A rock burst time series prediction model construction method based on small sample learning

Abstract

A kind of rock burst time series prediction model construction method based on small sample learning, the original microseismic data collected by microseismic sensor is transmitted to the host computer of ground control room, the original microseismic data is standardized by the host computer, then the microseismic data after standardization is processed using principal component analysis method and kernel density estimation method, corresponding two-dimensional time series data is obtained, the correlation between the impending large energy mine seismic event and the past mine seismic event is studied, the density index is constructed, and the data labeling is completed, finally the long short-term neural network based on small sample learning is constructed, the time series data and density index are put into the constructed neural network for training, and the prediction of the time period of future large energy event is completed.The application can reduce the dependence of long short-term memory recurrent neural network on data, improve the generalization of network, so that it can also predict large energy mine earthquake when the amount of data is small, and provide early warning for rock burst.

Description

TECHNICAL FIELD

[0001] The application relates to a rock burst prediction model, in particular to a rock burst time series prediction model construction method based on small sample learning, and belongs to the technical field of rock burst prediction. BACKGROUND

[0002] Due to the actual national conditions of China, coal has become the basic energy and important raw material in China, and the coal industry has become an important basic industry related to the economic lifeline and energy security of the country. In recent years, with the rapid increase of the mining depth of coal mines in China and the increasingly complex geological and mining conditions, rock burst disasters have become increasingly severe and have become a key factor restricting the safety production of coal mines. Rock burst has the characteristics of suddenness, instantaneous vibration and great destructiveness, so it is difficult to predict the time, place and degree of its occurrence in advance. Rock burst has a great degree of harm and a wide range of influence, and once it occurs, it is easy to cause other chain accidents such as abnormal gas emission, coal and gas outburst, water inrush and gas explosion, causing great property losses and casualties. However, the complexity of the time, place, region and source of rock burst and the suddenness of rock burst make its prediction extremely difficult and complex, becoming a world problem that needs to be solved urgently.

[0003] Many scholars have made great progress in parameterizing mine earthquake data to predict mine earthquake disasters. However, in the past research process, it is often given for the mining geological conditions, which leads to the inability to provide a general applicability rule for rock burst. The occurrence of rock burst events is often accompanied by the generation of large energy microseismic events, but due to the complexity of rock burst and the limitation of human factors, it is difficult to obtain the data required for establishing a rock burst prediction model, which leads to the inability to master some mine microseismic data conditions and conduct scientific rock burst management. The problem of difficult rock burst data acquisition, missing sample data, makes it difficult to train machine learning models, making the rock burst research in data-deficient areas more challenging.

[0004] In recent years, with the development of computer technology, the application of rock burst prediction in deep learning has been greatly developed. Among them, the deep recurrent neural network (RNN) is particularly widely used in time series prediction, which can predict the current time with the help of previous information. However, it forgets some relatively distant information for a long time, and is prone to gradient disappearance and gradient explosion problems. Followed by the deformation of the deep recurrent neural network (RNN), the long short-term memory recurrent neural network (LSTM), which can selectively remember important information and forget unimportant information through the gate control unit, has achieved good results in time series prediction.

[0005] There are two kinds of existing rock burst prediction methods, one is a traditional rock burst prediction model established based on microseismic data, and the other is a rock burst prediction model established based on deep learning method. Although these models have achieved a series of successes, there are still the following disadvantages: 1) due to the influence of multiple factors on the occurrence of rock burst, the rock burst time series presents a highly complex nonlinearity, which limits the application of the traditional rock burst prediction model; at the same time, due to the different conditions of each mine and the great influence of human factors, although some indicators can be used to judge, a unified conclusion cannot be obtained. 2) the traditional deep learning model represented by long short-term neural network needs a large amount of data training to obtain a high enough prediction accuracy. For some mines with insufficient microseismic data, a large amount of microseismic data cannot be provided, and a good effect cannot be obtained. SUMMARY

[0006] The purpose of the present application is to provide a rock burst time series prediction model construction method based on small sample learning, which can reduce the dependence of long short-term memory recurrent neural network (LSTM) on data, improve the generalization of the network, and enable it to predict large energy mine earthquakes when the data volume is small, thereby providing early warning for rock burst.

[0007] In order to achieve the above purpose, the present application provides a rock burst time series prediction model construction method based on small sample learning, comprising the following steps:

[0008] Step 1, mine earthquake data processing: install microseismic sensors around the coal mining face, and transmit the collected original microseismic data to the upper computer in the ground control room by the microseismic sensors. The original microseismic data includes microseismic data occurrence time, energy size and seismic source coordinates. The original microseismic data is processed by the upper computer, first the original microseismic data is processed according to time period to obtain the microseismic data in each time period, then the microseismic data is standardized to obtain the time series data of each time period;

[0009] Step 2, density index establishment: the density index establishment includes principal component analysis PCA processing and kernel density estimation KDE processing, the time series data of step 1 is processed by principal component analysis PCA, and the original four-dimensional time series data is reduced to two-dimensional time series data; the two-dimensional time series data is processed by kernel density estimation KDE processing, and the probability density of each microseismic data is obtained after the two-dimensional time series data is processed, then the maximum probability density in each time period is selected as the density index for microseismic data time series analysis, and the general rule of large energy occurrence time is obtained, that is, the maximum probability density of more than 70% in this time period is selected, when the peak value appears, it is speculated that the future time period is possible to have a large energy event, so as to complete the construction of microseismic data label, the data label construction is as follows: 1 for large energy event, and 0 for other time period data label construction, that is, no large energy event occurs;

[0010] Step 3, long short-term memory recurrent neural network LSTM model based on small sample learning is established: long short-term memory recurrent neural network LSTM is used to model and extract features of the density index, and regression loss and data features are obtained; then, a small sample learning prototype network is used, the data features processed by long short-term memory recurrent neural network LSTM are clustered, and the center points of different class sample features in each time period are calculated, and the distance between the hidden feature vector of each sample and the corresponding hidden feature center point is used as the network classification loss measured by the Euclidean distance; finally, in the prediction module, the above loss is connected, gradient back propagation and optimization are carried out, the deep learning model based on small sample learning is constructed, and the prediction of large energy mine earthquake time period is finally output.

[0011] The mine earthquake data processing in step 1 of the application is as follows:

[0012] The fixed time window statistical data is used to construct a time series data set, the time series data includes time and coordinate positions X, Y and Z, then the time series data in each time period is standardized, the data is processed into standard normal distribution data with mean value 0 and standard deviation 1, the purpose is to scale the data, remove the unit limit of the data, and convert it into dimensionless pure value, so that different units or orders of magnitude indicators can be compared and weighted.

[0013] The specific method of standardization processing is as follows:

[0014]

[0015] The average value calculation method is as follows:

[0016]

[0017] The standard deviation calculation method is as follows:

[0018]

[0019] In the formula: x i Is the i-th original data, including the time and coordinate position of the i-th microseismic data;

[0020] Is the average value of the original data;

[0021] Sigma (x) is the standard deviation of the original data;

[0022] X i Is the updated data of the original data;

[0023] N is the total amount of data contained in the time period.

[0024] The density index in step 2 of the application is established as:

[0025] The principal component analysis method PCA is based on data dimension reduction, and the main idea of the method is to recombine M indexes with certain correlation to form a new set of N indexes that are not related to each other, and the N indexes obtained after recombination are linear combinations of the original M indexes and meet M>N, and the new N indexes are used instead of the original M indexes as the input of the model for subsequent analysis, and the specific steps are as follows:

[0026] Suppose there are p pieces of q-dimensional standardized data, first, the standardized data is grouped by column to form a q-row p-column matrix D, then each row of the matrix D is zero-mean, that is, the mean value of the row is subtracted, then the covariance matrix is solved According to the solved covariance matrix, the eigenvalue and the corresponding eigenvector are solved, the eigenvectors are arranged in a matrix according to the size of the corresponding eigenvalue from large to small, and the first k rows are taken to form a matrix P, and finally E=PD is the data after dimension reduction to k;

[0027] The data processed by the principal component analysis method PCA is subjected to kernel density estimation KDE, and the kernel density estimation function is as follows:

[0028]

[0029] In the formula: K(.) is a kernel function (non-negative, integral is 1, meets the probability density property, and the mean value is 0).

[0030] There are many kinds of kernel functions, for example, the Gaussian function is a commonly used kernel function;

[0031] H>0 is a smoothing parameter, called bandwidth, also known as window;

[0032] To scale the kernel function;

[0033] In , the h denominator inside the K function is used to adjust the width of the kernel density estimation KDE curve, and the h denominator outside the K function is used to ensure that the area under the curve meets the rules of kernel density estimation KDE;

[0034] After selecting a suitable kernel function and bandwidth, the kernel density estimation KDE can simulate the real probability distribution curve and obtain a smooth and beautiful result;

[0035] The present application adopts Gaussian function in the selection of kernel function, and the selection of wide band h according to Silverman rule is as follows:

[0036]

[0037] In the formula: n is the total amount of data contained in the time period;

[0038] D is the data dimension;

[0039] After the above kernel density estimation KDE processing, the probability density of each data point can be obtained, and by griding the principal component analysis PCA coordinate axis, the probability density of each time period can be plotted on the principal component analysis PCA coordinate axis.

[0040] The long short-term memory recurrent neural network LSTM model based on small sample learning in step 3 of the present application is established, and the specific steps are as follows:

[0041] First, the prediction model is generated based on the density index created by the training data set after step 2 processing, and the time series data set is constructed, and the present application takes 6h as the time window to count the data, and assumes that the data record calculated by the ith time window is m i , which can be expressed as:

[0042] , which can be expressed as:

[0043]

[0044] In the formula: id is the time window number;

[0045] is the average value of the occurrence place in the microseismic data time period;

[0046] t is the later time in the microseismic data time period;

[0047] For example, when the time period is 0 to 6, t is selected as 6 o'clock;

[0048] After the m i The data obtained after the standardization processing according to the method of step 1 is added to the corresponding p of each time periodmax Density Index, configured to

[0049] For the problem of data missing in a time period, a nearest neighbor interpolation technique can be used to supplement the missing data, that is, when the data in the current time period is missing, the data in the previous time period is supplemented to the data in the current time period;

[0050] Based on the above method of traversing the time window, the time data sequence data set M can be obtained:

[0051] M = [m0, m1, m2,..., m n-1 ] (7)

[0052] The label set T corresponding to the time period is represented as:

[0053] T = [t0, t1, t2,..., t n-1 ] (8)

[0054] In the formula: t i is 1, otherwise t i = 0; i

[0055] The deep learning model based on small sample learning is as follows:

[0056] The deep learning model based on small sample learning mainly includes three modules of convolution layer, long short-term memory recurrent neural network LSTM and small sample prototype network, the data set is selected as 85% for training set, 5% for verification set and 10% for test set;

[0057] The convolution layer is as follows:

[0058] The data first flows into the convolution module, which is used to analyze the short-range dependence relationship between 20 (single sequence length) microseismic events and the relationship between the 5 attributes of the events, the convolution kernel size is determined as 6x5, the number of convolution kernels is 32, and the convolution layer activation function is selected as ReLU function, whose expression is as follows:

[0059] ReLU(x) = max(0, x) (9)

[0060] The convolution kernel operation process is as follows:

[0061] h k = ReLU(W k *X + b k ) (10)

[0062] In the formula: W k is the kth convolution kernel matrix;​

[0063] X is an input time series matrix;

[0064] b k is a bias term;

[0065] * is a Hamiltonian operator;

[0066] The calculation result is zero-padded to ensure the consistency of the dimensions before and after input;

[0067] The long short-term memory recurrent neural network LSTM is as follows:

[0068] The output of the convolutional layer flows into the long short-term memory recurrent neural network LSTM recurrent neural network, which is used to mine long-range dependencies between microseismic event sequences and can better capture long-range dependencies through the "gate mechanism" while avoiding the problem of gradient explosion.

[0069] For each long short-term memory recurrent neural network LSTM neural unit, the data x t input at the current time, the output h t-1 of the neural unit at the previous time, and the hidden memory state C t-1 of the previous long short-term memory recurrent neural network LSTM neural unit correspond to the output value h t at the current time and the memory state C t of the long short-term memory recurrent neural network LSTM neural unit at the current time; there are three gate control doors in the long short-term memory recurrent neural network LSTM neural unit: a forgetting door, an input door, and an output door; the forgetting door screens and retains the memory state C t-1 at the previous time, the input door determines how much data x t at the current time can be memorized by the memory unit, and the output door controls how much memory state C t can be output to the output h t at the current time, and the state formula of each door is as follows:

[0070] Forgetting door:

[0071] f t =σ(W f ·[h t-1 ,x t ]+b f ) (11)

[0072] Input door:

[0073] i t =σ(W i ·[h t-1 ,x t ]+bi ) (12)

[0074] Cell state input:

[0075]

[0076] Cell state:

[0077]

[0078] Output gate:

[0079] o t = σ(W o · [h t-1 , x t ] + b o ) (15)

[0080] Current time instant hidden state output:

[0081] h t = o t * tanh(C t ) (16)

[0082] In the formula, W f , W i , W c , W o represent weight matrices;

[0083] b f , b i , b c , b o represent bias vectors;

[0084] tanh is an activation function;

[0085] σ represents a sigmoid activation function;

[0086] [] represents vector connection;

[0087] * is a Hamiltonian operator;

[0088] The small sample prototype network is as follows:

[0089] The hidden features extracted by the long short-term memory recurrent neural network LSTM are clustered, and the hidden feature center points of different category samples 0 and 1 are calculated. By projecting unknown data into the clustering space, the distances between the hidden feature vectors of each unknown data and the center points of different category samples are calculated by the Euclidean distance. The distances of the target to the hidden feature center points of the samples 0 and 1 are compared, the probability of belonging to each sample is obtained by using Softmax, so that the category of the target is analyzed, and the Euclidean distance is taken as the prototype network classification loss. In the training process, it is continuously minimized and the distance to the corresponding class is continuously enlarged, and the distance to other classes is continuously enlarged. The specific calculation formula of the Euclidean distance is as follows:

[0090]

[0091] In the formula, x1 and y1 are the coordinates of the hidden feature vectors of each sample;

[0092] x2 and y2 are the coordinates of the corresponding hidden feature center points;

[0093] The training process is as follows:

[0094] The above time series data M is put into the long short-term memory recurrent neural network LSTM for training, the neural network is constructed, and the hidden features are extracted. The prototype network module is combined with the small sample learning, the loss value of the long short-term memory recurrent neural network LSTM model is calculated by the weighted cross-entropy loss function, and the loss value calculated by the Euclidean distance is added and fused. The parameters in the neural network model are continuously updated by using the recursive backpropagation method, so that the loss of the model on the training data set is minimized. In the training process, sgd is used as the optimizer, and the accuracy is used as the evaluation function. The hyperparameters are set as follows: the training round is 100, the batch size is 400, the decay rate is 0.8, the Softmax is used for normalization processing, and the probability of occurrence of the large energy event is obtained. If the probability of occurrence of the large energy event is greater than the probability of occurrence of the large energy event, 1 is output, otherwise 0 is output. Finally, the model is built, and the prediction model is generated. The cross-entropy loss function used in the application is as follows:

[0095]

[0096]

[0097] In the formula, L i represents the loss value of the i-th precursor mode sequence;

[0098] N represents the number of precursor mode sequences;

[0099] z0 and z1 represent the learning weights of the two categories respectively;

[0100] w0 and w1 represent the sample distribution weights of class 0 and 1 respectively;

[0101] If the label of the i-th precursor mode sequence is a small energy event, y io = 1, y i1 = 0, otherwise y io = 0, y i1 = 1; p i0 is the predicted probability that the observed sample i is of class 0, and p i1 is the predicted probability that the observed sample i is of class 1.

[0102] Compared with the prior art, the present application installs microseismic sensors around the coal mining face, the microseismic sensors transmit the collected original microseismic data to the upper computer of the ground control room, the upper computer performs standardization processing on the original microseismic data, then the microseismic data after standardization is processed by using the principal component analysis method PCA and the kernel density estimation method KDE, corresponding two-dimensional time series data is obtained, the correlation between the impending large energy mine earthquake event and the past mine earthquake event is studied, the density index is constructed, and the data label is completed, finally the long short-term neural network based on small sample learning is constructed, the time series data and the density index are put into the constructed neural network for training, and the prediction of the time period of the future large energy event is completed. The present application can reduce the dependence of the long short-term memory recurrent neural network (LSTM) on data, improve the generalization of the network, so that it can also predict large energy mine earthquakes when the amount of data is small, and provide early warning for rock burst. BRIEF DESCRIPTION OF DRAWINGS

[0103] Figure 1 It is a flow chart of the method of the present application;

[0104] Figure 2 It is a graph of the proportion of each part of the PCA processed parameters of the present application;

[0105] Figure 3 It is a micro-integrated data distribution point on the PCA coordinate of the present application;

[0106] Figure 4 It is a time series graph of the density function of the present application;

[0107] Figure 5 It is a long short-term memory recurrent neural network (LSTM) structure diagram of the present application;

[0108] Figure 6 It is a small sample learning prototype network structure diagram of the present application;

[0109] Figure 7 It is a long short-term memory recurrent neural network (LSTM) network model based on small sample learning of the present application. DETAILED DESCRIPTION

[0110] The application will be further described below with reference to the accompanying drawings.

[0111] As Figures 1-4 shown, a rock burst time series prediction model construction method based on small sample learning, comprising the following steps:

[0112] Step 1, mine earthquake data processing: install microseismic sensors around the coal mining face, the microseismic sensors transmit the collected original microseismic data to the upper computer in the ground control room, the original microseismic data includes microseismic data occurrence time, energy size, and seismic source coordinates, the original microseismic data is processed by the upper computer, first, the original microseismic data is processed according to time period to obtain microseismic data in each time period, then the microseismic data is standardized to obtain time series data in each time period;

[0113] Step 2, density index establishment: density index establishment includes principal component analysis PCA processing and kernel density estimation KDE processing, the time series data of step 1 is processed by principal component analysis PCA to reduce dimension, the original four-dimensional time series data is reduced to two-dimensional time series data; the two-dimensional time series data is processed by kernel density estimation KDE, the probability density of each microseismic data is obtained after processing the two-dimensional time series data, then the maximum probability density in each time period is selected as the density index for microseismic data time series analysis, as Figure 4 shown, the general rule of large energy occurrence time is obtained, that is, the maximum probability density of more than 70% in this time period is selected, when the peak value appears, it is predicted that a large energy event may occur in the future time period (such as within 3 days), thereby completing the construction of microseismic data label, the data label construction is specific: 1 for large energy event, and 0 for other time period data label construction, that is, no large energy event occurs;

[0114] Step 3, establish long short-term memory recurrent neural network LSTM model based on small sample learning: first, the results obtained by the above data analysis (such as Figure 4As shown, the maximum probability density in the time period is selected, and when a peak value appears, the future time period (such as 3 days) is predicted to possibly have a large energy event, and the data label is constructed. The long short-term memory recurrent neural network LSTM is used to model and extract features of the density index, and the regression loss and data features are obtained. Then, the small sample learning prototype network is used. The data features processed by the long short-term memory recurrent neural network LSTM are clustered, and the center point of the privacy feature of different categories of samples in each time period is calculated. The distance between the hidden representation vector of each sample and the corresponding hidden representation center point is measured as the network classification loss. Finally, in the prediction module, the above losses are connected, gradient backpropagation and optimization are performed, the deep learning model based on small sample learning is constructed, and the prediction of the large energy mine earthquake time period is finally output.

[0115] The mine earthquake data processing in step 1 is as follows:

[0116] The fixed time window statistical data is used to construct the time series data set. The time series data includes time and coordinate positions X, Y, and Z (the time series data is all microseismic data contained in the time period in step 1, for example, all microseismic data contained in 0 to 6 hours in a 6-hour time period. The data energy size is only used to determine whether the large energy event is contained in the time window, and is used to derive the rule of the density index subsequently. It does not need to be put into training). Then, the time series data in each time period is standardized to process the data into a standard normal distribution with a mean of 0 and a standard deviation of 1. The purpose is to scale the data by proportion, remove the unit limit of the data, and convert it into a dimensionless pure number, so that different units or orders of magnitude of the index can be compared and weighted. The standardization processing method is as follows:

[0117]

[0118] The average value calculation method is as follows:

[0119]

[0120] The standard deviation calculation method is as follows:

[0121]

[0122] In the formula, x i is the i-th original data, including the time and coordinate position of the i-th microseismic data;

[0123] is the average value of the original data;

[0124] σ(x) is the standard deviation of the original data;

[0125] xi the updated data for the original data;

[0126] n is the total amount of data contained in the time period.

[0127] The density index in step 2 is established as:

[0128] The principal component analysis (PCA) processing method is based on data dimension reduction. The main idea of the method is to recombine the original indexes (for example, M indexes) having certain correlation to form a new set of indexes (for example, N indexes) that are not correlated with each other. The recombined indexes are linear combinations of the original M indexes and M>N. The new indexes are used to replace the original indexes as the input of the model for subsequent analysis. The specific steps are as follows:

[0129] Suppose there are p pieces of standardized data of q dimensions. First, the standardized data is arranged into a q-row-p-column matrix D. Then, the zero mean of each row of the matrix D is calculated, that is, the mean value of the row is subtracted. Then, the covariance matrix is calculated. According to the calculated covariance matrix, the eigenvalues and the corresponding eigenvectors are calculated. The eigenvectors are arranged into a matrix in descending order of the corresponding eigenvalues. The first k rows are taken to form a matrix P. Finally, E=PD is the data after being reduced to k dimensions.

[0130] In this paper, the energy value greater than 105 in the microseismic data is defined as a large energy event. As shown in the description Figure 2 As shown in the figure, the results of the variance contribution rate of the 4 indexes (X, Y, Z, t) after the PCA processing of the 16 large energy events are shown. The variance contribution rate represents the proportion of the variable caused by a single factor in the total variable, which indicates the influence of the common factor on the dependent variable. The gray line in the figure represents the variance contribution rate of the PCA indexes converted from the data containing the large energy events, and the black line represents the average value of the variance contribution rate of the PCA indexes of all large energy events. Therefore, from the figure, it can be analyzed that the first two eigenvalues can contain more than 70% of the information characteristics of the overall microseismic events. The figure provides a certain theoretical basis for reducing the microseismic data to two-dimensional data and then performing kernel density estimation (KDE) two-dimensional plane analysis.

[0131] The kernel density estimation (KDE) processing method is used to estimate the unknown density function in probability theory, which is one of the non-parametric test methods. It uses a smooth peak function ("kernel") to fit the observed data points, thereby simulating the real probability distribution curve. The specific steps are as follows:

[0132] The data processed by the principal component analysis (PCA) is subjected to kernel density estimation (KDE) processing in the present application. The kernel density estimation function is as follows:

[0133]

[0134] In the formula, K(.) is a kernel function (non-negative, integral is 1, meets the probability density property, and the mean is 0). There are many kinds of kernel functions, for example, a Gaussian function is a commonly used kernel function;

[0135] h>0 is a smoothing parameter, called bandwidth, also known as window;

[0136] is a scaled kernel function;

[0137] In the formula, h in the denominator inside the K function is used to adjust the width of the kernel density estimation KDE curve, and h in the denominator outside the K function is used to ensure that the area under the curve meets the kernel density estimation KDE rule; After selecting a suitable kernel function and bandwidth, the kernel density estimation KDE can simulate the real probability distribution curve and obtain a smooth and beautiful result;

[0138] The present application adopts a Gaussian function in the selection of the kernel function, and the selection of the wideband h according to the Silverman rule is as follows:

[0139]

[0140] In the formula, n is the total amount of data contained in the time period;

[0141] d is the data dimension;

[0142] After the above kernel density estimation KDE processing, the probability density of each data point can be obtained, and by griding the principal component analysis PCA coordinate axis, the probability density of each time period can be plotted on the principal component analysis PCA coordinate axis, as shown in the accompanying drawings of the specification.

[0143] In the contour map, the higher probability density indicates that the mine shock event has a stronger clustering effect in the PCA space, that is, the aggregation degree of the past mine shock event. The high-energy event is located in the high-probability region and is close to the maximum probability density, which indicates that the high-energy event has a strong correlation with the microseismic event cluster in the time period and is not isolated. Through this point, it is provided that the high-energy event is not isolated, and mathematical basis is provided. If the development of the event cluster is discovered early enough, rock burst prediction can be carried out. Figure 3 Before the occurrence of the high-energy event, a series of events often occur, and according to the above analysis, the probability density of each microseismic event is obtained, and the maximum probability density in each time period is taken to draw a broken line graph for analysis, as shown in the accompanying drawings of the specification.

[0144] Figure 4 ​As shown, the black dots in the figure represent time periods without large energy events, and the red pentagrams represent time periods with large energy events, and 70% is the size between the maximum and minimum values of the maximum probability density in the entire time period. By the peak of the maximum probability density greater than 70%, it is possible to predict that a large energy event may occur in the future time period, thereby predicting the time of the microseismic large energy event. The density index is the maximum probability density of each time period Figure 4 The time interval adopted in the present application is 6h, and the prediction is for the next 3 days, i.e. 12 time periods. The density index is the maximum probability density of each time period max .

[0145] In step 3, the long short-term memory recurrent neural network LSTM model based on small sample learning is established, and the specific process is as follows:

[0146] First, the prediction model is generated based on the training data set created by the density index processed in step 2, and a time series data set is constructed. The present application adopts 6h as the time window to count the data, and assumes that the data record calculated by the ith time window is m i , which can be expressed as:

[0147]

[0148] In the formula, id is the time window number;

[0149] is the average value of the occurrence location in the microseismic data time period;

[0150] t is the later time in the range of the microseismic data time period;

[0151] For example, when the time period is from 0 to 6, t is selected as 6 o'clock;

[0152] The data m i is added to the corresponding ρ max density index of each time period after standardized processing according to the method of step 1, and is constructed as

[0153] For the problem of data missing in the time period, the nearest neighbor interpolation technique can be used to supplement the missing data, i.e. when the data of the current time period is missing, the data of the previous time period is supplemented to the data of the current time period;

[0154] Based on the above method, the time data sequence data set M can be obtained by traversing the time window:

[0155] M=[m0,m1,m2,......,m n-1 ] (7)

[0156] The label set T corresponding to the time period is represented as:

[0157] T = [t0, t1, t2,..., t n-1 ](8)

[0158] In the formula: t i For each time period, when the time period meets the result after the analysis of the above kernel density estimation KDE processing, a large energy event occurs, at which time t i = 1, otherwise t i = 0;

[0159] The long short-term memory recurrent neural network LSTM module based on small sample learning is as follows:

[0160] The long short-term memory recurrent neural network LSTM module based on small sample learning mainly includes a convolutional layer, a long short-term memory recurrent neural network LSTM, and a small sample prototype network. The data set is selected as 85% for training, 5% for verification, and 10% for testing.

[0161] The convolutional layer is as follows:

[0162] Data first flows into the convolutional module to analyze the short-range dependence relationship between 20 (single sequence length) microseismic events and the relationship between 5 attributes of the events. The convolution kernel size is determined to be 6x5, the number of convolution kernels is 32, and the convolution layer activation function is selected to be the ReLU function, whose expression is as follows:

[0163] ReLU(x) = max(0, x) (9)

[0164] The convolution kernel operation process is as follows:

[0165] h k = ReLU(W k *X + b k ) (10)

[0166] In the formula: W k is the kth convolution kernel matrix;

[0167] X is the input time series matrix;

[0168] b k is the bias;

[0169] * is the Hamiltonian operator;

[0170] The calculation result adopts zero padding method to ensure the consistency of the dimensions before and after input;

[0171] The long short-term memory recurrent neural network LSTM is as follows:

[0172] The output of the convolutional layer flows into a long short-term memory recurrent neural network (LSTM) recurrent neural network. The LSTM is used to mine long-range dependencies between microseismic event sequences. The LSTM can better capture long-range dependencies through a "gate mechanism" and can avoid the problem of gradient explosion. The structure of the LSTM is shown in FIG. 8. Figure 5

[0173] For each LSTM neural unit, the data x t input at the current time, the output h t-1 of the neural unit at the previous time, and the memory state C t-1 of the previous LSTM neural unit are inputted, and the corresponding output is the output value h t at the current time and the memory state C t of the LSTM neural unit at the current time. The LSTM neural unit has three gate control doors: a forgetting door, an input door, and an output door. The forgetting door filters and retains the memory state C t-1 at the previous time, the input door determines how much data x t at the current time can be memorized by the memory unit, and the output door controls how much memory state C t can be output to the output h t at the current time. The state formula of each door is as follows:

[0174] Forgetting door:

[0175] f t = σ(W f · [h t-1 , x t ] + b f ) (11)

[0176] Input door:

[0177] i t = σ(W i · [h t-1 , x t ] + b i ) (12)

[0178] Cell state input:

[0179]

[0180] Cell state:

[0181]

[0182] Output door: ​

[0183] o t =σ(W o ·[h t-1 ,x t ]+b o (15)

[0184] Current hidden state output:

[0185] h t =o t *tanh(C t (16)

[0186] In the formula: W f W i W c W o Represents the weight matrix;

[0187] b f b i b c b o Represents the bias vector;

[0188] tanh is the activation function;

[0189] σ represents the sigmoid activation function;

[0190] [] indicates vector concatenation;

[0191] * represents the Hamiltonian operator;

[0192] The small sample prototype network is as follows:

[0193] Few-shot prototype networks are used to address the dependency of Long Short-Term Memory (LSTM) recurrent neural networks on large amounts of data, improving the network's generalization ability. Examples of few-shot prototype networks are attached. Figure 6 As shown, the implicit features extracted by the Long Short-Term Memory (LSTM) recurrent neural network are clustered, and the implicit feature centroids of samples 0 and 1 of different categories are calculated. Unknown data is projected into the clustering space, and the distance between the implicit feature vectors of each unknown data point and the centroids of samples of different categories is calculated using Euclidean distance. The distances from the target to the implicit feature centroids of samples 0 and 1 are compared, and the probability of belonging to each sample is obtained using Softmax, thus analyzing the target's category. Simultaneously, Euclidean distance is used as the classification loss of the prototype network. During training, the distance between the target and its corresponding class is continuously minimized, while the distance between the target and other classes is increased. The specific formula for calculating Euclidean distance is as follows:

[0194]

[0195] In the formula: x1 and y1 are the coordinates of the implicit feature vectors of each sample;

[0196] x2 and y2 are coordinates of corresponding implicit feature center points;

[0197] The long short-term memory recurrent neural network LSTM network model based on small sample learning is as shown in the accompanying drawings. Figure 7

[0198] The training process is specifically as follows:

[0199] The above time series data M is put into the long short-term memory recurrent neural network LSTM for training, neural network construction and implicit feature extraction, combined with a small sample learning prototype network module, the loss value of the long short-term memory recurrent neural network LSTM model is calculated through a weighted cross-entropy loss function, which is added and fused with the loss value calculated by the Euclidean distance, and the recursive backpropagation method is used to continuously update the parameters in the neural network model, so that the loss of the model on the training data set is minimized, and sgd is used as the optimizer during the training process, and the accuracy is used as the evaluation function, and the hyperparameters are set as follows: the training round is 100, the batch size is 400, the decay rate is 0.8, and Softmax is used for normalization processing, and the probability of occurrence of a large energy event is obtained. If the probability of a large energy event is greater than the probability of no large energy event, 1 is output, otherwise 0 is output. Finally, the model is built to generate a prediction model, and the cross-entropy loss function used in the present application is as follows:

[0200]

[0201]

[0202] In the formula: L i represents the loss value of the i-th precursor mode sequence;

[0203] N represents the number of precursor mode sequences;

[0204] z0 and z1 represent the learning weights of the two categories, respectively;

[0205] w0 and w1 represent the sample distribution weights of categories 0 and 1, respectively;

[0206] If the label of the i-th precursor mode sequence is a small energy event, then y io = 1, y i1 = 0, otherwise y io = 0, y i1 = 1; p i0 is the predicted probability that the observed sample i is category 0, and p i1 is the predicted probability that the observed sample i is category 1.​

Claims

1. A method for constructing a rock burst timing prediction model based on small sample learning, characterized in that, Comprise the following steps: Step 1, mine shock data processing: install microseismic sensors around the coal mining face, the microseismic sensor transmits the collected original microseismic data to the host computer in the ground control room, the original microseismic data includes microseismic data occurrence time, energy size, and focal coordinates, the original microseismic data is processed by the host computer, first, the original microseismic data is processed according to the time period to obtain the microseismic data in each time period, then the microseismic data is standardized to obtain the time series data of each time period; Step 2, density index establishment: the density index establishment includes principal component analysis PCA processing and kernel density estimation KDE processing, the time series data of step 1 is processed by principal component analysis PCA, the original four-dimensional time series data is reduced to two-dimensional time series data; the two-dimensional time series data is processed by kernel density estimation KDE, and the probability density of each microseismic data is obtained after processing the two-dimensional time series data, then the largest probability density in each time period is selected as the density index for microseismic data time series analysis, the law of large energy occurrence time is obtained, when there is a peak value, it is predicted that the future time period will have a large energy event, thereby completing the construction of the microseismic data label, and the large energy event is 1, and the data label of other time periods is 0, that is, no large energy event occurs; Step 3, establish a long short-term memory recurrent neural network LSTM model based on small sample learning: use the long short-term memory recurrent neural network LSTM to model and extract features of the density index, obtain the regression loss and data features; then use the small sample learning prototype network, cluster the data features processed by the long short-term memory recurrent neural network LSTM, and calculate the center point of the privacy features of different categories of samples in each time period, and use the Euclidean distance to measure the distance between the hidden feature vector of each sample and the corresponding hidden feature center point as the network classification loss; finally, connect the above loss on the prediction module, perform gradient backpropagation and optimization, complete the construction of the deep learning model based on small sample learning, and finally output the prediction of the large energy mine shock time period.

2. The rock burst timing prediction model construction method based on small sample learning according to claim 1, characterized in that, The mine shock data processing in step 1 is: Fixed time window statistics data is used to construct a time series data set, which includes time and coordinate positions X, Y, and Z, then the time series data in each time period is standardized, and the data is processed into a standard normal distribution with a mean of 0 and a standard deviation of 1, the specific method of standardization is as follows: （1） The average value calculation method is as follows: （2） The standard deviation calculation method is as follows: （3） In the formula: is the first original data, including the first time and coordinate position of the microseismic data; is the average value of the original data; is the standard deviation of the raw data; updated data for the original data; Total amount of data contained in the time period.

3. The rock burst timing prediction model construction method based on small sample learning according to claim 2, characterized in that, The density index establishment in step 2 is: The principal component analysis PCA processing method is based on data dimension reduction, which recombines the original M indicators with correlation to form a new set of N indicators that are not related to each other, the N indicators obtained after recombination are linear combinations of the original M indicators and satisfy M>N, the new N indicators replace the original M indicators as the input of the model for subsequent analysis, the specific steps are as follows: Assuming there are p pieces of q-dimensional standardized data, first the standardized data is grouped by column to form a matrix D of q rows and p columns, then each row of the matrix D is zero-meaned, and then the covariance matrix is calculated According to the calculated covariance matrix, the eigenvalues and their corresponding eigenvectors are calculated, the eigenvectors are arranged in rows in descending order of the corresponding eigenvalues to form a matrix, the first k rows are taken to form a matrix P, and finally E = PD is the data reduced to k dimensions. The data after PCA dimension reduction is processed by kernel density estimation (KDE). The kernel density estimation function is as follows: （4） In the formula: is a kernel function; is a smoothing parameter, called bandwidth, also known as window; to scale the kernel function; In the denominator inside the function adjusts the width of the kernel density estimation KDE curve, while the denominator outside the function ensures that the area under the curve meets the rules of kernel density estimation KDE. After selecting a suitable kernel function and bandwidth, kernel density estimation (KDE) can simulate the real probability distribution curve. The Gaussian function is adopted in the selection of the kernel function, and the wideband selection is shown as follows: （5） In the formula: is the total amount of data contained in the time period; Data is the dimension; After kernel density estimation (KDE) processing, the probability density of each data point can be obtained. By gridding the principal component analysis (PCA) coordinate axis, the probability density of each time period can be plotted on the principal component analysis (PCA) coordinate axis.

4. The rock burst timing prediction model construction method based on small sample learning according to claim 3, characterized in that, The long short-term memory (LSTM) model based on small sample learning in step 3 is established as follows: First, the prediction model is generated based on the training data set created by the density index processed in step 2, and a time series data set is constructed. A 6h time window is taken to count the data, and it is assumed that the data record calculated by the first time window is , which can be expressed as: ​ （6） In the formula: is the time window number; , , is the average of the occurrence locations in the time period of the microseismic data; is the time period for the microseismic data; and t is the time period for the microseismic data; and Will The data obtained after standardization in step 1 is added to the corresponding data for each time period. Density index, constructed ; When the data of the current time period is missing, the data of the previous time period is used to supplement the data of the current time period. Based on the above method, the time window is traversed to obtain a time data sequence data set : （7） The time period corresponds to a label set is represented as: （8） In the formula: For each time period, when the time period meets the result after the kernel density estimation (KDE) processing analysis, a large energy event occurs, at which time , otherwise ; The deep learning model based on small sample learning is as follows: The deep learning model based on small sample learning includes convolutional layer, long short-term memory (LSTM), and small sample prototype network. The data set is selected as follows: 85% as training set, 5% as validation set, and 10% as test set. The convolutional layer is as follows: The data first flows into the convolution module to analyze the short-range dependence between 20 microseismic events and the relationship between the 5 attributes of the events. The convolution kernel size is determined to be 6x5, the number of convolution kernels is 32, and the activation function of the convolution layer is selected as The function has the following expression: （9） The convolution kernel operation process is as follows: （10） In the formula: is the first convolution kernel matrix; is the input time series matrix; To be paranoid; * is the Hamiltonian operator; The calculation result is padded with zeros to ensure the consistency of the input and output dimensions. The long short-term memory (LSTM) is as follows: The output of the convolutional layer flows into the long short-term memory (LSTM) recurrent neural network. The long short-term memory (LSTM) is used to mine the long-range dependencies between microseismic event sequences. For each long short-term memory recurrent neural network LSTM neural unit, input data at current time , output of neural unit at last time , and hidden memory state of last long short-term memory recurrent neural network LSTM neural unit , corresponding output is output value at current time , and memory state of long short-term memory recurrent neural network LSTM neural unit at current time ; there are three control gates in long short-term memory recurrent neural network LSTM neural unit: forget gate, input gate and output gate; forget gate filters memory state at last time , input gate determines data input at current time , and output gate controls how much memory state can be output to output at current time , and how much can be memorized by memory unit , state formulas of forget gate, input gate and output gate are as follows: Forget gate: (11) Input gate: (12) Cell state input: (13) Cell state: (14) Output gate: (15) Current time hidden state output: (16) In the formula: , , , represents a weight matrix; , , , denotes a bias vector; is an activation function; denotes a sigmoid activation function; [] represents vector connection; * is the Hamiltonian operator; The small sample prototype network is as follows: By clustering the hidden features extracted by the long short-term memory (LSTM) and calculating the hidden feature center points of different category samples 0 and 1, the unknown data is projected into the clustering space. The Euclidean distance between the hidden feature vector of each unknown data and the center point of different category samples is calculated. By comparing the distance between the target and the hidden feature center points of samples 0 and 1, the probability of belonging to each sample is obtained by using Softmax, so as to analyze the category of the target. At the same time, the Euclidean distance is used as the prototype network classification loss, which is continuously minimized during the training process. It also enlarges the distance between it and other categories. The specific calculation formula of Euclidean distance is as follows: (17) wherein: and are the coordinates of the implicit feature vector of each sample; and are coordinates of the corresponding implicit feature center point; The training process is as follows: The time series data above Put into long short-term memory recurrent neural network LSTM training, construct the neural network and extract the implicit features, combine the small sample learning prototype network module, calculate the loss value of the long short-term memory recurrent neural network LSTM model through the weighted cross-entropy loss function, add and fuse it with the loss value calculated by the Euclidean distance, and constantly update the parameters in the neural network model by using the recursive back propagation method, so that the loss of the model on the training data set is minimized, sgd is used as the optimizer during the training process, and the accuracy rate is used as the evaluation function, the hyperparameters are set as follows: the training round is 100, the batch size is 400, the decay rate is 0.8, the Softmax is used for normalization processing, and the probability of occurrence of large energy event and no large energy event is obtained. If the probability of large energy event is greater than the probability of no large energy event, output 1, otherwise output 0, finally complete the construction of the model, thereby generating a prediction model, and the cross-entropy loss function used is as follows: （18） （19） In the formulae: denotes the loss value of the first precursor pattern sequence; represents the number of precursor pattern sequences; and respectively denote two categories of learning weights; and denote the sample distribution weights for classes 0 and 1, respectively; If the tag of the nth precursor pattern sequence is a small energy event, then , , , else , ; is the observed sample is the predicted probability for class 0, is the observed sample is the predicted probability for class 1.

Images