Underground mine rock mass instability precursor intelligent decision-making platform based on LSTM

Abstract

The invention provides an underground mine rock mass instability precursor intelligent decision-making platform based on LSTM, and belongs to the technical field of mine rock mass monitoring. An LSTM neural network is used for constructing a rock mass instability precursor intelligent identification model IIM, 16 acoustic emission time sequence parameters are used as input vectors, and three states of non-early-warning [0, 0], primary early-warning [1, 0] and secondary early-warning [1, 1] are used as output ends; the rock mass instability precursor intelligent identification model effectively identifies the rock shear slip instability and judges the early warning level, and the identification effect is good; a rock instability precursor intelligent decision-making platform IDMP is constructed based on a rock instability precursor intelligent identification model, is composed of an early warning identification layer, an early warning analysis layer and an early warning decision-making layer, and can carry out intelligent decision making of whether early warning is carried out or not and judging the early warning level.

Description

Technical Field

[0001] The invention provides an underground mine rock mass instability precursor intelligent decision-making platform based on LSTM, belonging to the technical field of mine rock mass monitoring. Background Art

[0002] Real-time monitoring of mine rock masses is critical for early warning of instability. Large-scale mining operations cause great disturbances to mine rock masses, which are severely deformed and damaged in complex forms, further increasing the difficulty of monitoring and early warning. Mining is developing towards large-scale and large-scale operations, and the complexity of mine rock mass monitoring and early warning has increased dramatically. The factors affecting the stability of mine rock masses are extremely complex, and the selection of precursor warning characterization parameters relies on experience, and the monitoring accuracy and early warning accuracy are still not guaranteed. From the perspective of intelligent prediction, it is of great practical significance to carry out the identification of rock instability precursor warning information.

[0003] After years of development, the stability monitoring and instability precursor warning of mine rock mass have formed a relatively stable research idea. In the 1970s, the Geological Society of London and Deere proposed the Pugh hardness coefficient method and RQD method based on single factors. In view of the incompleteness of the stability monitoring information of mine rock mass, considering that the application of single factors in the stability monitoring of mine rock mass is prone to low representation efficiency and large errors, many mathematical methods such as multi-index grey correlation analysis, multi-index analytical analysis, and multi-index comprehensive analysis have been applied. With the development of information technology and computer generalization capabilities, artificial neural networks have the characteristics of self-learning ability, nonlinear dynamic processing and adaptive mode, and artificial neural networks that can effectively identify complex nonlinear systems have been applied. When the computing power of computers is further enhanced, the diversification of the representation forms of the stability of mine rock mass engineering can be considered. The combined time series monitoring information and image monitoring information have the advantages of being intuitive and easy to understand. The stability monitoring of mine rock mass and the acquisition of instability characterization information are more comprehensive, and data mining technology has been successfully applied in this field. The above research fully considers the discreteness and multidimensionality of monitoring information, and has been applied in underground goafs and open-pit slopes.

[0004] However, the factors affecting the stability of rock mass in mines are extremely complex. The selection of traditional precursor warning parameters relies heavily on engineering experience, resulting in insufficient monitoring and warning accuracy and strong randomness, and the monitoring and warning effect cannot be guaranteed. With the help of big data thinking, the importance of building an intelligent decision-making platform for rock instability precursor warning is self-evident. Summary of the invention

[0005] Based on the idea of ​​"integrated development of deep learning and mine rock mass monitoring", the present invention proposes an intelligent decision-making platform for rock instability precursors based on LSTM. Based on the idea of ​​big data, data cleaning is no longer performed. An intelligent identification model (IIM) for rock instability precursors is constructed using an LSTM neural network, with 16 acoustic emission time series parameters as input vectors and three states of no warning [0,0], first-level warning [1,0], and second-level warning [1,1] as output ends; the intelligent identification model (IIM) for rock instability precursors can effectively identify rock shear slip instability and determine the warning level, with good recognition effect; finally, based on the intelligent identification model (IIM) for rock instability precursors, an intelligent decision-making platform (IDMP) for rock instability precursors is constructed. IDMP consists of a warning recognition layer, a warning analysis layer, and a warning decision layer, and can make intelligent decisions on whether to warn and determine the warning level.

[0006] The specific technical solution is:

[0007] An intelligent decision-making platform for underground mine rock instability precursors based on LSTM includes three modules: early warning identification layer, early warning analysis layer, and early warning decision layer.

[0008] The early warning recognition layer normalizes the collected monitoring information to form the input vector of LSTM. The early warning information intelligent recognition layer, based on the big data concept, directly converts the collected monitoring information E (e1, e2, ..., e n ) is normalized to form the input vector [x1, x2, ..., x i ,...x N ], no longer selecting feature data. The LSTM neural network is used to build an intelligent identification model for rock instability precursors, optimize the recognition accuracy and efficiency of the model from multiple angles, and output the output sample K (k1, k2, ..., k t )’s representation vector [0, 0], [1, 0], [1, 1].

[0009] The warning analysis layer designs an analysis circuit based on the warning identification results to provide warning information for the warning decision layer. Output [0, 0], k0 is closed, k1 and k2 are open, and the green light is on; output [1, 0], k1 is closed, k0 and k2 are open, and one red light is on, a first-level warning; output [1, 1], k1 and k2 are closed, k2 is open, two red lights are on, a second-level warning.

[0010] The warning decision layer is based on the output results of the warning analysis layer analysis circuit, and the warning level is determined by the decision maker D. If the output result is not 1, no warning is given, and the decision is "normal production". If there is 1, the warning mode is entered, [1,0] is the first-level warning, and the decision is "production, focus on the changes in monitoring information"; [1,1] is the second-level warning, and the decision is "stop production, take measures such as pressure relief and support key areas".

[0011] Among them, the LSTM neural network constructs an intelligent identification model for rock instability precursors, including the following steps:

[0012] (1) Input sample E(e1, e2, ..., e n )Build

[0013] According to the monitoring information collection of the monitoring means, an n-dimensional input sample E (e1, e2, ..., e n ).

[0014] (2) Output samples K (k1, k2, ..., k t )Build

[0015] According to the purpose of precursor identification, a t-dimensional output sample K (k1, k2, ..., k t ).

[0016] (3) LSTM model training and testing

[0017] Use input sample E and output sample K to train the LSTM model. Assume that the initial learning rate of the training is set to 1×10 -4 The training batch size is set to 32, the total number of training rounds is set to 100 epochs, the loss function uses the binary cross entropy loss function BCELoss, and the optimizer uses the Adam optimizer, so that the model can dynamically adjust the learning rate to optimize the training effect.

[0018] It also includes model performance evaluation, using the mean absolute error (MAE) and root mean square error (RMSE) as evaluation indicators. MAE is the average of the absolute errors between the predicted value and the true value, as shown in Formula 1; RMSE is the square root of the average of all squared errors, as shown in Formula 2:

[0019]

[0020] In the above formula, n is the number of samples. is the predicted value, y i is the true value.

[0021] Formula 3 is the discriminant formula of the model effect:

[0022]

[0023] If formula 3 holds true, it means that the model is qualified and the recognition effect of the model meets the requirements.

[0024] The technical effects of the present invention are as follows:

[0025] (1) Based on the idea of ​​big data, data cleaning is no longer performed. An intelligent identification model for rock instability precursors is constructed based on LSTM. The model is input by all the time series parameters of acoustic emission and outputs three warning levels: no warning, first-level warning, and second-level warning.

[0026] (2) An intelligent decision-making platform for rock instability precursors was constructed. The platform consists of a warning identification layer, a warning analysis layer, and a warning decision-making layer, which enables intelligent decision-making on whether to issue a warning and the identification of the warning level.

[0027] (3) Through the whole process of "early warning identification layer → early warning analysis layer → early warning decision layer", a full process is provided for the early warning work of rock instability precursors, which includes identifying early warning information from monitoring information, analyzing the early warning information to determine the early warning level, and finally making early warning decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Schematic diagram of the invented LSTM structure;

[0029] Figure 2 This is an intelligent decision-making platform for precursors of rock mass instability in underground mines of the present invention. DETAILED DESCRIPTION

[0030] The specific technical solution of the present invention is explained in conjunction with embodiments.

[0031] 1. Construction of intelligent identification model of rock instability precursors based on LSTM;

[0032] Long Short-Term Memory Network (LSTM) is a special type of Recurrent Neural Network (RNN), which is usually used to deal with long-term dependency problems. LSTM includes input gate, forget gate, output gate and memory cell. Figure 1 .

[0033] The memory unit is used to save and update the long-term dependency information of the entire sequence processing process; the input gate is used to decide which information is stored in the memory unit; the forget gate is used to decide which information should be deleted from the memory unit; the output gate is used to decide which information will be output to the next time step. Each gate uses the sigmoid activation function to determine the flow of information.

[0034] LSTM overcomes the gradient vanishing and gradient exploding problems of traditional RNN in long sequence learning. LSTM is highly flexible. LSTM can process input sequences of different lengths and has been widely used in various fields, such as financial market forecasting, weather forecasting, earthquake forecasting and other time series forecasting fields.

[0035] (1) Input sample E(e1, e2, ..., e n )Build;

[0036] The learning sample is related to the redundancy of the model input vector, which is an important part of improving the recognition accuracy and optimizing the learning performance of the neural network. According to the monitoring information collection of the monitoring means, an n-dimensional input sample E (e1, e2, ..., e n ).

[0037] (2) Output samples K (k1, k2, ..., k t )Build;

[0038] According to the purpose of precursor identification, a t-dimensional output sample K (k1, k2, ..., k t ).

[0039] (3) LSTM model training and testing;

[0040] Use input sample E and output sample K to train the LSTM model. Assume that the initial learning rate of the training is set to 1×10 -4 The training batch size is set to 32, the total number of training rounds is set to 100 epochs, the loss function uses the binary cross entropy loss function (BCE Loss), and the optimizer uses the Adam optimizer, so that the model can dynamically adjust the learning rate to optimize the training effect.

[0041] Rock fracture acoustic emission data has the characteristics of large amount in a short time, huge information, and high dimension, which just conforms to the idea of ​​big data. The learning sample construction of the present invention draws on the idea of ​​big data and no longer performs data cleaning. According to the acoustic emission timing information obtained from the experiment in Table 1, there are 16 feature vectors including rise time, count, energy, duration, amplitude, average frequency, RMS, ASL, peak frequency, threshold, back-calculated frequency, initial frequency, signal strength, absolute energy, center frequency, and peak frequency. The learning sample consists of these 16 feature vectors.

[0042] The rock mass instability precursor intelligent identification model proposed in the present invention is based on the idea of ​​big data, and no data cleaning is performed. All the collected effective feature information is used for the input layer feature vector.

[0043] The input vector is composed of 16 parameters of rock failure acoustic emission time series, and the output vector has three states, namely no warning, first-level warning, and second-level warning. Assume that "00" is no warning, "10" is first-level warning, and "11" is second-level warning. The input vector is composed of 16 acoustic emission time series parameters [rise time, count, energy, duration, amplitude, average frequency, RMS, ASL, peak frequency, threshold, back-calculated frequency, initial frequency, signal strength, absolute energy, center frequency, peak frequency] after normalization; the output vector is represented by [00,10,11].

[0044] The dataset (Table 2) is divided into a training set (Table 3) and a test set (Table 4). The training set is used for model training, and the test set is used to evaluate the performance of the model. The initial learning rate of the training is set to 1×10-4, the training batch size is set to 32, the total number of training rounds is set to 100 epochs, the loss function uses the binary cross entropy loss function (BCE Loss), and the optimizer uses the Adam optimizer, so that the model can dynamically adjust the learning rate to optimize the training effect.

[0045]

[0046]

[0047] 2. Build an intelligent decision-making platform for underground mine rock instability precursors based on LSTM;

[0048] Based on the construction and testing of qualified LSTM model, an intelligent decision-making platform is built for intelligent decision-making of precursors of rock instability. Figure 2 This is a schematic diagram of the intelligent decision-making platform for rock instability precursors. The platform includes three modules, namely the early warning identification layer, the early warning analysis layer, and the early warning decision-making layer.

[0049] Early warning recognition layer: The collected monitoring information is normalized to form the input vector of LSTM. The early warning information intelligent recognition layer, based on the big data concept, directly converts the collected monitoring information E (e1, e2, ..., e n ) is normalized to form the input vector [x1, x2, ..., x i ,...x N ], no longer selecting feature data. The LSTM neural network is used to build an intelligent identification model for rock instability precursors, optimize the recognition accuracy and efficiency of the model from multiple angles, and output the output sample K (k1, k2, ..., k t )’s representation vector [0, 0], [1, 0], [1, 1].

[0050] Early warning analysis layer: Based on the early warning identification results, an analysis circuit is designed to provide early warning information for the early warning decision layer. Output [0, 0], k0 is closed, k1 and k2 are open, and the green light is on; output [1, 0], k1 is closed, k0 and k2 are open, one red light is on, and the first-level early warning; output [1, 1], k1 and k2 are closed, k2 is open, two red lights are on, and the second-level early warning.

[0051] Early warning decision layer: Based on the output results of the early warning analysis layer, the decision maker D determines the early warning level. If the output result is not 1, no early warning is given and the decision is "normal production". If there is 1, the early warning mode is entered, [1,0] is the first-level early warning, and the decision is "production, focus on changes in monitoring information"; [1,1] is the second-level early warning, and the decision is "stop production, take measures such as pressure relief and support key areas".

[0052] In short, through the whole process of "early warning identification layer → early warning analysis layer → early warning decision layer", a full process is provided for the early warning work of rock instability precursors, which includes identifying early warning information from monitoring information, analyzing the early warning information to determine the early warning level, and finally making early warning decisions.

[0053] 3. Model recognition results

[0054] The mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation indicators. MAE is the average of the absolute errors between the predicted value and the true value (Formula 1), and RMSE is the square root of the average of all error squares (Formula 2):

[0055]

[0056] In the above formula, n is the number of samples. is the predicted value, y i is the true value.

[0057] Formula 3 is the discriminant formula of the model effect:

[0058]

[0059] If formula 3 holds true, it means that the model is qualified and the recognition effect of the model meets the requirements.

[0060] The model recognition effect discriminant constructed by formula 3 is used to calculate the T and K values, T≈K≈0, and the model test results are shown in Table 4. The intelligent early warning information recognition model constructed by the present invention can use the rock acoustic emission time series information to effectively identify the precursor information of rock shear slip instability and determine the early warning level, and the recognition effect is good. It shows that the model is qualified and the recognition effect of the model meets the requirements.

Claims

1. An intelligent decision-making platform for underground mine rock mass instability precursors based on LSTM, characterized in that: It includes three modules: early warning identification layer, early warning analysis layer, and early warning decision layer; The early warning recognition layer normalizes the collected monitoring information to form the input vector of LSTM. The early warning information intelligent recognition layer, based on the big data concept, directly converts the collected monitoring information E (e1, e2, ..., e n ) is normalized to form the input vector [x1, x2, ..., x i ,...x N ], no longer selecting feature data; constructing the rock mass instability precursor intelligent identification model through LSTM neural network, optimizing the recognition accuracy and efficiency of the model from multiple angles, and outputting the output sample K (k1, k2, ..., k t )’s representation vector [0,0], [1,0], [1,1]; The warning analysis layer designs an analysis circuit based on the warning identification results to provide warning information to the warning decision layer; Output [0, 0], k0 is closed, k1 and k2 are open, and the green light is on; output [1, 0], k1 is closed, k0 and k2 are open, and a red light is on, a first-level warning; Output [1, 1], k1, k2 are closed, k2 is open, two red lights are on, level 2 warning; The warning decision layer determines the warning level through the decision maker D based on the output result of the warning analysis layer analysis circuit; If the output result does not have 1, there is no warning, and the decision is "normal production"; if there is 1, the warning mode is entered, [1,0] is the first-level warning, and the decision is "production, focus on changes in monitoring information"; [1,1] is the second-level warning, and the decision is "stop production, take measures such as pressure relief and support key areas".

2. According to claim 1, an LSTM-based intelligent decision-making platform for underground mine rock mass instability precursors is characterized in that: The LSTM neural network constructs an intelligent identification model for rock mass instability precursors, including the following steps: (1) Input sample E(e1, e2, ..., e n )Build; According to the monitoring information collection of the monitoring means, an n-dimensional input sample E (e1, e2, ..., e n ); (2) Output samples K (k1, k2, ..., k t )Build; According to the purpose of precursor identification, a t-dimensional output sample K (k1, k2, ..., k t ); (3) LSTM model training and testing; Using input sample E and output sample K, train the LSTM model; assuming that the initial learning rate of the training is set to 1×10 -4 The training batch size is set to 32, the total number of training rounds is set to 100 epochs, the loss function uses the binary cross entropy loss function BCELoss, and the optimizer uses the Adam optimizer, so that the model can dynamically adjust the learning rate to optimize the training effect.

3. The LSTM-based intelligent decision-making platform for underground mine rock mass instability precursors according to claim 2 is characterized in that: It also includes model performance evaluation, which is done by using mean absolute error (MAE) and root mean square error (RMSE) as evaluation indicators; MAE is the average of the absolute errors between the predicted value and the true value, as shown in formula 1; RMSE is the square root of the average of all squared errors, as shown in formula 2: In the above formula, n is the number of samples. is the predicted value, y i is the true value; Formula 3 is the discriminant formula of the model effect: If formula 3 holds true, it means that the model is qualified and the recognition effect of the model meets the requirements.

Images