavelet decomposition of N layer, with the energy of each floor height frequency coefficient of wavelet decomposition sequence for element forms proper vector, gather the overall target signal of the various running status of characterization system as much as possible, after feature extraction, proper vector is stored in storer;

S2, design neural network structure, according to input, the output layer nodes of the dimension of input feature value and the status number determination wavelet neural network of karst collapse col umn, if karst has N kind existence, so network output layer nodes just elects N as; The activation function of hidden layer and output layer selects Morlet wavelet function and sigmoid function respectively;

S3, neural network training, with signal energy proper vector for training sample input vector, training sample output vector defining method is: karst collapse col umn differentiates that state suppose there is 2 kinds of states: there is karst collapse col umn and there is not karst collapse col umn, i.e. state 1, state 2, network exports as { y ₁, y ₂, if differentiation state is 1, namely there is karst collapse col umn, then make y _j=y ₁, and all the other are zero, then the output vector of network is { 1,0}; If differentiation state is 2, namely there is not karst collapse col umn, then network output vector is { 0,1}; Utilize the training sample in storer to train network, make error sum of squares be less than error target, and the network weight trained and deviation are deposited in storer;

S4, differentiates type of prediction, utilizes the network weight and deviation that train, the feature parameter vectors input neural network to be discriminated, is exported by network and makes comparisons with training sample output vector, differentiate type of prediction.

By the technical scheme of the invention described above, the Genesis of Karst Subsided Column Forecasting Methodology based on wavelet neural network of the present invention has wavelet analysis method and the well-bound feature of BP neural network, and the differentiation of the position of Genesis of Karst Subsided Column predicts that science is with accurate more.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of Genesis of Karst Subsided Column Forecasting Methodology of the present invention;

Fig. 2 is the FB(flow block) of the defining method of overall target;

Fig. 3 is the FB(flow block) of neural network training;

Fig. 4 is the FB(flow block) differentiating prediction;

Fig. 5 is Genesis of Karst Subsided Column integrated forecasting index system schematic diagram.

Embodiment

Below in conjunction with accompanying drawing, the specific embodiment of the present invention is described in detail:

Based on a Genesis of Karst Subsided Column Forecasting Methodology for wavelet neural network, comprise the steps: to use small wave converting method and neural net model method construction Genesis of Karst Subsided Column forecast model, and then determine that Genesis of Karst Subsided Column differentiates type, realize Genesis of Karst Subsided Column prediction.

Further, as shown in Figure 1, Figure 2, Figure 3 and Figure 4, the described Genesis of Karst Subsided Column Forecasting Methodology based on wavelet neural network specifically comprises the steps:

S1, data acquisition and feature extraction.According to overall target signal wave feature, gather the fluctuation that can reflect overall target characteristic, after its discretize, carry out the wavelet decomposition of N layer, with the energy of each floor height frequency coefficient of wavelet decomposition sequence for element forms proper vector.Although neural network has stronger self study and associative ability, find that the associative ability of neural network is limited, overrun in practice, network adopts the mode of mistake to associate by not providing any prompting.Therefore, when this system of practical application, the overall target signal of the various running status of characterization system should be gathered as much as possible, after feature extraction, proper vector is stored in storer.

S2, design neural network structure.According to input, the output layer nodes of the dimension of input feature value and the status number determination wavelet neural network of karst collapse col umn.If karst has N kind existence, so network output layer nodes just elects N as; The activation function of hidden layer and output layer selects Morlet wavelet function and sigmoid function respectively.

S3, neural network training.In order to neural network can be identified Genesis of Karst Subsided Column existence, first must train network.With signal energy proper vector for training sample input vector, training sample output vector defining method is: karst collapse col umn differentiates that state suppose there is 2 kinds of states: there is karst collapse col umn and there is not karst collapse col umn, be i.e. state 1, state 2, and network exports as { y ₁, y ₂, if differentiation state is 1, namely there is karst collapse col umn, then make y _j=y ₁, and all the other are zero, then the output vector of network is { 1,0}; If differentiation state is 2, namely there is not karst collapse col umn, then network output vector is { 0,1}; Utilize a large amount of training samples in storer to train network, make error sum of squares be less than error target, and the network weight trained and deviation are deposited in storer.

S4, differentiates type of prediction, utilizes the network weight and deviation that train, the feature parameter vectors input neural network to be discriminated, is exported by network and makes comparisons with training sample output vector, differentiate type of prediction.

Wherein, as shown in Figure 2 and Figure 5, overall target in described step S1 comprises 10 two-level index, wherein, as shown in Figure 5, overall target comprises 2 first class index (i.e. geological structure factors, karsts developing area dynamic condition) and 10 two-level index (namely mining area structure form, fold are grown, to tiltedly growing scale, tensional fault growth, watery, the conditong of recharge and flowoff and draining, water temperature, dissolubility always fix TDS, calcium magnesium ratio and pH value) to be respectively its defining method as follows:

S11, determines 10 two-level index by following expression formula, that is:

$A_i=\frac{a_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i/n}---\left(1\right)$

$B_i=\frac{b_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{b}_i/n}---\left(2\right)$

$C_i=\frac{c_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{c}_i/n}---\left(3\right)$

$D_i=\frac{d_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{d}_i/n}---\left(4\right)$

$E_i=\frac{e_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{e}_i/n}---\left(5\right)$

$F_i=\frac{f_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{f}_i/n}---\left(6\right)$

$G_i=\frac{g_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{g}_i/n}---\left(7\right)$

$H_i=\frac{h_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{h}_i/n}$

$I_i=\frac{i_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{i}_i/n}$

$J_i=\frac{j_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{j}_i/n}---\left(10\right)$

Wherein A _i, B _i, C _i..., J _irepresent mining area structure morphological index sequence, fold development index sequence respectively, always fix TDS index series, calcium magnesium ratio index series and pH value index series, a to oblique growth scale merit sequence, tensional fault development index sequence, watery index series, the conditong of recharge and flowoff and draining index series, water temperature index series, dissolubility _i, b _i, c _i..., j _irepresent mining area structure morphological index, fold development index respectively, to tiltedly growing scale merit, tensional fault development index, watery index, the conditong of recharge and flowoff and draining index, water temperature index, dissolubility always fix TDS index, calcium magnesium ratio index and pH value index, the element number of n for comprising in the sequence selected;

S12, overall target is synthesized by above-mentioned 10 two-level index, and its expression formula is:

$Z_i=\frac{A_i+B_i+C_i+D_i+E_i+F_i+G_i+H_i+I_i+J_i}{10}---\left(10\right)$

Wherein, Z _irepresent i-th element in overall target sequence.

The present invention is the theoretical a kind of new karst collapse col umn forecast model constructed with the thought of artificial neural network of combined with wavelet transformed.Because wavelet transformation has good time frequency localization character, and neural network has self-learning function, and have stronger approximation capability and fault-tolerant ability, a kind of New Wavelet Neural Network after wavelet analysis method is fully combined with BP neural network, the advantage of the two is combined, karst collapse col umn integrated forecasting is had great importance.

Through studying the multianalysis of Carboniferous-permian, north China about coalfield Genesis of Karst Subsided Column formation condition and distribution rule, Carboniferous-permian, north China coalfield Genesis of Karst Subsided Column integrated forecasting index system (as shown in Figure 5) is proposed.This index system for Main Basis, in conjunction with the universal law of Genesis of Karst Subsided Column distribution, specifically comprises 2 first class index and 10 two-level index with Genesis of Karst Subsided Column formation condition.Because each two-level index is different on the order of magnitude and unit of quantity, therefore when by this 10 two-level index synthesis overall targets, takes standardized method, then synthesize overall target; Next differentiate that the overall target Z of prediction carries out analysis and prediction around Genesis of Karst Subsided Column.

The present invention is a kind of method carrying out differentiating prediction based on wavelet neural network to Genesis of Karst Subsided Column, and it specifically differentiates that process is: the acquisition being first discriminant information, in Genesis of Karst Subsided Column prediction, utilize overall target signal to obtain discriminant information; Then be feature extraction, namely adopt the extracting method of suitable science, from characteristic signal, extract the differentiation feature of Genesis of Karst Subsided Column; Be finally state recognition and differentiation prediction, namely adopt mode identification technology to classify, to determine to differentiate type, its essence is exactly the process of a Model Identification.

The above; be only the present invention's preferably embodiment, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; the change that can expect easily or replacement, all should be encompassed within protection scope of the present invention.Therefore, protection scope of the present invention should be as the criterion with the protection domain of claims.

Claims (3)

1., based on a Genesis of Karst Subsided Column Forecasting Methodology for wavelet neural network, it is characterized in that, comprise the steps:

Use small wave converting method and neural net model method construction Genesis of Karst Subsided Column forecast model, and then determine that Genesis of Karst Subsided Column differentiates type, realize Genesis of Karst Subsided Column prediction.

2. the Genesis of Karst Subsided Column Forecasting Methodology based on wavelet neural network according to claim 1, is characterized in that, comprise the steps:

S1, data acquisition and feature extraction, according to overall target signal wave feature, collection can reflect the fluctuation of overall target characteristic, after its discretize, carries out the wavelet decomposition of N layer, with the energy of each floor height frequency coefficient of wavelet decomposition sequence for element forms proper vector, gather the overall target signal of the various running status of characterization system as much as possible, after feature extraction, proper vector is stored in storer;

S2, design neural network structure, according to input, the output layer nodes of the dimension of input feature value and the status number determination wavelet neural network of karst collapse col umn, if karst has N kind existence, so network output layer nodes just elects N as; The activation function of hidden layer and output layer selects Morlet wavelet function and sigmoid function respectively;

S3, neural network training, with signal energy proper vector for training sample input vector, training sample output vector defining method is: karst collapse col umn differentiates that state suppose there is 2 kinds of states: there is karst collapse col umn and there is not karst collapse col umn, i.e. state 1, state 2, network exports as { y ₁, y ₂, if differentiation state is 1, namely there is karst collapse col umn, then make y _j=y ₁, and all the other are zero, then the output vector of network is { 1,0}; If differentiation state is 2, namely there is not karst collapse col umn, then network output vector is { 0,1}; Utilize the training sample in storer to train network, make error sum of squares be less than error target, and the network weight trained and deviation are deposited in storer;

S4, differentiates type of prediction, utilizes the network weight and deviation that train, the feature parameter vectors input neural network to be discriminated, is exported by network and makes comparisons with training sample output vector, differentiate type of prediction.

3. the Genesis of Karst Subsided Column Forecasting Methodology based on wavelet neural network according to claim 2, is characterized in that, the overall target in described step S1 comprises 10 two-level index, and its defining method is as follows:

S11, determines 10 two-level index by following expression formula, that is:

$A_i=\frac{a_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i/n}$

$B_i=\frac{b_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{b}_i/n}$

$C_i=\frac{c_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{c}_i/n}$

$D_i=\frac{d_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{d}_i/n}$

$B_i=\frac{b_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{b}_i/n}$

$F_i=\frac{f_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{f}_i/n}$

$G_i=\frac{g_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{g}_i/n}$

$H_i=\frac{h_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{h}_i/n}$

$I_i=\frac{i_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{i}_i/n}$

$J_i=\frac{j_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{j}_i/n}$

Wherein A _i, B _i, C _i..., J _irepresent mining area structure morphological index sequence, fold development index sequence respectively, always fix TDS index series, calcium magnesium ratio index series and pH value index series, a to oblique growth scale merit sequence, tensional fault development index sequence, watery index series, the conditong of recharge and flowoff and draining index series, water temperature index series, dissolubility _i, b _i, c _i..., j _irepresent mining area structure morphological index, fold development index respectively, to tiltedly growing scale merit, tensional fault development index, watery index, the conditong of recharge and flowoff and draining index, water temperature index, dissolubility always fix TDS index, calcium magnesium ratio index and pH value index, the element number of n for comprising in the sequence selected;

S12, overall target is synthesized by above-mentioned 10 two-level index, and its expression formula is:

$Z_i=\frac{A_i+B_i+C_i+D_i+E_i+F_i+G_i+H_i+I_i+J_i}{10}$

Wherein, Z _irepresent i-th element in overall target sequence.