WO2023056695A1 - Deep recurrent neural network-based coal and gas outburst accident prediction and recognition method - Google Patents

Abstract

The present invention relates to the technical field of coal mine safety, and relates to a deep recurrent neural network-based coal and gas outburst accident prediction and recognition method. The present invention provides a deep recurrent neural network-based coal and gas outburst accident prediction and recognition method. Data features and inherent correlations thereof are zoomed in while the data dimension is reduced; by using a deep recurrent neural network training method, the inherent features of data can be effectively learned, and a large amount of model training proves that the method is quite robust; a method of combining multiple model evaluation indicators is used, so that a selected model is more representative.

Description

A Coal and Gas Outburst Accident Prediction and Identification Method Based on Deep Recurrent Neural Network technical field

The invention relates to the technical field of coal mine safety, and relates to a coal and gas outburst accident prediction and identification method based on a deep cycle neural network.

Background technique

Coal and gas outburst accidents refer to the phenomenon that the coal and gas in the mine suddenly and continuously burst into the roadway and working face in a very short period of time. It is an extremely complex and dangerous dynamic disaster, which will not only endanger the underground The life safety of the staff will also damage the expensive excavation equipment, which is one of the major accidents caused by coal mine safety. Therefore, how to predict and identify this phenomenon is very important. Based on the deep learning algorithm, this method analyzes the gas data in real time, trains the deep cycle neural network model, and looks for abnormal features in the newly collected gas data. Divided into normal data and abnormal data, the normal data will participate in the next model training, and the abnormal data will alarm to remind the mine staff. Traditional coal mine gas data identification methods include support vector machines, decision trees, and regression methods. In addition to methods based on machine learning, they also include statistical learning methods such as time series decomposition algorithms. Deep recurrent neural network algorithms include but are not limited to Recurrent Neural Network (RNN for short), Long Short-Term Memory (LSTM for short) and Gated Recurrent Unit (GRU for short).

At present, in view of the time-varying and non-linear characteristics of coal mine gas data, combined with the characteristics of high complexity and large amount of gas data, it is difficult for traditional data analysis methods to accurately identify the features contained in the data and at the same time There is generalization ability in time. However, the deep recurrent neural network has enough neurons to process high-dimensional gas data, and the network model can regularly forget and update the learned gas data characteristics, so the deep recurrent neural network has certain advantages in processing gas data. However, coal mine gas data is still highly unbalanced data, which makes the importance of data preprocessing and representation methods as well as model selection and training more important. Appropriate methods are used to process gas data uploaded from underground mines, and data reconstruction and model Structure selection and training often determine the accuracy of this method for gas data prediction and identification, so reasonable data processing and model selection are very necessary.

Contents of the invention

(1) Solved technical problems

The purpose of the present invention is to provide a coal and gas outburst accident prediction and identification method based on a deep cycle neural network to avoid coal and gas outburst accidents caused by gas accumulation.

(2) Technical solutions

A coal and gas outburst accident prediction and identification method based on deep cyclic neural network, comprising the following steps:

S1, real-time acquisition of underground coal mine gas data, preprocessing and reconstruction of the gas data, to obtain a coal mine gas data set that can be used to train deep cyclic neural networks;

S2, use the deep recurrent neural network to learn the coal mine gas data set, and get n models according to the model structure and training method;

S3, measure n deep recurrent neural network models through different measurement indicators, compare the loss size and correlation between the actual value of the model and the predicted value, and select the model with strong generalization ability and the smallest loss;

S4. Deploy the optimal model to the coal mine server, and input the latest gas data into the model; if the calculation result is not in the range of the three measurement indicators, it can be judged that the gas at the data collection location is abnormal, otherwise it is normal.

Preferably, S1 includes:

S101, collect the gas sensors of each working face in real time from the mine central server, including the sensor data of the upper corner (T0), T1, T2, T3 and other positions);

S102, respectively preprocessing the obtained data according to the sensor position;

S103, reconstructing the preprocessed gas data;

S104, loop through the entire gas data to obtain a reconstructed gas data set D _data .

Preferably, the pre-processed content in S102 includes:

1) read data;

2) Null filling, including linear interpolation, cubic interpolation, and zero-order interpolation;

3) Set the index and rearrange the data with time as the index;

4) Delete the adjustment value, where the adjustment value is a human factor, and the value is far greater than the methane content, and the characteristic is that the adjustment value will only appear once or twice;

5) Delete duplicate data, using time as an index, delete duplicate data on the timeline, where the definition of duplication means that the year, month, day, hour, minute, and second are completely consistent.

Preferably, the reconstruction method in S103 is to use a connected segment of values to represent the next value of the segment in linear time, to form data in matrix format, written as lag _m , and m means to use consecutive m values to represent the mth +1 value.

Preferably, S2 includes:

S201, the gas data set D _data is divided into a training set D _train and a test set D _test ;

S202, select the architecture of the deep recurrent neural network model (in the model training process, determine _L2 as the loss function, and Adam as the optimizer of the neural network);

S203, determine the number of cells, the batch batch, the model learning rate lr, and the number of training rounds Epoch, each individual change of the above parameters will generate a new model, and the selection of these parameters will affect the training results of the model, and have It may lead to overfitting and underfitting, and the parameters should be fine-tuned at this time.

Preferably, the gas data set D _data in S201 is divided into a training set D _train and a test set D _test according to a ratio of 7:3.

Preferably, in S202, three recurrent neural network structures are selected but not limited to, namely, RNN, LSTM and GRU.

Preferably, S3 includes:

S301, respectively input the test set D _test into the n models obtained in S2, and obtain the model prediction value D'_test;

S302, calculate the MSE, RMSE loss between the real test value and the predicted value of different models; calculate the R ² between the real test value and the predicted value of different models;

S303, screening out the minimum loss of MSE and RMSE and the maximum value of R ² as the optimal model.

Preferably, the formula for calculating MSE is:

The formula for calculating RMSE is:

The formula to calculate ^R2 is:

in

is the predicted value of the model, y _t is the real test value,

is the average of the real values.

Preferably, in order to ensure the accuracy and generalization ability of the model, the optimal model is retrained once in a while.

(3) Beneficial effects

The invention provides a coal and gas outburst accident prediction and identification method based on a deep cyclic neural network, which enlarges the data features and their inherent correlation while reducing the data dimension; the deep cyclic neural network training method can effectively learn At the same time, a large number of model training proves that the method is quite robust; the method of combining multiple model evaluation indicators makes the selected model more representative.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings that are required for the description of the embodiments. Obviously, the drawings in the following description are only for the present invention and protect some embodiments. For those skilled in the art, other drawings can also be obtained based on these drawings without creative effort.

Fig. 1 is a basic flow chart shown according to the embodiment of this specification.

Fig. 2 is a basic frame diagram according to the embodiment of this specification.

Fig. 3 is a general structure diagram of a deep recurrent neural network according to an embodiment of the present specification.

Fig. 4 is a graph showing experimental results according to an embodiment of the present specification.

Fig. 5 is the result map of the first 100 points enlarged in Fig. 4.

Detailed ways

The technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Apparently, the described embodiments are some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In the description of the present invention, it should be noted that if the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" appear ", etc., the indicated orientation or positional relationship is based on the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation , constructed and operated in a particular orientation and therefore should not be construed as limiting the invention. In addition, if the terms "first", "second", and "third" appear, they are used for descriptive purposes only, and should not be understood as indicating or implying relative importance.

In the description of the present invention, it should be noted that, unless otherwise clearly stipulated and limited, if the terms "installation", "connection" and "connection" appear, they should be understood in a broad sense, for example, it can be a fixed connection or a It is a detachable connection or an integral connection; it may be a mechanical connection or an electrical connection; it may be a direct connection or an indirect connection through an intermediary, and it may be the internal communication of two components. Those of ordinary skill in the art can understand the specific meanings of the above terms in the present invention in specific situations.

Example

Referring to accompanying drawing 1, a kind of coal and gas outburst accident prediction and identification method based on deep cycle neural network, this method comprises:

S1, real-time acquisition of underground gas data, preprocessing and reconstruction of the gas data, to obtain a gas data set that can be used to train the deep loop neural network;

Among them, the gas data obtained from the underground comes from the gas sensors of each working face, and the sensor positions include but not limited to the upper corner (T0), T1, T2, T3 and other positions; data preprocessing is to read the collected data, And set the index based on the timestamp. For empty values, two methods are used: filling interpolation and direct deletion. The interpolation methods include linear interpolation, cubic interpolation, zero-order interpolation, etc.; direct deletion includes deleting adjustment values and deleting Duplicate data.

The basic frame diagram shown in this embodiment is shown in Figure 2; gas data reconstruction generally chooses continuous m values to represent the m+1th value, which is written as lag _m , which can effectively convert the gas data in time In connection, the trend of a piece of data determines a value. Generally speaking, the larger the value of m, the greater the amount of information contained, but too large a value of m will lead to a proportional increase in time complexity. In this embodiment In , the value of m is 128, and the expression of the reconstructed gas data set D _data is as follows:

D _data ＝{[X ₁ :(x ₁ ,x ₂ ,...,x ₁₂₈ ),Y ₁ :(x ₁₂₉ )],[X ₂ :(x ₂ ,x ₃ ,...,x ₁₂₉ ) ,Y ₂ :(x ₁₃₀ )],...,[X _n :(x _n+1 ,x _n+2 ,...,x _n+127 ),Y _n :(x _n+128 )]} .

S2, use the deep recurrent neural network to learn the coal mine gas data set, and get n models according to the model structure and training method;

Divide the reconstructed data D _data into a training set D _train and a test set D _test , and divide them according to the ratio. The purpose of segmentation is to fit the deep cyclic neural network as much as possible, so that it can learn the potential features of the data set as much as possible without overfitting. In this embodiment, the ratio used is the training set and the test set Set 7: 3, that is, D _train =D _data ×0.7, D _test =D _data −D _train .

The training set D _train of the gas data set is input into the deep recurrent neural network for training; the data is m-dimensional stacking in time (m is the value selected for data reconstruction), and it is one-dimensional data in space, so the deep recurrent neural network The network only accepts one-dimensional arrays; in the present embodiment, the deep recurrent neural network selected is one layer, and i cells are connected, and the different numbers of cells determine the length and performance of the deep recurrent neural network. In this embodiment, The number of selected cells is 64, 128, 256 respectively; other parameters such as learning rate lr, batch batch and number of training rounds Epoch will also affect the performance of the model. In this embodiment, the parameter selection is respectively: learning rate lr= [0.0001,0.0003,0.0005,0.0009,0.001], batch batch=[32,64,128], and the number of training rounds Epoch=[20,30,40];

Among them, the deep recurrent neural network structure used in this embodiment includes RNN, LSTM, GRU; the general structure of the model based on RNN is shown in Figure 3; the optimization function used is Adam, and the activation function of each layer of neural network is softmax and Relu.

S3, measure n deep recurrent neural network models through different measurement indicators, compare the loss size and correlation between the actual value of the model and the predicted value, and select the model with strong generalization ability and the smallest loss;

The obtained n models are comprehensively measured according to the three values of mean square error, root mean square error and coefficient of determination, wherein the mean square error and root mean square error should be as small as possible and the coefficient of determination should be as large as possible; Specifically, the test set D _test of the gas data set is input into the model to obtain the predicted value D′ _test , and the test set D _test and the predicted value D′ _test are calculated using three metrics (MSE, RMSE loss, R ² ), Get the optimal model and the intervals of the three measurement indicators; the data and models obtained after learning can be used in coal mine accident analysis, coal mine gas prediction and other fields;

The results of this embodiment are shown in Figure 4, the comparison chart between the 200 data points predicted by the three models and the real gas data, and the right side of the figure is the result of zooming in on the first 100 points.

S4. Deploy the optimal model to the coal mine server, and input the latest gas data into the model. If the calculation result is not in the range of the three measurement indicators, it can be judged that the gas at the data collection location is abnormal, otherwise it is normal;

Among them, in order to ensure the accuracy and generalization ability of the model, the optimal model is retrained once in a while; in this embodiment, the optimal model is retrained every seven days, that is, the normal data of the past seven days and the past Three months of data to retrain the model; considering that the characteristics of underground gas data will change over time, in this embodiment, the length of the training set is selected as three months;

In this example, the underground data obtained are collected from several real mines in Anhui Province; including several working face gas data from 2020 to 2021, each sensor collects once or dozens of times per minute depending on the settings of the underground sensors , the time is not uniform, so the data preprocessing and data reconstruction mentioned in this method are necessary.

In this embodiment, the W3220 working face of a certain mine in Anhui is selected as the gas data set during the period from 2021-05-01 to 2021-07-20, including a total of 160,412 pieces of data.

From the experimental results in Figures 4 and 5, among the three models used in this embodiment, the GRU model has the best effect, and has the characteristics of high prediction accuracy, small loss value and high robustness. However, according to the selected parameters and model structure, RNN and LSTM also have the above characteristics in some cases, which further illustrates the practicability and research significance of deep recurrent neural network in the fields of gas prediction and gas anomaly analysis. From the above results, it can be seen that for coal mine gas data, the method proposed by the present invention can effectively analyze and learn the potential characteristics of gas data, and has certain practical and future significance for ensuring safe coal mine collection.

In the description of this specification, descriptions with reference to the terms "one embodiment", "example", "specific example" and the like mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment of the present invention. In an embodiment or example. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

The preferred embodiments of the invention disclosed above are only to help illustrate the invention. The preferred embodiments do not exhaust all details nor limit the invention to only specific embodiments. Obviously, many modifications and variations can be made based on the contents of this specification. This description selects and specifically describes these embodiments in order to better explain the principle and practical application of the present invention, so that those skilled in the art can well understand and utilize the present invention. The invention is to be limited only by the claims, along with their full scope and equivalents.

Claims (10)

1. A coal and gas outburst accident prediction and identification method based on a deep cyclic neural network, characterized in that it includes the following steps:

   S1, real-time acquisition of underground coal mine gas data, preprocessing and reconstruction of the gas data, to obtain a coal mine gas data set that can be used to train deep cyclic neural networks;

   S2, use the deep recurrent neural network to learn the coal mine gas data set, and get n models according to the model structure and training method;

   S3, measure n deep recurrent neural network models through different measurement indicators, compare the loss size and correlation between the actual value of the model and the predicted value, and select the model with strong generalization ability and the smallest loss;

   S4. Deploy the optimal model to the coal mine server, and input the latest gas data into the model; if the calculation result is not in the range of the three measurement indicators, it can be judged that the gas at the data collection location is abnormal, otherwise it is normal.
2. A coal and gas outburst accident prediction and identification method based on deep cyclic neural network according to claim 1, characterized in that S1 includes:

   S101, collect the gas sensors of each working face in real time from the mine central server;

   S102, respectively preprocessing the obtained data according to the sensor position;

   S103, reconstructing the preprocessed gas data;

   S104, loop through the entire gas data to obtain a reconstructed gas data set D _data .
3. A coal and gas outburst accident prediction and identification method based on deep cyclic neural network according to claim 2, characterized in that, the content of preprocessing in S102 includes:

   1) read data;

   2) Null filling, including linear interpolation, cubic interpolation, and zero-order interpolation;

   3) Set the index and rearrange the data with time as the index;

   4) Delete the adjustment value, where the adjustment value is a human factor, and the value is far greater than the methane content, and the characteristic is that the adjustment value will only appear once or twice;

   5) Delete duplicate data, using time as an index, delete duplicate data on the timeline, where the definition of duplication means that the year, month, day, hour, minute, and second are completely consistent.
4. A method for predicting and identifying coal and gas outburst accidents based on deep cyclic neural network according to claim 2, characterized in that, the reconstruction method in S103 is in linear time, representing the value of the segment with a connected segment of values The latter value, which forms the data in matrix format, is written as lag _m , where m means to use consecutive m values to represent the m+1th value.
5. A coal and gas outburst accident prediction and identification method based on deep cyclic neural network according to claim 1, characterized in that S2 includes:

   S201, the gas data set D _data is divided into a training set D _train and a test set D _test ;

   S202, select the architecture of the deep recurrent neural network model;

   S203, determine the number of cells, the batch batch, the model learning rate lr, and the number of training rounds Epoch.
6. A coal and gas outburst accident prediction and identification method based on deep cyclic neural network according to claim 5, characterized in that in S201, the gas data set D data is divided into training sets D _train and D _data according to the ratio of 7:3 The test set D _test .
7. A coal and gas outburst accident prediction and identification method based on a deep cyclic neural network according to claim 5, characterized in that, in S202, three cyclic neural network structures are selected but not limited to, namely RNN, LSTM and GRU .
8. A coal and gas outburst accident prediction and identification method based on deep cyclic neural network according to claim 1, characterized in that S3 includes:

   S301, respectively input the test set D _test into the n models obtained in S2, and obtain the model prediction value D'_test;

   S302, calculate the MSE, RMSE loss between the real test value and the predicted value of different models; calculate the R ² between the real test value and the predicted value of different models;

   S303, screening out the minimum loss of MSE and RMSE and the maximum value of R ² as the optimal model.
9. A coal and gas outburst accident prediction and identification method based on a deep cyclic neural network according to claim 8, wherein the formula for calculating MSE is:

   

   The formula for calculating RMSE is:

   

   The formula to calculate ^R2 is:

   

   in

   

   is the predicted value of the model, y _t is the real test value,

   

   is the average of the real values.
10. A coal and gas outburst accident prediction and identification method based on deep cyclic neural network according to claim 1, characterized in that, in order to ensure the accuracy and generalization ability of the model, the optimal model is retrained once in a while .

Images