CN-121980839-A - Rock burst simulation prediction method and device based on graph Transformer

CN121980839ACN 121980839 ACN121980839 ACN 121980839ACN-121980839-A

Abstract

The application relates to a rock burst simulation prediction method and a device based on a graph Transformer, wherein the method comprises the steps of constructing a mapping relation between input features and output features of each sub-model according to a finite element model in rock burst simulation; generating a node training sample and a unit training sample respectively through finite element calculation and a mapping relation, training a displacement directed graph neural network by using the node training sample to obtain an initial displacement sub-model, training a stress directed graph neural network by using the unit training sample to obtain an initial stress sub-model, training and optimizing the initial displacement sub-model and the initial stress sub-model by applying physical boundary constraint to the initial displacement sub-model and the initial stress sub-model and constructing a loss function to obtain a target displacement sub-model and a target stress sub-model, and predicting rock burst by using the target displacement sub-model and the target stress sub-model. The problem that rock burst cannot be rapidly and accurately predicted is solved.

Inventors

WANG SHOUGUANG
GAO FUQIANG
WU RUI
ZHAO CHENXI
DONG SHUANGYONG
PENG XIANGYUAN

Assignees

中煤科工开采研究院有限公司

Dates

Publication Date: 20260505
Application Date: 20251128

Claims (10)

1. The rock burst simulation prediction method based on the graph Transformer is characterized by comprising the following steps of: Constructing a mapping relation between input features and output features of each sub-model according to a finite element model in rock burst simulation, wherein the sub-models comprise an initial displacement sub-model and an initial stress sub-model, and the sub-models are obtained by combining a directed graph neural network and a transducer model; Generating node training samples and unit training samples respectively through finite element calculation and the mapping relation; Training a displacement directed graph neural network by using the node training sample to obtain the initial displacement sub-model, and training a stress directed graph neural network by using the unit training sample to obtain the initial stress sub-model; the initial displacement sub-model and the initial stress sub-model are subjected to physical boundary constraint, and the initial displacement sub-model and the initial stress sub-model are subjected to training optimization by constructing a loss function, so that a target displacement sub-model and a target stress sub-model are obtained; And predicting rock burst by using the target displacement submodel and the target stress submodel.
2. The method according to claim 1, wherein the constructing the mapping relation between the input feature and the output feature of each sub-model according to the finite element model in the rock burst simulation comprises: Extracting node input features of finite element nodes and unit input features of finite element units according to the finite element model, wherein the node input features and the unit input features are 9-dimensional vectors; And mapping the node output tensors corresponding to the node input features to obtain node mapping relations, and mapping the unit output tensors corresponding to the unit input features to obtain unit mapping relations, wherein the node mapping relations represent the input and output structures of the initial displacement submodel, and the unit mapping relations represent the input and output structures of the initial stress submodel.
3. The method according to claim 2, wherein generating node training samples and unit training samples from the mapping relation by finite element calculation includes: Generating a plurality of sets of computation samples using the random input information; performing finite element calculation on the calculation sample to obtain original node data matched with the node mapping relation and original unit data matched with the unit mapping relation; And carrying out standardization processing on the original node data to obtain the node training sample, and carrying out standardization processing on the original unit data to obtain the unit training sample.
4. The method of claim 1, wherein training the displacement directed graph neural network using the node training samples to obtain the initial displacement submodel comprises: Constructing the displacement directed graph neural network according to the node connection relation of the finite element model, wherein the graph vertex of the displacement directed graph neural network is a node, and the connection between the nodes is a directed edge; mapping the displacement input in the node training sample to a hidden space of the displacement directed graph neural network to obtain a first hidden representation; encoding first position information of nodes into the first hidden representation to form a first feature representation, and extracting deep node features of the first feature representation by using a transducer encoder; constructing a first output head through the deep node characteristics, and outputting a displacement prediction result according to the first output head; And under the condition that the displacement prediction result does not meet a first preset condition, adjusting network parameters of the displacement directed graph neural network according to the displacement prediction result, and performing iterative training by using the adjusted displacement directed graph neural network until a new displacement prediction result output by the displacement directed graph neural network meets the first preset condition, so as to obtain the initial displacement sub-model.
5. The method of claim 1, wherein training the stress directed graph neural network using the cell training samples to obtain the initial stress sub-model comprises: Constructing the stress directed graph neural network according to the node connection relation of the finite element model, wherein graph vertices of the stress directed graph neural network are units, and connection among the units is directed edges; Mapping stress input in the unit training sample to a hidden space of the stress directed graph neural network to obtain a second hidden representation; encoding second position information of the units into the second hidden representation, forming a second feature representation, and extracting deep unit features of the second feature representation using a transducer encoder; constructing a second output head through the deep unit characteristics, and outputting a stress prediction result according to the second output head; And under the condition that the stress prediction result does not meet a second preset condition, adjusting network parameters of the stress directed graph neural network according to the stress prediction result, and performing iterative training by using the adjusted stress directed graph neural network until a new stress prediction result output by the stress directed graph neural network meets the second preset condition, so as to obtain the initial stress sub-model.
6. The method of claim 1, wherein the training and optimizing the initial displacement sub-model and the initial stress sub-model by applying physical boundary constraints to the initial displacement sub-model and the initial stress sub-model and by constructing a loss function to obtain a target displacement sub-model and a target stress sub-model comprises: constructing a boundary constraint matrix, and correcting the predicted displacement output of the initial displacement sub-model by using the boundary constraint matrix to obtain a target displacement output; correcting the predicted stress output of the initial stress sub-model by using the boundary constraint matrix to obtain target stress output; defining a displacement loss function of the initial displacement sub-model according to the actual displacement output and the target displacement output in the node training sample, and defining a stress loss function of the initial stress sub-model according to the actual stress output and the target stress output in the unit training sample; weighting the displacement loss function and the stress loss function based on regularization to obtain a total loss function; and iteratively updating model parameters of the initial displacement sub-model and the initial stress sub-model through a preset parameter updating rule until the total loss function converges to obtain the target displacement sub-model and the target stress sub-model.
7. The method of claim 1, wherein the predicting rock burst using the target displacement sub-model and the target stress sub-model comprises: Constructing a target prediction model through the target displacement sub-model and the target stress sub-model; Inputting node characteristics and unit characteristics of a scene to be predicted into the target prediction model to obtain a main stress field, a displacement field, a plastic region, an unbalanced force field and a point safety degree field which are output by the prediction model, wherein the displacement field is output by the target displacement sub-model, and the main stress field is output by the target stress sub-model; Respectively constructing risk indexes of the main stress field, the displacement field, the plastic region, the unbalanced force field and the point safety degree field, and carrying out weighted calculation on each risk index to obtain a comprehensive rock burst risk index; and determining a dangerous threshold according to the comprehensive rock burst danger index and the actual working condition, and predicting the occurrence area and the danger level of the rock burst according to the dangerous threshold.
8. The rock burst simulation prediction device based on the graph transducer is characterized by comprising the following components: The construction module is used for constructing a mapping relation between input features and output features of each sub-model according to a finite element model in rock burst simulation, wherein the sub-models comprise an initial displacement sub-model and an initial stress sub-model, and the sub-models are obtained by combining a directed graph neural network and a transducer model; The generation module is used for respectively generating a node training sample and a unit training sample through finite element calculation and the mapping relation; The training module is used for training the displacement directed graph neural network by using the node training sample to obtain the initial displacement sub-model, and training the stress directed graph neural network by using the unit training sample to obtain the initial stress sub-model; The optimization module is used for performing training optimization on the initial displacement sub-model and the initial stress sub-model by applying physical boundary constraint on the initial displacement sub-model and the initial stress sub-model and constructing a loss function to obtain a target displacement sub-model and a target stress sub-model; and the prediction module is used for predicting rock burst by using the target displacement submodel and the target stress submodel.
9. An electronic device comprising a memory, a processor, a communication interface and a communication bus, said memory storing a computer program executable on said processor, said memory, said processor communicating with said communication interface via said communication bus, characterized in that said processor, when executing said computer program, implements the steps of the method according to any of the preceding claims 1 to 7.
10. A computer readable medium having non-volatile program code executable by a processor, the program code causing the processor to perform the method of any one of claims 1 to 7.

Description

Rock burst simulation prediction method and device based on graph Transformer Technical Field The application relates to the technical field of coal mine safety, in particular to a rock burst simulation prediction method and device based on a graph Transformer. Background Rock burst is a common dynamic disaster in deep mining of coal mines, and the occurrence mechanism of the rock burst relates to surrounding rock stress concentration, plastic region expansion, energy accumulation, instantaneous release and other multi-field coupling processes. The existing rock burst prediction method mainly depends on numerical simulation, experience criteria and on-site monitoring means, wherein finite element numerical simulation can provide high-precision stress field and displacement field calculation results, but the calculation amount is large, the solving time is long, and the requirement of on-site rapid early warning is difficult to meet. Meanwhile, the existing prediction method based on machine learning generally depends on a small amount of statistical features extracted manually, so that the three-dimensional space structure and multi-field coupling effect of the coal and rock mass are difficult to truly reflect, and the accuracy of a prediction result is limited. In addition, because the coal seam structure is complicated, the uncertainty of material parameters is strong, traditional model is difficult to realize quick calculation under the prerequisite of guaranteeing the precision. As can be seen, rock burst is currently not predicted quickly and accurately. In view of the above problems, no effective solution has been proposed at present. Disclosure of Invention The application provides a rock burst simulation prediction method, device, equipment and medium based on a graph Transformer, which are used for solving the technical problem that rock burst cannot be rapidly and accurately predicted. According to one aspect of the embodiment of the application, the application provides a graph-Transformer-based rock burst simulation prediction method, which comprises the steps of constructing a mapping relation between input features and output features of each sub-model according to a finite element model in rock burst simulation, wherein the sub-models comprise initial displacement sub-models and initial stress sub-models, the sub-models are obtained by combining a directed graph neural network and a Transformer model, node training samples and unit training samples are respectively generated through finite element calculation and the mapping relation, the node training samples are used for training a displacement directed graph neural network to obtain an initial displacement sub-model, the unit training samples are used for training a stress directed graph neural network to obtain an initial stress sub-model, physical boundary constraint is applied to the initial displacement sub-model and the initial stress sub-model, training optimization is carried out on the initial displacement sub-model and the initial stress sub-model through construction loss functions to obtain a target displacement sub-model and a target stress sub-model, and the target displacement sub-model and the target stress sub-model are used for rock burst prediction. The method comprises the steps of extracting node input features of finite element nodes and unit input features of finite element units according to a finite element model, wherein the node input features and the unit input features are 9-dimensional vectors, mapping node output tensors corresponding to the node input features to obtain node mapping relations, and mapping unit output tensors corresponding to the unit input features to obtain unit mapping relations, wherein the node mapping relations represent input and output structures of initial displacement submodels, and the unit mapping relations represent input and output structures of initial stress submodels. The method comprises the steps of generating a node training sample and a unit training sample through finite element calculation and a mapping relation respectively, generating a plurality of groups of calculation samples by using random input information, carrying out finite element calculation on the calculation samples to obtain original node data matched with the node mapping relation and original unit data matched with the unit mapping relation, carrying out standardization processing on the original node data to obtain the node training sample, and carrying out standardization processing on the original unit data to obtain the unit training sample. Optionally, training a displacement directed graph neural network by using a node training sample to obtain an initial displacement sub-model, wherein the initial displacement sub-model comprises the steps of constructing the displacement directed graph neural network according to the node connection relation of the finite element model, wherein graph vertices o