Search

CN-121982860-A - Deformation prediction and risk early warning method for underground factory building

CN121982860ACN 121982860 ACN121982860 ACN 121982860ACN-121982860-A

Abstract

The invention discloses an underground plant deformation prediction and risk early warning method, which comprises the steps of firstly obtaining static characteristics representing rock mass attributes and underground plant geometric attributes and dynamic time sequence data representing construction disturbance, constructing a physical driving type input variable comprising space effect characteristics of a displacement sensor and time-varying equivalent support rigidity, then inputting a static-dynamic double-flow fusion deep neural network, deeply coupling static geological genes and dynamic construction disturbance, finally outputting a future displacement sequence by adopting an anti-noise increment direct multi-step prediction strategy, and judging a risk level based on a relative displacement rate. The method effectively overcomes the defect that the traditional model ignores a physical mechanism and a construction driving force, and realizes high-precision, robust prediction and low-cost intelligent early warning of surrounding rock deformation under non-uniform construction and complex geological environment.

Inventors

  • DAI FENG
  • XIONG ZHILIN
  • Song Dingran
  • ZHOU ANXIANG
  • DA YUXIN
  • LIU YI

Assignees

  • 四川大学

Dates

Publication Date
20260505
Application Date
20260123

Claims (10)

  1. 1. The deformation prediction and risk early warning method for the underground factory building is characterized by comprising the following steps of: s1, acquiring multi-source heterogeneous data of different monitoring sections of an oversized underground plant hole group, and structuring the multi-source heterogeneous data into a static tensor and a dynamic tensor form which can be identified by a model; s2, denoising the data obtained by the processing in the step S1, constructing a time sequence supervision learning data set based on multi-source heterogeneous data, carrying out data normalization, and generating a training sample, a verification sample and a test sample with the input window length of L and the prediction window length of H based on a time sliding window mechanism and a time sequence blocking strategy; s3, constructing a multisource data fusion and deep learning prediction model by adopting a PyTorch framework, and training the model by combining a Huber loss function, a AdamW optimizer and a gradient clipping algorithm based on time sequence supervision learning data; S4, outputting a future displacement increment prediction sequence based on the trained model, reconstructing accumulated total displacement at a future moment, constructing an engineering risk early warning model based on a relative displacement rate, and dividing engineering risk grades according to the accumulated total displacement to output corresponding construction control suggestions.
  2. 2. The method for predicting deformation and pre-warning risk of an underground powerhouse according to claim 1, wherein the multi-source heterogeneous data comprises static geological geometrical feature data and dynamic construction monitoring time sequence data.
  3. 3. The method for predicting deformation and early warning risk of an underground powerhouse according to claim 2, wherein the static geological geometrical characteristic data comprise rock mass quality score, burial depth of a cavity, cavity height, cavity width, rock weight, compressive strength, tensile strength, poisson ratio, deformation modulus, internal friction angle and cohesive force, and the dynamic construction monitoring time sequence data comprise accumulated excavation footage length, spatial effect characteristic values of displacement sensors, time-dependent support stiffness coefficient and historical displacement increment.
  4. 4. The method for predicting deformation and pre-warning risk of underground powerhouse according to claim 3, wherein the distance between the displacement sensor and the face is subjected to nonlinear reciprocal transformation to obtain the space effect characteristic value of the displacement sensor The expression is: ; ; In the above-mentioned method, the step of, The space effect characteristic value of the displacement sensor at the moment t is represented, and the larger the value is, the more severe the excavated disturbance is, and the value range (0, 1) is represented; The actual physical distance of the section distance from the current excavation face is monitored at the moment t; The method comprises the steps of determining a cavity width, wherein k is a surrounding rock disturbance attenuation form factor which is not a fixed constant but is an adaptive parameter based on surrounding rock geological parameters, and RMR is a quantitative scoring index of surrounding rock geological integrity and strength, and the value range of the RMR is [0,100]; , e is the base number of natural logarithm, and represents the nonlinear characteristic of the influence of geological properties on disturbance attenuation; the time-varying equivalent support rigidity is The concrete is characterized by being formed by coupling two parts of the aging rigidity of sprayed concrete and the space constraint rigidity of an anchor rod, the value range is usually between [0,1], and the expression is as follows: ; ; ; ; ; In the above-mentioned method, the step of, And The weight is allocated with coefficients, which satisfy The coefficient reflects the relative contribution rate of the shotcrete and the anchor rod in the primary support; And The theoretical maximum rigidity of the spray layer and the anchor rod are respectively used for realizing dimension normalization of data and eliminating the influence of numerical value difference on model training; the instantaneous elastic modulus of the sprayed concrete at the time t; Equivalent radial support rigidity provided for the anchor rod system at the moment t; standard modulus of elasticity of shotcrete for 28 days; is a cement type coefficient and is used for representing the hydration reaction rate; for the effective age of the sprayed concrete, when When the temperature is less than or equal to 0, i.e. the slurry is not sprayed yet; The current monitoring time is the current monitoring time; the actual construction time of the concrete is sprayed for the section; The elastic modulus of the anchor rod body; Is the cross section area of a single anchor rod; And The arrangement intervals of the anchor rods in the circumferential direction and the longitudinal direction are respectively; An activation function in the form of an aging Sigmoid of the anchor rod grouting body, which ensures a smooth transition of the anchor rod stiffness from 0 to the design value; Is the grouting solidification rate factor; Offset for initial set time; Denoising the historical displacement incremental data, if the missing duration is not more than a preset threshold value for 3 days, complementing the historical displacement incremental data by linear interpolation, and if the missing duration is more than the threshold value, cutting off the corresponding sequence to avoid introducing larger errors.
  5. 5. The method for predicting deformation and pre-warning risk of underground powerhouse according to claim 4, wherein the static tensor is characterized by Comprises 11 dimensions, specifically defined as: ; In the formula, RMR is rock mass quality score, Is the burial depth of the cavity, Is the height of the chamber, Is the width of the cavity, Is of rock weight, Is compressive strength, Is poisson's ratio, Tensile strength, E is deformation modulus, Is the internal friction angle and c is the cohesive force; Dynamic tensor The method comprises the steps of constructing a dynamic feature matrix changing along with a time step t aiming at each monitoring sensor, wherein each time step comprises 4 dimensions, and the method is specifically defined as follows: ; In the above-mentioned method, the step of, The accumulated excavation length at the moment t is expressed; The characteristic value of the space effect of the displacement sensor at the moment t is expressed, and the larger the value is, the more severe the excavated disturbance is; expressed as equivalent support stiffness at time t; the historical displacement increment at the last moment denoted t.
  6. 6. The method for predicting deformation and warning risk of an underground powerhouse according to claim 5, wherein in step S2, denoising is performed by processing spike noise generated by explosion vibration or electromagnetic interference in displacement time sequence data by a pointer and adopting a sliding median filtering method, wherein the size of a filtering window is set to 5, and the filtering window is used for effectively eliminating outliers and simultaneously preserving step deformation characteristics of displacement data.
  7. 7. The method for predicting deformation and early warning risk of underground powerhouse according to claim 6, wherein in step S2, the time-series supervised learning data set is obtained by dividing long-term monitoring data into a plurality of sample sets according to the increment displacement for predicting future H days by using the past L days; The data normalization is performed by RobustScaler, and the data is centered and scaled by using the median and the fourth quantile of the dataset, and the expression is as follows: ; In the above formula, x is represented as the original monitoring data, Q 1 、Q 2 、Q 3 is the range of the data of the first quartile 25%, the median 50% and the 3 rd quartile 75%, and Q 3 -Q 1 is the range of the data of the middle 50% in the data distribution.
  8. 8. The method for predicting deformation and pre-warning risk of an underground powerhouse according to claim 7, wherein in the step S2, the time sliding window mechanism is used for inputting data from day 1 to day L, the tag is used for inputting real data from day L+1 to day L+H, namely the first sample, and so on; inputting window length l=60 days, predicting window length h=30 days; Training set data of the time sequence blocking strategy accounts for 70% of total data, verification set data accounts for 15% of total data, and test set data accounts for 15% of total data.
  9. 9. The method for predicting deformation and pre-warning risk of underground powerhouse according to claim 8, wherein in step S3, the architecture of the multi-source data fusion and deep learning prediction model comprises three parts, namely a static encoder, a dynamic sequence encoder and a fusion and decoder; The static encoder utilizes a multi-layer perceptron to extract nonlinear characteristics of static geological parameters, maps 11-dimensional static geological geometric characteristics which do not change with time into a high-dimensional embedded vector s through two fully-connected layers, and has the expression: ; ; In the above-mentioned method, the step of, Is an input static feature vector; , A weight matrix and a bias term for a first full-connection layer; an intermediate hidden layer feature of a static branch; , The method comprises the steps of obtaining a weight matrix and a bias term of a second full-connection layer, wherein a ReLU is a modified linear unit activation function, a calculation formula is max (0, x), dropout is a random inactivation operation, and s is an output static feature embedded vector; The LSTM only extracts the hidden state corresponding to the last time step as the integral representation of the sequence after the complete time sequence is processed, and the expression is as follows: ; ; ; ; ; ; ; In the above-mentioned method, the step of, The dynamic input characteristic at the time t; , the hidden state is the hidden state at the previous moment and the current moment; , , Respectively a forgetting door, an input door and an output door; Is a candidate cell state; , Cell status at the previous time and the current time; the tan h is a hyperbolic tangent activation function; Is Hadamard product; , , , , , , , , , the learning weights and the biasing of each gate control in the long-term and short-term memory network are adopted; The hidden state is the L-th time step, L is the length of the sliding window; taking the hidden state from the last time step of the sequence as the dynamic context vector; The fusion and decoder is used for splicing and fusing the static embedded features and the dynamic features to obtain a fused multi-source feature vector z, and then directly mapping the fused features into displacement increments of 30 days in the future, wherein the specific expression is as follows: ; ; ; in the above formula, z is a fused multisource feature vector which simultaneously contains static geological geometric features and dynamic time sequence monitoring features; And Respectively the weight and bias of the middle layer of the decoding layer; an intermediate layer feature that is a decoding layer; And Respectively the weight and the bias of the output layer; for predicting output vector, i.e. [ ,..., ]; The Huber loss function is used for calculating an error between a model predicted value and a true value, and the specific formula is as follows: ; In the above-mentioned method, the step of, Expressed as a total loss function; The loss value expressed as the i-th sample, when the prediction error is smaller than 1, adopts the mean square error to obtain a smooth gradient, when the error is larger than 1, adopts the linear absolute error to reduce the punishment weight of the outlier noise to model training, and the specific formula is as follows: ; In the above-mentioned method, the step of, Representing the actual total displacement increment value of the ith sample in the future 30 days; representing model predicted total displacement increment values for the ith sample over the future 30 days; a threshold parameter for Huber loss, which is 1; the AdamW optimizer is used for iteratively updating the weight parameters of the neural network, and decouples the weight attenuation items and the gradient updating process, so that the weight attenuation directly acts on the parameters, and the specific expression is as follows: ; ; ; ; ; In the above formula, itera represents the time step of the current iteration; And Respectively representing model weight parameters before updating and after updating; Representing the gradient corresponding to the current time step; And Respectively representing a first moment estimated value and a second moment estimated value of the gradient; And Respectively representing the first moment estimated value and the second moment estimated value after deviation correction; And Is the exponential decay rate; is the learning rate; Is the weight attenuation coefficient; Is a numerical stability constant.
  10. 10. The method for predicting deformation and warning risk of an underground powerhouse according to claim 9, wherein in step S4, the specific formula of the accumulated total displacement at the future time is: In the above-mentioned method, the step of, Monitoring the obtained actual measurement value of the accumulated total displacement at the current moment; predicting a total increment of displacement on a future K day for the model; Cumulative total displacement after 30 days as predicted; The expression of the relative displacement rate R d is: ; In the above-mentioned method, the step of, The unit is mm for predicting the obtained future final accumulated displacement value; The unit is m for the width of the monitored chamber; The concrete contents of the engineering risk early warning and construction control advice are that R d is less than 0.2%, the system advice is low in risk and normal in working condition, R d is less than or equal to 0.2%, the system advice is medium in risk, the advice is used for enhancing monitoring, the stress of an anchor rod is checked, R d is less than or equal to 0.5 and less than or equal to 1.0%, the system advice is high in risk, the advice is used for suspending the progress of a face, the reinforcement support is implemented, R d is more than or equal to 1.0%, the system advice is used for ultra-high risk, the destabilization risk exists, the immediate shutdown of people is advice, and the emergency plan is started.

Description

Deformation prediction and risk early warning method for underground factory building Technical Field The invention relates to the technical field of intelligent construction and safety monitoring of underground engineering, in particular to a deformation prediction and risk early warning method for an underground plant. Background In order to save construction cost, reduce head loss and fully exert self bearing capacity of surrounding rocks, most current diversion power generation systems mostly adopt an underground structure form, namely an underground factory building is constructed. Accurate prediction of surrounding rock deformation of an underground factory building is a key for evaluating stability of a chamber and guiding dynamic design and construction. At present, a prediction method for underground cavern displacement mainly comprises numerical simulation, theoretical analysis and a statistical analysis and machine learning method based on monitoring data. Although the above-described methods have wide application in engineering practice, there are still limitations that are difficult to overcome in the face of complex "geological-construction" coupling environments for very large underground works, including in particular the following problems: (1) While the theoretical analysis method is limited by idealized assumptions such as homogeneous isotropy, nonlinear deformation under a complex geological structure is difficult to quantitatively describe, the numerical simulation is limited by uncertainty of mechanical parameter acquisition and lengthy period of modeling calculation, so that dynamic feedback cannot be carried out in real time along with daily excavation progress, serious timeliness lag exists, and the requirement of on-site high-frequency safety monitoring is difficult to meet. (2) The existing time sequence prediction model mostly adopts univariate time sequence extrapolation, and falsely regards surrounding rock deformation as a spontaneous process with natural lapse of time, and neglects the control action of static characteristics such as rock mass, burial depth, cavity size and the like on deformation magnitude and the direct driving action of excavation footage on deformation rate. Therefore, the model cannot sense nonlinear influence caused by geological difference, and cannot capture working condition mutation caused by shutdown, rework or explosion fluctuation, and finally serious trend deviation and phase lag of a predicted result are caused. (3) The existing prediction model usually only adopts an absolute physical distance to simply mark the position of a sensor, so that the physical connection between fixed-point monitoring data and dynamic excavation unloading is cut off. The linearization process ignores nonlinear attenuation rules of excavation disturbance to specific sensor positions under different lithology and cavern scale, so that the model cannot adaptively sense stress release degree according to geological differences, and severe distortion is predicted in soft and hard rock interactive strata. (4) The existing prediction model only uses discrete digital labels to mark the supporting state, and ignores the physical facts that sprayed concrete and grouting body are hydrated and hardened along with time. The model regards the support as a constant constraint which takes effect instantly, and the process of gradually increasing the support rigidity from zero cannot be reflected, so that the predicted value and the actual stress state of the model are seriously distorted in a critical safety window period of the support in the initial stage of construction. (5) The existing prediction architecture has poor noise immunity and serious error accumulation, and the medium-long term prediction is not only divergent but also unreliable. The underground engineering monitoring data is often accompanied with spike noise such as blasting vibration, the mean square error loss function adopted by the conventional model is extremely sensitive to abnormal values and is easy to cause training divergence, and meanwhile, the adopted full-displacement recursion prediction strategy generally causes single-step errors to be accumulated exponentially along with the prediction step length. Along with the continuous improvement of monitoring technology and informatization level, a great amount of multisource monitoring information can be acquired in the construction and operation processes of the underground factory building, wherein the multisource monitoring information comprises surrounding rock displacement data of different parts and parameters such as accumulated excavation distance reflecting construction disturbance characteristics, supporting completion conditions, excavation rate and the like. These data contain important rules of deformation evolution of surrounding rock, but the conventional method is difficult to fully excavate and comprehensively utilize the surrounding rock