CN117094235B - Ground pressure disaster prediction method and device based on multitasking learning - Google Patents

Ground pressure disaster prediction method and device based on multitasking learning Download PDF

Info

Publication number
CN117094235B
CN117094235B CN202311361552.0A CN202311361552A CN117094235B CN 117094235 B CN117094235 B CN 117094235B CN 202311361552 A CN202311361552 A CN 202311361552A CN 117094235 B CN117094235 B CN 117094235B
Authority
CN
China
Prior art keywords
input
data
neural network
gate
task
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202311361552.0A
Other languages
Chinese (zh)
Other versions
CN117094235A (en
Inventor
李杨
夏文浩
宋卫东
付建新
曹帅
谭玉叶
汪杰
张理
马军
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Science and Technology Beijing USTB
Original Assignee
University of Science and Technology Beijing USTB
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Science and Technology Beijing USTB filed Critical University of Science and Technology Beijing USTB
Priority to CN202311361552.0A priority Critical patent/CN117094235B/en
Publication of CN117094235A publication Critical patent/CN117094235A/en
Application granted granted Critical
Publication of CN117094235B publication Critical patent/CN117094235B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28—Processing seismic data, e.g. for interpretation or for event detection
    • G01V1/30—Analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G06F18/15—Statistical pre-processing, e.g. techniques for normalisation or restoring missing data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Remote Sensing (AREA)
  • Probability & Statistics with Applications (AREA)
  • Medical Informatics (AREA)
  • Geometry (AREA)
  • Computer Hardware Design (AREA)
  • Acoustics & Sound (AREA)
  • Environmental & Geological Engineering (AREA)
  • Geology (AREA)
  • General Life Sciences & Earth Sciences (AREA)
  • Geophysics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a ground pressure disaster prediction method and device based on multi-task learning, which are applied to the technical field of data analysis and prediction and comprise the following steps: extracting microseism time sequence data from parameter data of a microseism focus; preprocessing microseismic time sequence data to obtain an input vector, wherein the preprocessing comprises data resampling, missing data processing and data normalization; and inputting the input vector into a pre-trained ground pressure disaster prediction model to obtain the predicted event number, the accumulated energy, the accumulated apparent volume and the energy index, and performing ground pressure disaster prediction. The method can remarkably improve the efficiency and the precision of ground pressure disaster prediction, and can also reserve the real microseismic data characteristics.

Description

Ground pressure disaster prediction method and device based on multitasking learning
Technical Field
The invention relates to the technical field of data analysis and prediction, in particular to a ground pressure disaster prediction method and device based on multi-task learning.
Background
Along with the gradual exhaustion of shallow mineral resources, the mineral resource development in China gradually enters a deep mining stage, and along with the increase of mining depth, mining disturbance is easier to destroy the original stress balance of rock mass, and rock burst, roof fall, collapse and other ground pressure disasters are induced. The sudden, random and dangerous nature of the ground pressure disasters not only greatly limit the construction efficiency, but also seriously threaten the safety of workers and equipment.
The microseismic monitoring is used as a non-contact, real-time and dynamic space three-dimensional monitoring means, has the advantages of wide monitoring range and large information quantity, and is widely applied to ground pressure disaster monitoring of underground projects such as metal nonmetallic mines, tunnels and the like, but the current ground pressure disaster prediction and early warning method based on microseismic monitoring data does not or less consider time factors, is mostly analyzed afterwards, has low early warning timeliness and poor advanced analysis capability, and influences the working efficiency and effect of ground pressure disaster prevention to a certain extent; in addition, the microseismic parameters before the occurrence of the ground pressure disasters are obviously deviated from the normal period, the ground pressure disasters are predicted and early-warned only by the current data, and the timeliness and the accuracy of the ground pressure disasters are greatly reduced. Therefore, by accurately predicting the change of the microseismic time sequence data, the time node of the microseismic event is moved forward, so that more sufficient time guarantee can be provided for the ground pressure disaster prevention and control work.
The microseismic time sequence data is dynamic and nonlinear time sequence data, is comprehensively influenced by various factors around a stope, is difficult to analyze and master in change rule, and is often related to each other in change of a plurality of types of microseismic time sequence data, for example, when the cumulative apparent volume is increased sharply, the number of microseismic events is increased sharply, and the energy index and the schmitt number are reduced rapidly. However, in the current research, only a certain index is considered separately and predicted separately, each index is split, the internal relation between the indexes cannot be reflected, and the complex relation between the indexes is difficult to discover. The current microseismic time sequence data prediction method can be roughly divided into three categories, namely classical time sequence modeling (ARMA, ARIMA and the like), deep learning (RNN, GRU, LSTM and the like) and machine learning (LightGBM, XGBoost, random Forest and the like). The main defects existing in the current ground pressure disaster prediction when facing a plurality of microseismic time sequence data prediction tasks are as follows: (1) The method has low efficiency, a plurality of models are needed to predict a plurality of microseismic time sequence data, when facing to a larger microseismic time sequence data set, the models have more parameters, and the model prediction time, the calculation cost, the storage cost and the maintenance cost are higher. (2) The effect is poor, and because partial microseismic time sequence data are sparse or unbalanced, for example, a plurality of unchanged or little-changed data exist in the accumulated view volume time sequence data set, the data characteristic extraction is affected, and the model overfitting is serious. (3) The prediction result deviates from the real microseismic data characteristics and is affected by factors such as rock drilling operation, blasting operation, mechanical vibration, electromagnetic interference and the like, some noise inevitably exists in microseismic time sequence data, the prediction value often inherits the noise characteristics in the original data, the real microseismic signal characteristics cannot be reflected, and therefore the accuracy and the reliability of subsequent data analysis are affected.
Disclosure of Invention
The embodiment of the invention provides a ground pressure disaster prediction method and device based on multi-task learning. The method and the device mainly aim at the problems that the ground pressure disaster prediction efficiency is low, the effect is poor, and the prediction result deviates from the real microseismic data characteristics, so that the ground pressure disaster prediction efficiency and accuracy can be improved, and the real microseismic data characteristics can be reserved. The technical scheme is as follows:
on one hand, the embodiment of the application provides a ground pressure disaster prediction method based on multi-task learning, which comprises the following steps:
s1: extracting microseism time sequence data from parameter data of a microseism focus;
s2: preprocessing microseismic time sequence data to obtain an input vector, wherein the preprocessing comprises data resampling, missing data processing and data normalization;
s3: inputting the input vector into a pre-trained ground pressure disaster prediction model to obtain the predicted event number, accumulated energy, accumulated apparent volume and energy index, carrying out ground pressure disaster prediction,
the ground pressure disaster prediction model is a multi-task learning model based on an Attention-LSTM-MTL neural network, the multi-task learning model comprises a two-layer stacked LSTM network, an Attention module and four full-connection layers, and the two-layer stacked LSTM network comprises a forgetting gate, an input gate and an output gate.
Preferably, the extracting the microseismic time sequence data from the parameter data of the microseismic source in S1 includes:
s11: acquiring microseism focus parameter data through a microseism monitoring system;
s12: microseismic timing data is extracted from microseismic source parameter data.
Preferably, the microseismic time sequence data comprises time sequence data of event number, accumulated energy, accumulated apparent volume and energy index.
Preferably, the preprocessing the microseismic time sequence data in S2 to obtain an input vector includes:
s21: resampling the microseismic time sequence data according to the formula (1) to obtain a fixed frequencyTime series data of (2)：
（1）
In the method, in the process of the invention,for mean aggregation function, +.>Is a fixed frequency +.>Time series data of->Is microseismic time sequence data;
s22: processing the missing data of the fixed frequency time sequence data to obtain complete time sequence data；
S23: carrying out data normalization processing on the complete time sequence data x_w through a formula (2) to obtain an input vector:
（2）
in the method, in the process of the invention,for inputting vectors, ++>For complete time series data->Minimum value of->For complete time series data->Is the maximum value of (a).
Preferably, the step S22 of performing missing data processing on missing data of the fixed frequency time series data to obtain complete time series data includes:
S221: selecting time sequence data with fixed frequencyData with missing value in the data to obtain missing data +.>；
S222: will miss dataSetting zero to obtain the complete timeSequence data->。
Preferably, inputting the input vector into a pre-trained ground pressure disaster prediction model at S3, to obtain a predicted event number, accumulated energy, accumulated apparent volume and energy index, including:
s31: input at time t in input vectorAnd the last time two-layer stacked LSTM network outputSplicing to obtain spliced vector ∈>；
S32: inputting the spliced vector into a forgetting gate, and mapping the vector between (0, 1) through a sigmoid neural network layer of the forgetting gate based on a formula (3) to obtain an output value of the forgetting gate:
（3）
in the method, in the process of the invention,output value for forgetting gate +.>For the input at time t in the input vector, < >>For the two-layer stacked LSTM network output of the last moment, < >>Weights of sigmoid neural network layer for amnestic gate, < ->For all that isDeviation of sigmoid neural network layer of forgetting gate;
s33: the spliced vector is input to an input gate for the first time, mapped to a (0, 1) interval through a sigmoid neural network layer according to a formula (4) to obtain a first output value of the input gate, mapped to a (-1, 1) interval through a tanh neural network layer to generate a state candidate vector:
（4）
In the method, in the process of the invention,for the first output value of the input gate, +.>For the input at time t in the input vector, < >>For the last moment LSTM network output, < + >>Is a state candidate vector, +.>Weights of sigmoid neural network layer for input gates, +.>Deviation of sigmoid neural network layer for input gate, +.>Weights of tanh neural network layer as input gates, +.>Deviations of the tanh neural network layer that are input gates;
s34: based on the forgetting gate output value, updating the state candidate through a formula (5) to obtain an updated state candidate:
（5）
in the method, in the process of the invention,is a state candidate vector, +.>For updated state candidates +.>Information transmitted for the time t-1 of selective forgetting,/>Information to be newly added;
s35: the spliced vector is input and output for the second time, mapped to a (0, 1) interval through a sigmoid neural network layer according to a formula (6), and integrated with the updated state candidate quantity to obtain the input gate second output value at the moment t:
（6）
in the method, in the process of the invention,for the second output value of the input gate, +.>For the input at time t in the input vector, < >>For the last moment LSTM network output, < + >>Weighting of sigmoid neural network layer for updated input gate +. >To input sigmoid in gateDeviation of the neural network layer;
s36: inputting the second output value of the input gate into an Attention module, and distributing corresponding weights according to the time step importance difference according to a formula (7) to obtain the Attention score of the second output value of the input gate:
（7）
in the method, in the process of the invention,for attention score, ++>Weight of full connection layer, +.>Deviation for full connection layer, +.>For LSTM network output, tanh is full connection layer activation function;
s37: based on the formula (8), the attention score is input, and the attention weight corresponding to the second output value of the input door at the moment t is obtained through calculation of a softmax function:
（8）
in the method, in the process of the invention,for attention score, ++>Is the attention weight;
s38: based on a formula (9), the Attention weight is input to carry out weighted summation on the second output value of the input gate at the moment t to obtain the output value of the Attention module:
（9）
in the method, in the process of the invention,for attention weight, ++>For LSTM network output, < >>The output value of the Attention module;
s39: inputting the output value of the Attention module into four fully connected layers to output a predicted value, wherein the predicted value comprises: the four full connection layers correspond to four tasks.
Preferably, before the step S3 of inputting the input vector into a pre-trained ground pressure disaster prediction model to obtain the predicted event number, the accumulated energy, the accumulated apparent volume and the energy index, the method further includes:
s0: training an initial multi-task learning model based on an Attention-LSTM-MTL neural network through training data to obtain a pre-trained ground pressure disaster prediction model;
the training of the initial multi-task learning model based on the Attention-LSTM-MTL neural network through training data to obtain a pre-trained ground pressure disaster prediction model comprises the following steps:
s01: collecting training data, wherein the training data comprises sample characteristics and sample labels of four tasks extracted from microseismic time sequence data;
s02: an initial multi-task learning model based on an attribute-LSTM-MTL neural network for predicting an earth pressure disaster is constructed, wherein the initial multi-task learning model based on the attribute-LSTM-MTL neural network comprises the following steps: the system comprises an input layer, a sharing layer, a linear layer and an output layer, wherein the input layer comprises four input vectors, and the four tasks are event number prediction, accumulated energy prediction, accumulated apparent volume prediction and energy index prediction;
S03: initializing an initial multi-task learning model based on an attribute-LSTM-MTL neural network for predicting an earth pressure disaster, setting super parameters, and randomly initializing weights of all layers;
s04: sample characteristics of a training set and sample labels of four corresponding tasks are input into an initial multi-task learning model based on an Attention-LSTM-MTL neural network, and the model is trained through dynamic weights;
s05: in the training process, the MSE loss function is adopted to respectively calculate the loss value of each task, and the four loss function values are weighted and summed based on formulas (10) and (11), so as to obtain the total loss value:
（10）
（11）
in the method, in the process of the invention,loss value representing the ith task, +.>Indicating total loss value, ++>Representing the actual value of the ith task at the jth moment,/->Indicating the predicted value of the ith task at the jth moment, n being the number of samples, +.>A weight representing an ith task;
s06: based on the formula (12), weighting the loss values of each task in a dynamic weight mode by utilizing the total loss value, and dynamically adjusting according to the training progress of the task to obtain a predicted value:
（12）
in the method, in the process of the invention,weight value representing the ith task to be put into the next m rounds of training, +. >Loss value for each task calculated for this round, < > j->The total loss value obtained by the calculation of the round is calculated;
s07: and after the predicted value is inversely normalized, judging the performance of the model through the evaluation index, and if the performance of the model meets the preset prediction precision requirement, storing and outputting model parameters to obtain a pre-trained ground pressure disaster prediction model.
In a second aspect, an embodiment of the present application provides a device for predicting an earth pressure disaster based on multitasking learning, including the following steps:
data unit: extracting microseism time sequence data from parameter data of a microseism focus;
pretreatment unit: the method comprises the steps of preprocessing microseismic time sequence data to obtain an input vector, wherein the preprocessing comprises data resampling, missing data processing and data normalization;
prediction unit: the ground pressure disaster prediction model is a multi-task learning model based on an Attention-LSTM-MTL neural network, the multi-task learning model comprises two layers of stacked LSTM networks, an Attention module and four fully connected layers, and the two layers of stacked LSTM networks comprise a forgetting gate, an input gate and an output gate.
In a third aspect, an embodiment of the present application provides an electronic device, including: the device comprises a shell, a processor, a memory, a circuit board and a power circuit, wherein the circuit board is arranged in a space surrounded by the shell, and the processor and the memory are arranged on the circuit board; a power supply circuit for supplying power to each circuit or device of the electronic apparatus; the memory is used for storing executable program codes; the processor executes a program corresponding to the executable program code by reading the executable program code stored in the memory for performing the method as described above.
In a fourth aspect, embodiments of the present application provide a computer readable storage medium, where one or more programs are stored, where the one or more programs are executable by one or more processors to implement a method as described above.
The technical scheme provided by the embodiment of the invention has the beneficial effects that at least: according to the scheme, a single model is adopted to conduct joint prediction of a plurality of microseismic time sequence data, so that the calculation efficiency of the ground pressure disaster prediction model is improved, and the calculation cost and the memory occupation are reduced. And learning the microseismic time sequence data set with smaller fluctuation together with other data sets, supplementing and restraining the shared information mutually, relieving the overfitting of the model, and improving the generalization capability of the model. And the method adopts a multi-task learning method to predict a plurality of microseismic time sequence data, effectively explores the coupling information among the plurality of microseismic time sequence data, and can obtain useful information and feedback from other tasks by each task so as to further improve the prediction precision of each microseismic time sequence data prediction task. Meanwhile, as different task noises tend to different directions, the multi-task learning counteracts part of the noises to a certain extent, plays a role in enhancing implicit data, and improves the model prediction effect and robustness. The multi-task learning process is trained through the dynamic weights, so that the model training process is more flexible, the problem that partial tasks are fitted and partial tasks are still not fitted is avoided, and each task can achieve a good prediction effect.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flowchart of a method for predicting an earth pressure disaster based on multi-task learning according to an embodiment of the present invention;
FIG. 2 is a schematic structural diagram of a multi-task learning model based on an Attention-LSTM-MTL neural network according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram of an LSTM unit according to an embodiment of the present invention;
FIG. 4 is a training flow chart of an earth pressure disaster prediction model based on multi-task learning provided by the embodiment of the invention;
FIG. 5 is a schematic diagram of steps for predicting microseismic time sequence data based on multi-task learning according to an embodiment of the present invention;
fig. 6 is a schematic diagram of a device for predicting an earth pressure disaster based on multi-task learning according to an embodiment of the present invention.
Fig. 7 is a schematic structural diagram of an electronic device 700 according to an embodiment of the present invention.
Detailed Description
In order to make the technical problems, technical solutions and advantages to be solved more apparent, the following detailed description will be given with reference to the accompanying drawings and specific embodiments.
The embodiment of the invention provides a ground pressure disaster prediction method based on multi-task learning, which can be realized by electronic equipment, wherein the electronic equipment can be a terminal or a server. As shown in fig. 1, a flow chart of a method for predicting an earth pressure disaster based on multi-task learning, the process flow of the method may include the following steps:
s1: extracting microseism time sequence data from parameter data of a microseism focus;
preferably, the S1 includes:
s11: acquiring microseism focus parameter data through a microseism monitoring system;
s12: microseismic timing data is extracted from microseismic source parameter data.
Preferably, the microseismic time sequence data comprises time sequence data of event number, accumulated energy, accumulated apparent volume and energy index;
in some embodiments, the mine microseism monitoring system is relied on to obtain microseism focus parameter data, and three-dimensional positioning coordinates (X, Y, Z) containing microseism events and occurrence time are constructedSensor trigger number,/->Wave energy and->Wave energy ratioEvent energy- >Moment of earthquake->Dynamic pressure drop->Static pressure drop->Corner frequency->Magnitude of vibrationVisual stress->Visual volume->Energy index->And a microseismic source parameter database of various parameters.
It should be noted that, the recommended standard of the energy industry in China, namely the technical specification of rock burst risk assessment of water and electricity engineering (NB/T10143-2019), proposes that the early warning indexes of the microseismic monitoring information are preferably the accumulated microseismic event number, the accumulated microseismic release energy, the accumulated microseismic apparent volume, the microseismic event rate, the microseismic release energy rate and the microseismic apparent volume rate; national recommended national Standard "rock burst determination, monitoring and control method part 4: the microseism monitoring method (GB/T25217.4-2019) indicates that the frequency of microseism and the total energy of the microseism are used as main judging indexes, and the maximum value of the energy of the microseism is used as an auxiliary judging index. Therefore, it is preferable to select a plurality of real-time or accumulated values of the event number, the energy, the apparent volume and the energy index for joint prediction, and finally select the event number, the accumulated energy, the accumulated apparent volume and the energy index, but the actual method is not limited to the four microseismic time sequence data for joint prediction, and can select a plurality of microseismic time sequence data with practical significance for joint prediction according to the characteristics of the mine.
S2: preprocessing microseismic time sequence data to obtain an input vector, wherein the preprocessing comprises data resampling, missing data processing and data normalization;
preferably, the S2 includes:
s21: resampling the microseismic time sequence data according to the formula (1) to obtain a fixed frequencyTime series data of (2)：
（1）
In the method, in the process of the invention,is mean value aggregationFunction of (A)/(B)>Is a fixed frequency +.>Time series data of->Is microseismic time sequence data;
it should be noted that, because the occurrence time of the microseismic event is usually uncertain, in order to better perform the prediction model construction, the irregular microseismic time sequence data needs to be resampled to a fixed frequency, and the daily frequency is selected in the application, if the microseismic event occurs more frequently, the daily frequency can also be resampled to the hourly frequency.
S22: carrying out missing data processing on missing data of time sequence data with fixed frequency to obtain complete time sequence data;
preferably, the S22 includes:
s221: selecting time sequence data with fixed frequencyData with missing value in the data to obtain missing data +.>；
S222: will miss dataSetting zero to obtain complete time sequence data +.>；
It should be noted that, immediately before a partial pressure disaster (such as time-lapse rock burst) occurs, the microseismic events are fewer, the apparent volume and energy index change is not obvious, and an obvious "quiet period" exists, so that partial microseismic time sequence data is lost. If linear, average, forward and backward interpolation methods are adopted, the microseismic signals of a 'quiet period' are subjected to blurring or smoothing treatment, key microseismic information is covered, so that the activity change before the occurrence of an earth pressure disaster cannot be accurately reflected, the method of directly zeroing missing data is adopted for treatment, no additional estimation or interpolation process is introduced, the original state of microseismic monitoring data is reserved, and the actual situation of the earthquake activity is more accurately reflected;
S23: carrying out data normalization processing on the complete time sequence data through a formula (2) to obtain an input vector:
（2）
in the method, in the process of the invention,for inputting vectors, ++>Time series data for fixed frequency->Minimum value of->Time series data for fixed frequency->Is the maximum value of (a).
It should be noted that, the sources and measurement units of the four microseismic time sequence data (the number of events, the accumulated energy, the accumulated apparent volume and the energy index) are different, the orders of magnitude are distributed between 100-107, the distribution range is greatly different, and the model is more sensitive to the characteristics with larger values, and the characteristics with smaller values are easily covered, so that the maximum and minimum normalization method is adopted for data normalization, the orders of magnitude difference is eliminated, and all the characteristics have similar importance in the training process.
S3: the input vector is input into a pre-trained ground pressure disaster prediction model to obtain the predicted event number, the accumulated energy, the accumulated apparent volume and the energy index, and ground pressure disaster prediction is carried out, as shown in fig. 2, the ground pressure disaster prediction model is a multi-task learning model based on an Attention-LSTM-MTL neural network, the multi-task learning model comprises two layers of stacked LSTM networks (shown in fig. 3), an Attention module and four fully connected layers, and the two layers of stacked LSTM networks comprise a forgetting gate, an input gate and an output gate.
Preferably, the multi-task learning model based on the Attention-LSTM-MTL neural network comprises a sharing module of the LSTM module and the Attention module, and parameters (weight and deviation) of the sharing module are shared among 4 micro-seismic time sequence data prediction subtasks. The LSTM module realizes the function of memorizing long-term information by introducing the concepts of door mechanism and cell unit, and avoids gradient explosion or gradient disappearance, and the structure of the LSTM unit is shown in figure 3.
Preferably, the S3 includes:
s31: input at time t in input vectorAnd the last time two-layer stacked LSTM network outputSplicing to obtain spliced vector +.>；
S32: inputting the spliced vector into a forgetting gate, and mapping the vector between (0, 1) through a sigmoid neural network layer of the forgetting gate based on a formula (3) to obtain an output value of the forgetting gate:
（3）
in the method, in the process of the invention,output value for forgetting gate +.>For the input at time t in the input vector, < >>For the two-layer stacked LSTM network output of the last moment, < >>Weights of sigmoid neural network layer for amnestic gate, < ->Deviation of sigmoid neural network layer of forgetting gate;
note that, forget the gate output value A value close to 0 means that the information should be discarded and a value close to 1 means that the information should be retained.
S33: the spliced vector is input to an input gate for the first time, mapped to a (0, 1) interval through a sigmoid neural network layer according to a formula (4) to obtain a first output value of the input gate, mapped to a (-1, 1) interval through a tanh neural network layer to generate a state candidate vector:
（4）
in the method, in the process of the invention,for the first output value of the input gate, +.>For the input at time t in the input vector, < >>For the last moment LSTM network output, < + >>Is a state candidate vector, +.>Weights of sigmoid neural network layer for input gates, +.>Deviation of sigmoid neural network layer for input gate, +.>Weights of tanh neural network layers respectively input gates, +.>Deviations of the tanh neural network layer that are input gates;
s34: based on the forgetting gate output value, updating the state candidate through a formula (5) to obtain an updated state candidate:
（5）
in the method, in the process of the invention,is a state candidate vector, +.>For updated state candidates +.>Information transmitted for the time t-1 of selective forgetting,/>Information to be newly added;
s35: the spliced vector is input and output for the second time, mapped to a (0, 1) interval through a sigmoid neural network layer according to a formula (6), and integrated with the updated state candidate quantity to obtain the input gate second output value at the moment t:
（6）
In the method, in the process of the invention,for the second output value of the input gate, +.>For the input at time t in the input vector, < >>For the last moment LSTM network output, < + >>Weighting of sigmoid neural network layer for updated input gate +.>A bias of a sigmoid neural network layer in an input gate;
s36: inputting the second output value of the input gate into an Attention module, and distributing corresponding weights according to the time step importance difference according to a formula (7) to obtain the Attention score of the second output value of the input gate:
（7）
in the method, in the process of the invention,for attention score, ++>Weight of full connection layer, +.>Deviation for full connection layer, +.>For LSTM network output, tanh is full connection layer activation function;
s37: based on the formula (8), the attention score is input, and the attention weight corresponding to the second output value of the input door at the moment t is obtained through calculation of a softmax function:
（8）
in the method, in the process of the invention,for attention score, ++>Is the attention weight;
s38: based on a formula (9), the Attention weight is input to carry out weighted summation on the second output value of the input gate at the moment t to obtain the output value of the Attention module:
（9）
in the method, in the process of the invention,for attention weight, ++>For LSTM network output, < >>The output value of the Attention module;
S39: inputting the output value of the Attention module into four fully connected layers to output a predicted value, wherein the predicted value comprises: the four full connection layers correspond to four tasks.
Preferably, before the step S3, the method further includes:
s0: training the neural network model of the initial multitask learning through training data to obtain a pre-trained ground pressure disaster prediction model;
as shown in fig. 4, the training the neural network model for initial multi-task learning by the training data in S0 to obtain a pre-trained ground pressure disaster prediction model includes:
s01: collecting training data, wherein the training data comprises sample characteristics and sample labels of four tasks extracted from microseismic time sequence data;
s02: an initial multi-task learning model based on an attribute-LSTM-MTL neural network for predicting an earth pressure disaster is constructed, wherein the initial multi-task learning model based on the attribute-LSTM-MTL neural network comprises the following steps: the system comprises an input layer, a sharing layer, a linear layer and an output layer, wherein the input layer comprises four input vectors, and the four tasks are event number prediction, accumulated energy prediction, accumulated apparent volume prediction and energy index prediction;
S03: initializing an initial multi-task learning model based on an attribute-LSTM-MTL neural network for predicting an earth pressure disaster, setting super parameters, and randomly initializing weights of all layers;
s04: sample characteristics of a training set and sample labels of four corresponding tasks are input into an initial multi-task learning model based on an Attention-LSTM-MTL neural network, and the model is trained through dynamic weights;
s05: in the training process, the MSE loss function is adopted to respectively calculate the loss value of each task, and the four loss function values are weighted and summed based on formulas (10) and (11), so as to obtain the total loss value:
（10）
（11）
in the method, in the process of the invention,loss value representing the ith task, +.>Indicating total loss value, ++>Representing the actual value of the ith task at the jth moment,/->Indicating the predicted value of the ith task at the jth moment, n being the number of samples, +.>A weight representing an ith task;
s06: based on the formula (12), weighting the loss values of each task in a dynamic weight mode by utilizing the total loss value, and dynamically adjusting according to the training progress of the task to obtain a predicted value:
（12）
in the method, in the process of the invention,weight value representing the ith task to be put into the next m rounds of training, +. >Loss value for each task calculated for this round, < > j->The total loss value obtained by the calculation of the round is calculated;
it should be noted that m may be adjusted according to the actual training effect, but m may not be too small or too large, where too small may cause frequent variation of the training loss function, resulting in unstable model training process, and too large may cause no change in the weight for a long time, so that the dynamic weight may not have a good effect.
The dynamic weight enables the training process to be more flexible, the problem that a part of tasks are fitted and a part of tasks are still not fitted is avoided, meanwhile, contributions of different tasks are balanced, and therefore good prediction effects can be achieved for all the tasks.
S07: and after the predicted value is inversely normalized, judging the performance of the model through the evaluation index, and if the performance of the model meets the preset prediction precision requirement, storing and outputting model parameters to obtain a pre-trained ground pressure disaster prediction model.
In some embodiments, the performance of the model is judged by an evaluation index MAE, RMSE, MAPE, R or the like.
In some embodiments, the first 80% of each microseismic time sequence data is first divided into a training set train, and the remaining 20% is divided into a test set test, so as to prevent the training set data from leaking into the test set when the sample features and the labels are segmented by adopting a sliding time window, and the performance evaluation of the model on the test set is caused to deviate.
It should be noted that, the sliding time window method is adopted to segment the features and the labels of the training set and the testing set, firstly, the sliding time window is determined, and 7 sliding time windows are taken, but the setting of the sliding time windows is not limited, and 3 days, 4 days or 5 days and the like can also be adopted in the embodiment of the application; the model adopts single-step prediction, namely, predicting data on day 8 according to microseismic time sequence data on the previous 7 days, as shown in fig. 5; and selecting samples from the sliding window in the microseismic time series data according to the set time window size, and dividing the samples into characteristic and label parts.
It should be further noted that, the event number prediction, the accumulated energy prediction, the accumulated view volume prediction and the energy index prediction are regarded as a task respectively, a multi-task learning model based on an Attention-LSTM-MTL neural network is constructed, the characteristics of various microseismic time sequence data are rapidly extracted through a two-layer stacked LSTM network, dropout is adopted between layers to prevent overfitting, an activation function is sigmoid, an Attention module is introduced after the LSTM network, the sharing layer formed by the LSTM module and the Attention module realizes the differential selection of different subtasks on the sharing characteristics, and finally the data is output through a Dense layer (full connection layer).
The foregoing is a description of embodiments of the method, and the following further describes embodiments of the device.
As shown in fig. 6, an embodiment of the present application provides a device for predicting an earth pressure disaster based on multitasking learning, including the following steps:
data unit 601: extracting microseism time sequence data from parameter data of a microseism focus;
preprocessing unit 602: the method comprises the steps of preprocessing microseismic time sequence data to obtain an input vector, wherein the preprocessing comprises data resampling, missing data processing and data normalization;
prediction unit 603: the ground pressure disaster prediction model is a multi-task learning model based on an Attention-LSTM-MTL neural network, the multi-task learning model comprises two layers of stacked LSTM networks, an Attention module and four fully connected layers, and the two layers of stacked LSTM networks comprise a forgetting gate, an input gate and an output gate.
An electronic device for ground pressure disaster prediction based on multitasking learning, the electronic device comprising: the device comprises a shell, a processor, a memory, a circuit board and a power circuit, wherein the circuit board is arranged in a space surrounded by the shell, and the processor and the memory are arranged on the circuit board; a power supply circuit for supplying power to each circuit or device of the electronic apparatus; the memory is used for storing executable program codes; the processor executes a program corresponding to the executable program code by reading the executable program code stored in the memory for performing the method described above.
A computer readable storage medium storing one or more programs executable by one or more processors to implement the method described above.
Aiming at the problems of low ground pressure disaster prediction efficiency, poor effect and deviation of prediction results from real microseismic data characteristics, the invention provides a multi-task learning model based on an Attention-LSTM-MTL neural network for ground pressure disaster prediction. The model considers the coupling information among a plurality of microseismic time sequence data, and can train a multi-task learning process through dynamic weights, so that the model training process is more flexible, and the problem that partial tasks are fitted and partial tasks are still not fitted is avoided. And the calculation efficiency and accuracy of the ground pressure disaster prediction model are comprehensively improved.
Fig. 7 is a schematic structural diagram of an electronic device 700 according to an embodiment of the present invention, where the electronic device 600 may have a relatively large difference due to different configurations or performances, and may include one or more processors (central processing units, CPU) 701 and one or more memories 702, where at least one instruction is stored in the memories 702, and the at least one instruction is loaded and executed by the processors 701 to implement the steps of the above-mentioned chinese text spell checking method.
In an exemplary embodiment, a computer readable storage medium, such as a memory including instructions executable by a processor in a terminal to perform the above-described chinese text spell checking method, is also provided. For example, the computer readable storage medium may be ROM, random Access Memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.
It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program for instructing relevant hardware, where the program may be stored in a computer readable storage medium, and the storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.
The following points need to be described:
(1) The drawings of the embodiments of the present invention relate only to the structures related to the embodiments of the present invention, and other structures may refer to the general designs.
(2) In the drawings for describing embodiments of the present invention, the thickness of layers or regions is exaggerated or reduced for clarity, i.e., the drawings are not drawn to actual scale. It will be understood that when an element such as a layer, film, region or substrate is referred to as being "on" or "under" another element, it can be "directly on" or "under" the other element or intervening elements may be present.
3) The embodiments of the invention and the features of the embodiments can be combined with each other to give new embodiments without conflict.
The foregoing description of the preferred embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and scope of the invention are intended to be included within the scope of the invention.

Claims (8)

1. The ground pressure disaster prediction method based on multitask learning is characterized by comprising the following steps of:
s1: extracting microseism time sequence data from parameter data of a microseism focus;
s2: preprocessing microseismic time sequence data to obtain an input vector, wherein the preprocessing comprises data resampling, missing data processing and data normalization;
s3: inputting the input vector into a pre-trained ground pressure disaster prediction model to obtain the predicted event number, accumulated energy, accumulated apparent volume and energy index, carrying out ground pressure disaster prediction,
the ground pressure disaster prediction model is a multi-task learning model based on an Attention-LSTM-MTL neural network, the multi-task learning model comprises a two-layer stacked LSTM network, an Attention module and four full-connection layers, and the two-layer stacked LSTM network comprises a forgetting gate, an input gate and an output gate:
Before the input vector is input into the pre-trained ground pressure disaster prediction model in the step S3 to obtain the predicted event number, the accumulated energy, the accumulated apparent volume and the energy index, the method further comprises:
s0: training an initial multi-task learning model based on an Attention-LSTM-MTL neural network through training data to obtain a pre-trained ground pressure disaster prediction model;
the training of the initial multi-task learning model based on the Attention-LSTM-MTL neural network through training data to obtain a pre-trained ground pressure disaster prediction model comprises the following steps:
s01: collecting training data, wherein the training data comprises sample characteristics and sample labels of four tasks extracted from microseismic time sequence data;
s02: an initial multi-task learning model based on an attribute-LSTM-MTL neural network for predicting an earth pressure disaster is constructed, wherein the initial multi-task learning model based on the attribute-LSTM-MTL neural network comprises the following steps: the system comprises an input layer, a sharing layer, a linear layer and an output layer, wherein the input layer comprises four input vectors, and the four tasks are event number prediction, accumulated energy prediction, accumulated apparent volume prediction and energy index prediction;
S03: initializing an initial multi-task learning model based on an attribute-LSTM-MTL neural network for predicting an earth pressure disaster, setting super parameters, and randomly initializing weights of all layers;
s04: sample characteristics of a training set and sample labels of four corresponding tasks are input into an initial multi-task learning model based on an Attention-LSTM-MTL neural network, and the model is trained through dynamic weights;
s05: in the training process, the MSE loss function is adopted to respectively calculate the loss value of each task, and the four loss function values are weighted and summed based on formulas (10) and (11), so as to obtain the total loss value:
（10）
（11）
in the method, in the process of the invention,loss value representing the ith task, +.>Indicating total loss value, ++>Representing the actual value of the ith task at the jth moment,/->Indicating the predicted value of the ith task at the jth moment, n being the number of samples, +.>A weight representing an ith task;
s06: based on the formula (12), weighting the loss values of each task in a dynamic weight mode by utilizing the total loss value, and dynamically adjusting according to the training progress of the task to obtain a predicted value:
（12）
in the method, in the process of the invention,weight value representing the ith task to be put into the next m rounds of training, +. >Loss value for each task calculated for this round, < > j->The total loss value obtained by the calculation of the round is calculated;
s07: after the predicted value is inversely normalized, judging the performance of the model through an evaluation index, and if the performance of the model meets the preset prediction precision requirement, storing and outputting model parameters to obtain a pre-trained ground pressure disaster prediction model;
and S3, inputting the input vector into a pre-trained ground pressure disaster prediction model to obtain the predicted event number, accumulated energy, accumulated apparent volume and energy index, wherein the method comprises the following steps of:
s31: input at time t in input vectorAnd the two-layer stacked LSTM network output of the last moment +.>Splicing to obtain spliced vector +.>；
S32: inputting the spliced vector into a forgetting gate, and mapping the vector between (0, 1) through a sigmoid neural network layer of the forgetting gate based on a formula (3) to obtain an output value of the forgetting gate:
（3）
in the method, in the process of the invention,output value for forgetting gate +.>For the input at time t in the input vector, < >>For the two-layer stacked LSTM network output of the last moment, < >>Weights of sigmoid neural network layer for amnestic gate, < ->Deviation of sigmoid neural network layer of forgetting gate;
s33: the spliced vector is input to an input gate for the first time, mapped to a (0, 1) interval through a sigmoid neural network layer according to a formula (4) to obtain a first output value of the input gate, mapped to a (-1, 1) interval through a tanh neural network layer to generate a state candidate vector:
（4）
In the method, in the process of the invention,for the first output value of the input gate, +.>For the input at time t in the input vector, < >>As a candidate vector for the state,weights of sigmoid neural network layer for input gates, +.>Deviation of sigmoid neural network layer for input gate, +.>Weights of tanh neural network layer as input gates, +.>Deviations of the tanh neural network layer that are input gates;
s34: based on the forgetting gate output value, updating the state candidate through a formula (5) to obtain an updated state candidate:
（5）
in the method, in the process of the invention,is a state candidate vector, +.>For updated state candidates +.>Information transmitted at time t-1 for selective forgetting;
s35: the spliced vector is input and output for the second time, mapped to a (0, 1) interval through a sigmoid neural network layer according to a formula (6), and integrated with the updated state candidate quantity to obtain the input gate second output value at the moment t:
（6）
in the method, in the process of the invention,for the second output value of the input gate, +.>For the updated sigmoid neural network layer weights of the input gates,for the bias of the sigmoid neural network layer in the input gate, +.>Outputting for an LSTM network;
s36: inputting the second output value of the input gate into an Attention module, and distributing corresponding weights according to different time step importance according to a formula (7) to obtain the Attention score of the second output value of the input gate:
（7）
In the method, in the process of the invention,for attention score, ++>Weight of full connection layer, +.>The deviation of the full connection layer is given, and tanh is the full connection layer activation function;
s37: based on the formula (8), the attention score is input, and the attention weight corresponding to the second output value of the input door at the moment t is obtained through calculation of a softmax function:
（8）
in the method, in the process of the invention,is the attention weight;
s38: based on a formula (9), the Attention weight is input to carry out weighted summation on the second output value of the input gate at the moment t, so as to obtain the output value of the Attention module:
（9）
in the method, in the process of the invention,the output value of the Attention module;
s39: inputting the output value of the Attention module into four fully connected layers to output a predicted value, wherein the predicted value comprises: the four full connection layers correspond to four tasks.
2. The method for predicting ground pressure disasters based on multi-task learning according to claim 1, wherein the extracting microseism time series data from the parameter data of the microseism source in S1 comprises:
s11: acquiring parameter data of a microseism focus through a microseism monitoring system;
s12: and extracting microseism time sequence data from the parameter data of the microseism focus.
3. The method for predicting ground pressure disasters based on multi-task learning of claim 1, wherein the microseismic timing data comprises: event number, accumulated energy, accumulated apparent volume, and energy index.
4. The method for predicting ground pressure disasters based on multi-task learning according to claim 1, wherein the preprocessing of the microseismic time sequence data in S2 to obtain an input vector comprises:
s21: resampling the microseismic time sequence data according to the formula (1) to obtain a fixed frequencyTime series data of->：
（1）
In the method, in the process of the invention,for mean aggregation function, +.>Is a fixed frequency +.>Time series data of->Is microseismic time sequence data;
s22: processing the missing data of the fixed frequency time sequence data to obtain complete time sequence data；
S23: for complete time series data through formula (2)Carrying out data normalization processing to obtain an input vector:
（2）
in the method, in the process of the invention,for inputting vectors, ++>For complete time series data->Minimum value of->Is complete time sequence dataIs the maximum value of (a).
5. The method for predicting ground pressure disasters based on multi-task learning according to claim 4, wherein the step S22 of performing missing data processing on missing data of fixed-frequency time series data to obtain complete time series data includes:
S221: selecting time sequence data with fixed frequencyMemory storageIn the data of the missing value, the missing data is obtained>；
S222: will miss dataSetting zero to obtain complete time sequence data +.>。
6. An earth pressure disaster prediction device based on multitasking learning, characterized in that the device is adapted for the method according to any of the preceding claims 1-5, the device comprising:
data unit: extracting microseism time sequence data from parameter data of a microseism focus;
pretreatment unit: the method comprises the steps of preprocessing microseismic time sequence data to obtain an input vector, wherein the preprocessing comprises data resampling, missing data processing and data normalization;
prediction unit: the method is used for inputting the input vector into a pre-trained ground pressure disaster prediction model to obtain a predicted event number, accumulated energy, accumulated apparent volume and energy index, wherein the ground pressure disaster prediction model is a multi-task learning model based on an Attention-LSTM-MTL neural network, the multi-task learning model comprises two layers of stacked LSTM networks, an Attention module and four fully connected layers, the two layers of stacked LSTM networks comprise a forgetting gate, an input gate and an output gate, and the method further comprises:
S0: training an initial multi-task learning model based on an Attention-LSTM-MTL neural network through training data to obtain a pre-trained ground pressure disaster prediction model;
the training of the initial multi-task learning model based on the Attention-LSTM-MTL neural network through training data to obtain a pre-trained ground pressure disaster prediction model comprises the following steps:
s01: collecting training data, wherein the training data comprises sample characteristics and sample labels of four tasks extracted from microseismic time sequence data;
s02: an initial multi-task learning model based on an attribute-LSTM-MTL neural network for predicting an earth pressure disaster is constructed, wherein the initial multi-task learning model based on the attribute-LSTM-MTL neural network comprises the following steps: the system comprises an input layer, a sharing layer, a linear layer and an output layer, wherein the input layer comprises four input vectors, and the four tasks are event number prediction, accumulated energy prediction, accumulated apparent volume prediction and energy index prediction;
s03: initializing an initial multi-task learning model based on an attribute-LSTM-MTL neural network for predicting an earth pressure disaster, setting super parameters, and randomly initializing weights of all layers;
S04: sample characteristics of a training set and sample labels of four corresponding tasks are input into an initial multi-task learning model based on an Attention-LSTM-MTL neural network, and the model is trained through dynamic weights;
s05: in the training process, the MSE loss function is adopted to respectively calculate the loss value of each task, and the four loss function values are weighted and summed based on formulas (10) and (11), so as to obtain the total loss value:
（10）
（11）
in the method, in the process of the invention,loss value representing the ith task, +.>Indicating total loss value, ++>Representing the actual value of the ith task at the jth moment,/->Indicating the predicted value of the ith task at the jth moment, n being the number of samples, +.>A weight representing an ith task;
s06: based on the formula (12), weighting the loss values of each task in a dynamic weight mode by utilizing the total loss value, and dynamically adjusting according to the training progress of the task to obtain a predicted value:
（12）
in the method, in the process of the invention,weight value representing the ith task to be put into the next m rounds of training, +.>Loss value for each task calculated for this round, < > j->The total loss value obtained by the calculation of the round is calculated;
s07: after the predicted value is inversely normalized, judging the performance of the model through an evaluation index, and if the performance of the model meets the preset prediction precision requirement, storing and outputting model parameters to obtain a pre-trained ground pressure disaster prediction model;
S31: input at time t in input vectorAnd the two-layer stacked LSTM network output of the last moment +.>Splicing to obtain spliced vector +.>；
S32: inputting the spliced vector into a forgetting gate, and mapping the vector between (0, 1) through a sigmoid neural network layer of the forgetting gate based on a formula (3) to obtain an output value of the forgetting gate:
（3）
in the method, in the process of the invention,output value for forgetting gate +.>For the input at time t in the input vector, < >>For the two-layer stacked LSTM network output of the last moment, < >>Weights of sigmoid neural network layer for amnestic gate, < ->Deviation of sigmoid neural network layer of forgetting gate;
s33: the spliced vector is input to an input gate for the first time, mapped to a (0, 1) interval through a sigmoid neural network layer according to a formula (4) to obtain a first output value of the input gate, mapped to a (-1, 1) interval through a tanh neural network layer to generate a state candidate vector:
（4）
in the method, in the process of the invention,for the first output value of the input gate, +.>For the input at time t in the input vector, < >>As a candidate vector for the state,weights of sigmoid neural network layer for input gates, +.>Deviation of sigmoid neural network layer for input gate, +. >Weights of tanh neural network layer as input gates, +.>Deviations of the tanh neural network layer that are input gates;
s34: based on the forgetting gate output value, updating the state candidate through a formula (5) to obtain an updated state candidate:
（5）
in the method, in the process of the invention,is a state candidate vector, +.>For updated state candidates +.>Information transmitted at time t-1 for selective forgetting;
s35: the spliced vector is input and output for the second time, mapped to a (0, 1) interval through a sigmoid neural network layer according to a formula (6), and integrated with the updated state candidate quantity to obtain the input gate second output value at the moment t:
（6）
in the method, in the process of the invention,for the second output value of the input gate, +.>For the updated sigmoid neural network layer weights of the input gates,for the bias of the sigmoid neural network layer in the input gate, +.>Outputting for an LSTM network;
s36: inputting the second output value of the input gate into an Attention module, and distributing corresponding weights according to different time step importance according to a formula (7) to obtain the Attention score of the second output value of the input gate:
（7）
in the method, in the process of the invention,for attention score, ++>Weight of full connection layer, +.>The deviation of the full connection layer is given, and tanh is the full connection layer activation function;
S37: based on the formula (8), the attention score is input, and the attention weight corresponding to the second output value of the input door at the moment t is obtained through calculation of a softmax function:
（8）
in the method, in the process of the invention,is the attention weight;
s38: based on a formula (9), the Attention weight is input to carry out weighted summation on the second output value of the input gate at the moment t, so as to obtain the output value of the Attention module:
（9）
in the method, in the process of the invention,the output value of the Attention module;
s39: inputting the output value of the Attention module into four fully connected layers to output a predicted value, wherein the predicted value comprises: the four full connection layers correspond to four tasks.
7. An electronic device, the electronic device comprising: the device comprises a shell, a processor, a memory, a circuit board and a power circuit, wherein the circuit board is arranged in a space surrounded by the shell, and the processor and the memory are arranged on the circuit board; a power supply circuit for supplying power to each circuit or device of the electronic apparatus; the memory is used for storing executable program codes; a processor executes a program corresponding to the executable program code by reading the executable program code stored in the memory for performing the method of any of the preceding claims 1 to 5.
8. A computer readable storage medium storing one or more programs executable by one or more processors to implement the method of any of the preceding claims 1-5.
CN202311361552.0A 2023-10-20 2023-10-20 Ground pressure disaster prediction method and device based on multitasking learning Active CN117094235B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311361552.0A CN117094235B (en) 2023-10-20 2023-10-20 Ground pressure disaster prediction method and device based on multitasking learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311361552.0A CN117094235B (en) 2023-10-20 2023-10-20 Ground pressure disaster prediction method and device based on multitasking learning

Publications (2)

Publication Number Publication Date
CN117094235A CN117094235A (en) 2023-11-21
CN117094235B true CN117094235B (en) 2024-02-20

Family

ID=88780188

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311361552.0A Active CN117094235B (en) 2023-10-20 2023-10-20 Ground pressure disaster prediction method and device based on multitasking learning

Country Status (1)

Country Link
CN (1) CN117094235B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113723595A (en) * 2021-09-02 2021-11-30 重庆大学 Coal mine rock burst intelligent early warning method based on microseismic event quantitative prediction
CN115545272A (en) * 2022-08-23 2022-12-30 辽宁大学 Long-term and short-term intelligent prediction and early warning method for microseismic frequency and energy of rock burst mine
CN115983493A (en) * 2023-02-13 2023-04-18 重庆大学 Water quality prediction method based on multitask learning and deep learning model
US11630227B1 (en) * 2022-03-21 2023-04-18 Liaoning University Method and system for recognizing mine microseismic event
CN116861789A (en) * 2023-07-13 2023-10-10 成都理工大学 Rock burst prediction method of microseismic parameter multi-input sequence prediction model

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113723595A (en) * 2021-09-02 2021-11-30 重庆大学 Coal mine rock burst intelligent early warning method based on microseismic event quantitative prediction
US11630227B1 (en) * 2022-03-21 2023-04-18 Liaoning University Method and system for recognizing mine microseismic event
CN115545272A (en) * 2022-08-23 2022-12-30 辽宁大学 Long-term and short-term intelligent prediction and early warning method for microseismic frequency and energy of rock burst mine
CN115983493A (en) * 2023-02-13 2023-04-18 重庆大学 Water quality prediction method based on multitask learning and deep learning model
CN116861789A (en) * 2023-07-13 2023-10-10 成都理工大学 Rock burst prediction method of microseismic parameter multi-input sequence prediction model

Also Published As

Publication number Publication date
CN117094235A (en) 2023-11-21

Similar Documents

Publication Publication Date Title
Xue et al. Predicting rockburst with database using particle swarm optimization and extreme learning machine
Gerstenberger et al. Probabilistic seismic hazard analysis at regional and national scales: State of the art and future challenges
Zhou et al. Predicting tunnel squeezing using support vector machine optimized by whale optimization algorithm
Zhou et al. Classification of rockburst in underground projects: comparison of ten supervised learning methods
Mahsuli Probabilistic models, methods, and software for evaluating risk to civil infrastructure
Lind et al. Artificial neural networks in drilling troubles prediction
Qiu et al. Short-term rockburst prediction in underground project: Insights from an explainable and interpretable ensemble learning model
Deng et al. Risk evaluation model of highway tunnel portal construction based on BP fuzzy neural network
Tzamos et al. Extending the Q system's prediction of support in tunnels employing fuzzy logic and extra parameters
CN115423167A (en) Subway deep foundation pit construction safety early warning and decision-making assisting method and system
CN111827982A (en) Method and device for predicting overflow and leakage working conditions of drilling well
Kothari et al. Machine learning: A novel approach to predicting slope instabilities
Fathipour-Azar New interpretable shear strength criterion for rock joints
CN116796825A (en) Rock risk assessment model training method, electronic equipment and readable storage medium
Zhao et al. Back analysis of surrounding rock parameters of tunnel considering displacement loss and space effect
CN113553356A (en) Drilling parameter prediction method and system
Rajabi et al. Prediction of blast-induced ground vibration using empirical models and artificial neural network (Bakhtiari Dam access tunnel, as a case study)
Liu et al. Optimized Adaptive Neuro‐Fuzzy Inference System Using Metaheuristic Algorithms: Application of Shield Tunnelling Ground Surface Settlement Prediction
CN110956543A (en) Method for detecting abnormal transaction
Qiu et al. Analysis and prediction of rockburst intensity using improved DS evidence theory based on multiple machine learning algorithms
Korobkin et al. Diagnosis in SEMS Based on Cognitive Models
Fattahi et al. Prediction of rock drillability using gray wolf optimization and teaching–learning-based optimization techniques
CN117094235B (en) Ground pressure disaster prediction method and device based on multitasking learning
Fang et al. A combined finite element and deep learning network for structural dynamic response estimation on concrete gravity dam subjected to blast loads
Hong et al. A Data‐Driven Method for Predicting the Cutterhead Torque of EPB Shield Machine

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant