Rock burst early warning method based on KAN neural network
Through the impact ground pressure early warning method based on KAN neural network, the problems of limited fitting ability and parameter complexity in the prior art are solved, and more efficient and interpretable impact ground pressure early warning results are achieved, which are suitable for multiple mine environments.
Patent Information
- Application Number
- CN202411026021.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-15
AI Technical Summary
The existing impact ground pressure early warning algorithm based on neural networks has problems with limited fitting capabilities, huge computing parameters, long training and inference time, and difficult to explain gray box state, resulting in poor warning results and difficult to apply to different mines.
Using the shock ground pressure early warning method based on KAN neural network, the microseismic monitoring data is preprocessed, and the four-layer neural network algorithm is constructed. The cross entropy loss function and the backpropagation algorithm are used to train the parameters, which reduces the number of training parameters and activates the function through SiLU. The and spline functions enhance nonlinear description capabilities and achieve the improvement of early warning results.
It realizes impact ground pressure warning with better performance without relying on artificially set warning indicators. The parameters are clear and the calculation formula is highly interpretable, and it is suitable for multiple mine environments.
Smart Images

Figure CN118964990A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of early warning of dangerous disasters such as rock burst in mines, and specifically to a rock burst early warning method based on a KAN neural network. Background Art
[0002] Rock burst is a typical dynamic disaster in deep mining, with the characteristics of violent destruction and sudden occurrence. Rock burst can cause casualties and damage to tunnels. At present, the use of mine monitoring data to provide real-time early warning of rock burst is the prerequisite for effectively preventing and controlling the occurrence of rock burst.
[0003] Microseismic monitoring data is a kind of high-quality monitoring data that can reflect the mining situation of mines. Due to the complexity of the rock burst warning task, it is difficult to obtain satisfactory warning results by directly using neural network algorithms to process microseismic monitoring data. In the existing rock burst warning algorithms based on neural networks (such as patent ZL202111558332.8), microseismic monitoring data is often used to construct rock burst warning indicators, and then the rock burst warning indicators are weighted by complex neural network algorithms to calculate the impact hazard. In the past, due to the limited fitting ability of the network, it is often necessary to rely on expert experience to manually determine the explicit rock burst warning indicators. These warning indicators can often only be used for a certain mine and have limited fitting ability. In the past, the neural network in the rock burst warning algorithm is often composed of convolutional neural networks and recurrent neural networks. The rock burst warning algorithms of these neural networks often have huge calculation parameters and have the disadvantages of long training and reasoning time. And due to the complexity of these networks themselves, these networks themselves are often in a gray box state during the rock burst warning process, and it is difficult to determine the meaning of the node parameters in the network. Summary of the invention
[0004] To solve the above problems, the present invention proposes a rock burst early warning method based on a KAN neural network, which is used to determine the probability that a future microseismic event will have a rock burst hazard, comprising the following steps:
[0005] S1: Preprocessing of raw microseismic monitoring data
[0006] The original microseismic monitoring data includes a number of continuously monitored and recorded original microseismic events, each of which includes the time of occurrence, the X coordinate of the location, the Y coordinate of the location, the Z coordinate of the location, the energy, and whether it is rock burst; the original microseismic events are preprocessed to obtain preprocessed microseismic events, the preprocessed microseismic events include the time interval, the X coordinate of the location, the Y coordinate of the location, the Z coordinate of the location, the energy, and whether it is rock burst, wherein the time interval is the time interval between the microseismic event and the previous microseismic event;
[0007] S2: Establishing a rock burst warning dataset
[0008] Extract M consecutive pre-processed microseismic events, and use the data of the mth to m+tth pre-processed microseismic events as the data source of a sample in the rock burst warning data set. 1≤m≤Mt, then at most Mt samples can be established, each sample includes features and labels, wherein the features of each sample include the time interval, location X coordinate, location Y coordinate, location Z coordinate and energy of the mth to m+t-1th pre-processed microseismic events, and the label of each sample is the rock burst hazard of the m+tth microseismic event; divide the rock burst warning data set into a training set and a test set;
[0009] S3: Constructing a rock burst warning algorithm based on KAN neural network
[0010] The rock burst warning algorithm based on KAN neural network has 4 layers of neural networks. Each layer of neural network includes input, calculation function and output. The output of the previous layer of neural network is the input of the next layer of neural network. The number of neurons in each layer of neural network is n1, n2, n3 and n4, n4 = 2.
[0011] The input of the rock burst warning algorithm is also the input of the first layer of the neural network, which is the sample feature of the rock burst warning data set. It includes time interval, location X coordinate, location Y coordinate, location Z coordinate and energy, which are respectively counted as x 0,1 、x 0,2 、x 0,3 、x 0,4 、x 0,5 ;
[0012] The number of computational functions of the first layer of neural network is equal to the product of its input number n0=5 and its number of neurons, and its output is the number of neurons of the first layer of neural network; the number of inputs of the second layer and subsequent neural networks is equal to the number of neurons of the previous layer of neural network; the number of computational functions of the second layer and subsequent neural networks is equal to the product of the number of neurons of the previous layer and the number of neurons of this layer, and the number of outputs of the second layer and subsequent neural networks is equal to its number of neurons;
[0013] Use (l,i l ) represents the i-th neuron of the l-th layer of the neural network, and the neuron (l, i l ) output Indicates; the lth layer of the neural network connects its input and its neurons (l, i l ) is calculated as For the l-th layer of the neural network, assuming that it has a total of n l neurons, whose input number is the number of neurons in the previous layer of neural network nl-1 , when it is the first layer of neural network, its input is sample feature x 0,i , input number n l-1 =5, neuron (l, i l ) is the sum of all inputs to the neuron:
[0014]
[0015] In the formula, For connection input The calculation function of the i-th neuron in the l-th layer of the neural network; is the i-th layer in the l-1th layer of the neural network l-1 The output of the neuron is also the input of the l-th layer of the neural network. When l = 1 Represents sample characteristics;
[0016] The rock burst early warning algorithm based on KAN neural network has 4 layers of neural networks. The output of the rock burst early warning algorithm based on KAN neural network is:
[0017]
[0018] Among them, y1, y2 represent the probability of no rock burst danger and the probability of rock burst danger respectively; i1, i2, i3 are the i-th neurons in the first, second and third layers of the neural network respectively; i0 represents the i-th input of the first layer of the neural network;
[0019] Among them, the calculation function is the sum of the basis function b(x) and the spline function:
[0020]
[0021] Where w is a trainable parameter; spline(x) is a spline function; b(x) is a basis function, and x is an input value;
[0022] b(x)=SiLU(x)=x / (1+e -x ) (4)
[0023] Where SiLU represents the Sigmoid Linear Unit activation function; e is a natural constant;
[0024] Among them, spline(x) is a linear spline function:
[0025]
[0026] In the formula, B z(·) represents the corresponding z-order term in the spline curve. This patent uses a cubic spline curve, so the maximum value of z is Z=3; c z Indicates that for B z The weight value of (·) is a trainable parameter;
[0027] S4: Training a rock burst warning algorithm based on a KAN neural network
[0028] Firstly, an initial value is given to each trainable parameter in the rock burst warning algorithm based on the KAN neural network. For the spline function in the calculation function, its trainable parameter is initialized to 0, and other trainable parameters are initialized according to the Xavier initialization rule. The training set samples are input into the rock burst warning algorithm based on the KAN neural network for training. The cross entropy loss function is used to evaluate the difference between the warning result of the rock burst warning algorithm and the sample label. The back propagation algorithm is used to update the parameters in the rock burst warning algorithm that need to be trained.
[0029] The performance of the trained rock burst warning algorithm is evaluated using the test set samples; the parameters of the rock burst warning algorithm with the best effect are saved to obtain the trained rock burst warning algorithm based on the KAN neural network.
[0030] Preferably, step S1 further comprises: standardizing the time interval, location X coordinate, location Y coordinate, location Z coordinate and energy of the pre-processed microseismic event.
[0031] Preferably, in step S2, the label needs to be processed by One-Hot encoding, and after the processing, the label of rock burst is 1, and the label of non-rock burst is 0.
[0032] Preferably, the method further comprises step S5: applying a rock burst warning algorithm based on a KAN neural network
[0033] The latest monitored continuous t original microseismic events are obtained, and data processing is performed according to steps S1-S2 to obtain a sample with only features but no labels. The features of the sample are input into the rock burst early warning algorithm based on the KAN neural network trained in step S4 to obtain the probability that a future microseismic event will have a rock burst hazard.
[0034] Beneficial effects of the present invention: The parameters that need to be trained in the rock burst warning algorithm based on the KAN neural network of the present invention are only w and c. z, compared with the previous neural network, the training parameters are greatly reduced; and because the calculation function in the KAN neural network has better nonlinear description ability, the rock burst warning algorithm based on the KAN neural network can achieve better performance than the traditional neural network with only a small number of parameters without relying on artificially set warning indicators. At the same time, since the calculation function and learnable parameters in the algorithm are clear, the formula of the algorithm itself can be expressed for the rock burst warning task (see formula 2), which has a certain degree of interpretability. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 Flowchart of rock burst warning method based on KAN neural network;
[0036] Figure 2 Framework diagram of rock burst warning algorithm based on KAN neural network. DETAILED DESCRIPTION
[0037] The technical solution of the present invention is described in more detail below in conjunction with the accompanying drawings.
[0038] A rock burst early warning method based on a KAN neural network is used to warn the probability that a future microseismic event has rock burst danger (is rock burst) without artificially setting early warning indicators, and includes the following steps:
[0039] S1: Preprocessing of raw microseismic monitoring data
[0040] The original microseismic monitoring data includes a number of continuously monitored and recorded original microseismic events, each of which includes the time of occurrence, the X coordinate of the location, the Y coordinate of the location, the Z coordinate of the location, the energy, and whether it is rock burst; the original microseismic events are preprocessed to obtain preprocessed microseismic events, the preprocessed microseismic events include the time interval, the X coordinate of the location, the Y coordinate of the location, the Z coordinate of the location, the energy, and whether it is rock burst, wherein the time interval is the time interval between the microseismic event and the previous microseismic event, and the time interval, the X coordinate of the location, the Y coordinate of the location, the Z coordinate of the location, and the energy are standardized; the standardized processing is a commonly used data processing method in the field, which will not be repeated here;
[0041] S2: Establishing rock burst warning data set
[0042] Extract M consecutive pre-processed microseismic events, and use the data of the mth to m+tth pre-processed microseismic events as the data source of a sample in the rock burst warning data set. 1≤m≤Mt, then at most Mt samples can be established. Each sample includes features and labels, wherein the features of each sample include the time interval, location X coordinate, location Y coordinate, location Z coordinate and energy of the mth to m+t-1th pre-processed microseismic events. The label of each sample is the rock burst hazard of a future microseismic monitoring event, that is, the rock burst hazard of the m+tth microseismic event (whether it is rock burst). The label needs to be processed by One-Hot encoding. After processing, the label of rock burst is 1, and the label of non-rock burst is 0; the rock burst warning data set is divided into a training set and a test set in a ratio of 7:3;
[0043] S3: Constructing a rock burst warning algorithm based on KAN neural network
[0044] like Figure 1-2 As shown, the rock burst warning algorithm based on the KAN neural network has a total of 4 layers of neural networks, and the number of neurons in each layer of the neural network is n1, n2, n3 and n4 respectively; in this embodiment, n1, n2, n3 and n4 are 8, 16, 8 and 2 respectively, wherein the number of neurons n4 in the last layer of the neural network is related to the rock burst hazard type. In this patent, the rock burst warning results only have two cases: rock burst hazard and no rock burst hazard, so the number of neurons n4 in the last layer of the neural network is 2; each layer of the neural network includes input, calculation function and output, wherein the number of calculation functions is an integer multiple of neurons;
[0045] The input of the rock burst warning algorithm is also the input of the first layer of the neural network, which is the sample feature x in the rock burst warning data set. 0,i0 , including time interval, location X coordinate, location Y coordinate, location Z coordinate and energy, respectively recorded as x 0,1 、x 0,2 、x 0,3 、x 0,4 、x 0,5 The output of the rock burst warning algorithm is also the output of the fourth layer of neural network, which is the probability that a future microseismic event is rock burst (there is rock burst danger) and the probability that it is not rock burst (there is no rock burst danger), corresponding to the label of the sample; the output of the previous layer of neural network is the input of the next layer of neural network;
[0046] The number of calculation functions of the first layer of neural network is equal to the product of its input number 5 and its number of neurons, which is 40 in this embodiment, and its output is the number of neurons of the first layer of neural network; the number of inputs of the second layer and subsequent neural networks is equal to the number of neurons of the previous layer of neural network; the number of calculation functions of the second layer and subsequent neural networks is equal to the product of the number of neurons of the previous layer (which is also the number of outputs of the previous neural network and the number of inputs of the current layer of neural network) and the number of neurons of the current layer, and the number of the second, third and fourth layers of neural networks in this embodiment are 128, 128 and 16 respectively; the number of outputs of the second layer and subsequent neural networks is equal to the number of neurons;
[0047] In the form of coordinates (l, i l ) represents the i-th neuron of the l-th layer of the neural network, and the neuron (l, i l ) output Indicates that at the same time It is also the input of the l+1th layer of the neural network; for example, neuron (2, 6) represents the 6th neuron of the 2nd layer of the neural network, x 2,6 Represents the output of neuron (2, 6), which is also the input of the third layer of neural network; the first layer of neural network connects its input and its neurons (l, i l ) is calculated as For example, the output of the 5th neuron in the second layer of the neural network is x 2,5 , which is also the input of the third layer of neural network. The third layer of neural network connects its input x 2,5 The calculation function of its sixth neuron is recorded as For the l-th layer of the neural network, assuming that it has a total of n l neurons, whose input number is the number of neurons in the previous layer of neural network n l-1 , when it is the first neural network, its input is the sample feature x 0,i , input number n l-1 =5, neuron (l, i l ) is the sum of all inputs to the neuron:
[0048]
[0049] In the formula, For connection input The calculation function of the i-th neuron in the l-th layer of the neural network; is the i-th layer in the l-1th layer of the neural network l-1 The output of the neuron is also the input of the l-th layer of the neural network. When l = 1 Represents sample characteristics; is the i-th layer in the l-th layer of the neural networkl The output of the neuron is also the input of the l+1th layer of neural network;
[0050] The rock burst early warning algorithm based on KAN neural network has 4 layers of neural networks. The features of the samples are used as input vectors. The output of the rock burst early warning algorithm based on KAN neural network is:
[0051]
[0052] Among them, y1 and y2 represent the probability of no rock burst danger and the probability of rock burst danger respectively; n0 represents the input vector, i.e., the sample feature There are five parameters: time interval, location X coordinate, location Y coordinate, location Z coordinate and energy. 0,1 、x 0,2 、x 0,3 、x 0,4 、x 0,5 , n0=5; i1, i2, i3 are the i-th neurons in the first, second, and third layers of the neural network respectively; i0 represents the i-th input of the first layer of the neural network;
[0053] Among them, the calculation function is the sum of the basis function b(x) and the spline function:
[0054]
[0055] Where w is a trainable parameter; spline(x) is a spline function; b(x) is a basis function, and x is an input value;
[0056] b(x)=SiLU(x)=x / (1+e -x ) (4)
[0057] Where SiLU represents the Sigmoid Linear Unit activation function; e is a natural constant;
[0058] In this embodiment, spline(x) is a linear spline function:
[0059]
[0060] In the formula, B z (·) represents the corresponding z-order term in the spline curve. This patent uses a cubic spline curve, so the maximum value of z is Z=3; c z Indicates for B z The weight value of (·) is a trainable parameter;
[0061] In the past, convolutional neural networks and recurrent neural networks firstly use a calculation function to calculate the input value as an output value, and then use an activation function to activate the output value. Due to the structure of the algorithm itself, the structure of these neural networks is difficult to express with a formula. The activation function of the KAN neural network constructed by the present invention is within the calculation function, and after being combined with learnable parameters, it has a stronger nonlinear description ability, and its structure can be expressed by formula (2);
[0062] S4: Training a rock burst warning algorithm based on a KAN neural network
[0063] Firstly, an initial value is given to each trainable parameter in the rock burst warning algorithm based on the KAN neural network. For the spline function in the calculation function, its trainable parameter is initialized to 0, and other trainable parameters are initialized according to the Xavier initialization rule. The training set samples are randomly shuffled and input into the rock burst warning algorithm based on the KAN neural network for training. The cross entropy loss function is used to evaluate the difference between the warning result of the rock burst warning algorithm and the sample label. The back propagation algorithm is used to update the parameters in the rock burst warning algorithm that need to be trained.
[0064] The performance of the trained rock burst warning algorithm was evaluated using the test set samples. After 200 training cycles, the best rock burst warning algorithm parameters were saved to obtain the trained rock burst warning algorithm based on the KAN neural network.
[0065] The method also includes step S5: application of rock burst warning algorithm based on KAN neural network
[0066] The latest monitored continuous t original microseismic events are obtained, and data processing is performed according to steps S1-S2 to obtain a sample with only features but no labels. The features of the sample are input into the rock burst warning algorithm based on the KAN neural network trained in step S4 to obtain the probability that a future microseismic event will have a rock burst hazard (is a rock burst).
[0067] The parameters that need to be trained in the rock burst warning algorithm based on the KAN neural network of the present invention are only w and c. z , compared with the previous neural network, the training parameters are greatly reduced; and because the calculation function in the KAN neural network has better nonlinear description ability, the rock burst warning algorithm based on the KAN neural network can achieve better performance than the traditional neural network with only a small number of parameters without relying on artificially set warning indicators. At the same time, since the calculation function and learnable parameters in the algorithm are clear, the formula of the algorithm itself can be expressed for the rock burst warning task (see formula 2), which has a certain degree of interpretability.
Claims
1. A rock burst early warning method based on a KAN neural network, used to determine the probability that a future microseismic event will have a rock burst hazard, characterized in that: The steps include: S1: Preprocessing of raw microseismic monitoring data The original microseismic monitoring data includes a number of continuously monitored and recorded original microseismic events, each of which includes the time of occurrence, the X coordinate of the location, the Y coordinate of the location, the Z coordinate of the location, the energy, and whether it is rock burst; the original microseismic events are preprocessed to obtain preprocessed microseismic events, the preprocessed microseismic events include the time interval, the X coordinate of the location, the Y coordinate of the location, the Z coordinate of the location, the energy, and whether it is rock burst, wherein the time interval is the time interval between the microseismic event and the previous microseismic event; S2: Establishing a rock burst warning dataset Extract M consecutive pre-processed microseismic events, and use the data of the mth to m+tth pre-processed microseismic events as the data source of a sample in the rock burst warning data set. 1≤m≤Mt, then at most Mt samples can be established, each sample includes features and labels, wherein the features of each sample include the time interval, location X coordinate, location Y coordinate, location Z coordinate and energy of the mth to m+t-1th pre-processed microseismic events, and the label of each sample is the rock burst hazard of the m+tth microseismic event; divide the rock burst warning data set into a training set and a test set; S3: Constructing a rock burst warning algorithm based on KAN neural network The rock burst warning algorithm based on KAN neural network has 4 layers of neural networks. Each layer of neural network includes input, calculation function and output. The output of the previous layer of neural network is the input of the next layer of neural network. The number of neurons in each layer of neural network is n1, n2, n3 and n4, n4 = 2. The input of the rock burst warning algorithm is also the input of the first layer of the neural network, which is the sample feature of the rock burst warning data set. It includes time interval, location X coordinate, location Y coordinate, location Z coordinate and energy, which are respectively counted as x 0,1 、x 0,2 、x 0,3 、x 0,4 、x 0,5 ; The number of computational functions of the first layer of the neural network is equal to the product of its input number n0=5 and its number of neurons, and its output is the number of neurons in the first layer of the neural network; The number of inputs of the second and subsequent neural networks is equal to the number of neurons in the previous layer of the neural network; the number of computational functions of the second and subsequent neural networks is equal to the product of the number of neurons in the previous layer and the number of neurons in this layer, and the number of outputs of the second and subsequent neural networks is equal to the number of neurons; Use (l,i l ) represents the i-th neuron of the l-th layer of the neural network, and the neuron (l, i l ) output is x l,il Indicates; the l-th layer of the neural network connects its input x l-1,il-1 and its neurons (l, i l ) is calculated as For the l-th layer of the neural network, assuming that it has a total of n l neurons, whose input number is the number of neurons in the previous layer of neural network n l-1 , when it is the first layer of neural network, its input is sample feature x 0,i , input number n l-1 =5, neuron (l, i l ) is the sum of all inputs to the neuron: In the formula, For connection input The calculation function of the i-th neuron in the l-th layer of the neural network; is the i-th layer in the l-1th layer of the neural network l-1 The output of the neuron is also the input of the l-th layer of the neural network. When l = 1 Represents sample characteristics; The rock burst early warning algorithm based on KAN neural network has 4 layers of neural networks. The output of the rock burst early warning algorithm based on KAN neural network is: Among them, y1, y2 represent the probability of no rock burst danger and the probability of rock burst danger respectively; i1, i2, i3 are the i-th neurons in the first, second and third layers of the neural network respectively; i0 represents the i-th input of the first layer of the neural network; Among them, the calculation function is the sum of the basis function b(x) and the spline function: Where w is a trainable parameter; spline(x) is a spline function; b(x) is a basis function, and x is an input value; b(x)=SiLU(x)=x / (1+e -x ) (4) Where SiLU represents the Sigmoid Linear Unit activation function; e is a natural constant; Among them, spline(x) is a linear spline function: In the formula, B z (·) represents the corresponding z-order term in the spline curve. This patent uses a cubic spline curve, so the maximum value of z is Z=3; c z Indicates that for B z The weight value of (·) is a trainable parameter; S4: Training a rock burst warning algorithm based on a KAN neural network Firstly, an initial value is given to each trainable parameter in the rock burst warning algorithm based on the KAN neural network. For the spline function in the calculation function, its trainable parameter is initialized to 0, and other trainable parameters are initialized according to the Xavier initialization rule. The training set samples are input into the rock burst warning algorithm based on the KAN neural network for training. The cross entropy loss function is used to evaluate the difference between the warning result of the rock burst warning algorithm and the sample label. The back propagation algorithm is used to update the parameters in the rock burst warning algorithm that need to be trained. The performance of the trained rock burst warning algorithm is evaluated using the test set samples; the parameters of the rock burst warning algorithm with the best effect are saved to obtain the trained rock burst warning algorithm based on the KAN neural network.
2. The rock burst warning method according to claim 1, characterized in that: Step S1 also includes: standardizing the time interval, location X coordinate, location Y coordinate, location Z coordinate and energy of the pre-processed microseismic event.
3. The rock burst warning method according to claim 1, characterized in that: In step S2, the label needs to be processed by One-Hot encoding. After the processing, the label of rock burst is 1, and the label of non-rock burst is 0.
4. The rock burst warning method according to any one of claims 1 to 3, characterized in that: The method also includes step S5: application of rock burst warning algorithm based on KAN neural network The latest monitored continuous t original microseismic events are obtained, and data processing is performed according to steps S1-S2 to obtain a sample with only features but no labels. The features of the sample are input into the rock burst early warning algorithm based on the KAN neural network trained in step S4 to obtain the probability that a future microseismic event will have a rock burst hazard.
Citation Information
Patent Citations
CNN-based micro-seismic event risk prediction method considering rock burst precursor information
CN113947197A
Rock burst early warning method for comparing long-term and short-term characteristics of micro-seismic data
CN117708705A
Rock burst prediction method based on comparison of different danger monitoring data
CN117849855A
Cited By
Forest fire point detection method based on KAN network and multi-sensor fusion
CN119296241A
