A True Triaxial Failure Prediction Method and System for Deep Engineering Dynamic Disturbance in Hard Rock

Abstract

This invention provides a method and system for predicting the failure of deep-buried hard rock under true triaxial dynamic disturbance, relating to the field of rock mechanics testing technology. The method includes: acquiring basic information; constructing a database based on the basic information to obtain a database of failure caused by true triaxial dynamic disturbance in deep-buried hard rock; reconstructing the stress sequence based on loading parameters to obtain a simulated three-dimensional principal stress sequence; obtaining a multi-source training dataset based on the simulated three-dimensional principal stress sequence and the database of failure caused by true triaxial dynamic disturbance in deep-buried hard rock; constructing a prediction model based on the multi-source training dataset to obtain a three-dimensional strain prediction model; and predicting the failure behavior based on the loading parameters of the condition to be predicted to obtain the prediction result. This invention achieves accurate and efficient intelligent prediction of the entire three-dimensional strain evolution behavior of deep-buried hard rock under true triaxial dynamic disturbance conditions, with only the input loading parameters.

Description

Technical Field

[0001] This invention relates to the field of rock mechanics testing technology, and more specifically, to a method and system for predicting true triaxial hard rock failure caused by dynamic disturbance in deep engineering. Background Technology

[0002] As underground engineering construction continues to extend deeper, numerous deep-buried tunnels, mine roadways, and underground water conservancy projects are highly susceptible to dynamic disasters such as rockbursts under high ground stress and intense excavation disturbances, seriously threatening construction safety. Therefore, establishing reliable methods for predicting rock mass failure is crucial. Currently, the main technical approaches for rockburst prediction have significant limitations: empirical criteria methods rely on macroscopic parameters and empirical formulas, making it difficult to reflect the entire process of rock mass failure evolution under complex stress paths and dynamic disturbance conditions; numerical simulation methods are difficult to achieve rapid prediction due to the complex geological conditions of deep rock masses, the difficulty in determining constitutive model parameters, and high computational costs.

[0003] With the development of artificial intelligence technology, some studies have begun to attempt to apply machine learning methods to rockburst prediction. For example, Chinese patent CN105260599A proposes a dynamic prediction method for rockburst based on BP neural network modeling. However, its inputs are mainly macroscopic rock mechanics parameters and geostress indices, lacking characterization of stress-strain evolution characteristics throughout the entire rock mass loading process. Another example is Chinese patent CN117972852A, which proposes a dynamic prediction method for rockburst based on multi-source heterogeneous data and machine learning. Although it can improve accuracy by utilizing multi-source monitoring data such as microseismic and acoustic emission, its predictions rely on field monitoring data, have weak control over experimental conditions and stress paths, and are difficult to directly use to analyze the failure evolution of deeply buried hard rock under complex true triaxial stress and dynamic disturbance conditions.

[0004] In experimental research, while true triaxial tests can more realistically simulate the three-dimensional stress state of deep rock masses and study the dynamic failure behavior of rock masses through periodic perturbation loading, these tests generate massive amounts of stress-strain data throughout the entire process. Directly using this data as input for prediction models would lead to exceptionally complex data processing and would prevent rapid prediction without prior experimental data. Therefore, existing technologies have the following shortcomings: existing rockburst prediction methods largely rely on macroscopic indicators or field monitoring data, making it difficult to reflect the full-process failure evolution characteristics of rock masses under complex stress paths and dynamic perturbation conditions; existing artificial intelligence prediction methods typically lack experimental databases supporting true triaxial stress states and dynamic perturbation conditions; and the massive scale of stress data obtained from true triaxial tests makes it difficult to directly use as input for prediction models, limiting the practical application of intelligent prediction methods.

[0005] Based on the shortcomings of the existing technologies, there is an urgent need for a true triaxial hard rock failure prediction method and system for deep engineering dynamic disturbance. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predicting failure of hard rock under dynamic disturbance in deep engineering, in order to improve the above-mentioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0007] Firstly, this application provides a method for predicting failure of hard rock in deep engineering under dynamic disturbance, comprising:

[0008] Acquire basic information, including mechanical response time-series data, failure characteristic information, and corresponding loading condition parameters of the true triaxial dynamic disturbance test of deep-buried hard rock.

[0009] Based on the aforementioned basic information, a database is constructed. By structurally integrating test data under multiple confining pressure and disturbance conditions, a database of true triaxial dynamic disturbance damage in deeply buried hard rock is obtained.

[0010] The stress sequence is reconstructed based on the loading condition parameters. Based on the multi-stage loading logic, the discrete time sequence of the three principal stress directions from the start of loading to the end of the disturbance is generated step by step to obtain the simulated three-dimensional principal stress sequence.

[0011] By aligning the simulated three-dimensional principal stress sequence with the mechanical response time series data in the deep-buried hard rock true triaxial dynamic disturbance failure database using time scales, and by unifying the sampling density of the simulated three-dimensional principal stress sequence to a time resolution consistent with the real experimental data, a multi-source training dataset is obtained.

[0012] A prediction model is constructed based on the multi-source training dataset. The three-dimensional strain prediction model is obtained by learning and fitting the nonlinear mapping relationship between the three-dimensional principal stress sequence and the three-dimensional principal strain sequence under multiple working conditions.

[0013] The failure behavior is predicted based on the loading parameters of the working condition to be predicted. The prediction result is obtained by inputting the loading parameters of the working condition to be predicted into the three-dimensional strain prediction model.

[0014] Secondly, this application also provides a true triaxial hard rock failure prediction system for deep engineering dynamic disturbance, comprising:

[0015] The acquisition module is used to acquire basic information, which includes mechanical response time-series data, failure characteristic information and corresponding loading condition parameters of the true triaxial dynamic disturbance test of deep buried hard rock.

[0016] The construction module is used to construct a database based on the basic information. By structurally integrating test data under multiple confining pressure and disturbance conditions, a database of true triaxial dynamic disturbance damage in deeply buried hard rock is obtained.

[0017] The reconstruction module is used to reconstruct the stress sequence according to the loading condition parameters. Based on the multi-stage loading logic, it generates discrete time sequences of the three principal stress directions from the start of loading to the end of disturbance step by step to obtain the simulated three-dimensional principal stress sequence.

[0018] The alignment module is used to align the time scale of the simulated three-dimensional principal stress sequence with the mechanical response time series data in the deep-buried hard rock true triaxial dynamic disturbance failure database. By unifying the sampling density of the simulated three-dimensional principal stress sequence to the same time resolution as the real experimental data, a multi-source training dataset is obtained.

[0019] The modeling module is used to construct a prediction model based on the multi-source training dataset. By learning and fitting the nonlinear mapping relationship between the three-dimensional principal stress sequence and the three-dimensional principal strain sequence under multiple working conditions, a three-dimensional strain prediction model is obtained.

[0020] The prediction module is used to predict the destructive behavior based on the loading parameters of the working condition to be predicted. The prediction result is obtained by inputting the loading parameters of the working condition to be predicted into the three-dimensional strain prediction model.

[0021] The beneficial effects of this invention are as follows:

[0022] This invention achieves accurate and efficient intelligent prediction of the three-dimensional strain evolution behavior of deeply buried hard rock under true triaxial dynamic disturbance conditions by constructing a standardized database based on true triaxial dynamic disturbance tests, introducing a simulated stress generation and time alignment mechanism driven by loading condition parameters, and establishing a three-dimensional strain prediction model. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a method for predicting true triaxial hard rock failure caused by dynamic disturbance in deep engineering, as described in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of a true triaxial hard rock failure prediction system for deep engineering dynamic disturbance as described in an embodiment of the present invention;

[0026] Figure 3 Schematic diagram of true triaxial loading for dynamic disturbance in deeply buried hard rock;

[0027] Figure 4 This is a schematic diagram of the three-axis dynamic disturbance loading path;

[0028] Figure 5 The full stress-strain curve for true triaxial dynamic disturbance;

[0029] Figure 6 This is a schematic diagram of the superposition of perturbation sine waves;

[0030] Figure 7 This is a schematic diagram simulating the stress generation mechanism;

[0031] Figure 8 This is a schematic diagram of the TTR prediction curve.

[0032] The diagram is labeled as follows: 901, Acquisition Module; 902, Construction Module; 903, Reconstruction Module; 904, Alignment Module; 905, Modeling Module; 906, Prediction Module. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0034] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0035] During the construction and operation of deep underground engineering projects, the rock mass is subjected to an extreme geological environment characterized by "three highs and one disturbance": high ground stress, high ground temperature, high osmotic pressure, and intense dynamic disturbance. Frequent dynamic disturbances induced by drilling and blasting (DBM) construction, adjacent cavern blasting, or far-field earthquakes couple with the existing complex three-dimensional stress field, making the mechanical response and failure mechanism of deeply buried hard rock exceptionally complex. In this environment, rock mass failure is not a simple static overload, but a dynamic damage accumulation process driven by the coupling of multiple factors, including excavation stress paths, confining pressure effects, and multi-source dynamic disturbances. Under a specific excavation scheme, what three-dimensional stress evolution path will the rock mass undergo, how will its deformation develop over time, and when will instability and failure occur? Traditional methods rely on post-completion monitoring feedback, a passive "after-the-fact" or "during-the-fact" response. This approach cannot conduct safety simulations and comparisons of multiple design schemes in advance, and it is even more difficult to provide a sufficiently long early warning window before failure occurs. Therefore, engineering practice urgently needs a predictive technology that can connect "design parameters" with "failure and instability," that is, by simply inputting the designed loading conditions (confining pressure, excavation stress path, disturbance parameters), it can proactively and comprehensively simulate the three-dimensional mechanical behavior and failure sequence of the rock mass under these conditions, thereby realizing a shift in decision-making mode from "passive response" to "proactive pre-control." This invention addresses this practical engineering need by providing a complete intelligent predictive solution.

[0036] Example 1:

[0037] This embodiment provides a method for predicting true triaxial hard rock failure caused by dynamic disturbance in deep engineering.

[0038] See Figure 1 The figure shows that the method includes steps S100 to S600.

[0039] Step S100: Obtain basic information, including mechanical response time series data, failure characteristic information and corresponding loading condition parameters of the true triaxial dynamic disturbance test of deep buried hard rock;

[0040] Understandably, mechanical response time-series data refers to the sequence of stress and strain changes over time along the three principal stress directions from the start of loading to specimen failure, synchronously acquired and recorded by sensors during the experiment. Failure characteristic information specifically refers to the key failure event points directly observed and recorded during the experiment, mainly including the moment the specimen fractures and the cumulative number of perturbation loadings up to that moment. Loading condition parameters are the pre-set physical conditions that control the experimental process, including confining pressure conditions (set values ​​for intermediate and minimum principal stresses), loading stress paths (such as loading rates and target stress values ​​at each stage), and dynamic perturbation parameters (perturbation frequency, perturbation amplitude, timing of perturbation application, and duration). All these data together constitute a complete information set from the experimental physical process to its digital characterization.

[0041] Step S200: Construct a database based on basic information. By structurally integrating test data under multiple confining pressure and disturbance conditions, a database of true triaxial dynamic disturbance damage in deeply buried hard rock is obtained.

[0042] It should be noted that in step S200, the original time-series data of each set of experiments is truncated based on the recorded rupture time. Only the effective data segment from the start of loading to the rupture time is retained, thereby eliminating the interference of post-rupture stage data on model learning. Then, all effective samples under different confining pressure combinations and different dynamic disturbance conditions are integrated according to a unified input parameter, time series data and output label structure to form a standardized multi-factor coupled database covering a variety of complex working conditions, providing a high-quality and consistent data foundation for subsequent model training.

[0043] Step S300: Reconstruct the stress sequence according to the loading condition parameters. Based on the multi-stage loading logic, generate discrete time sequences of the three principal stress directions from the start of loading to the end of disturbance step by step to obtain the simulated three-dimensional principal stress sequence.

[0044] Understandably, the stress sequence reconstruction in step S300 is a solution designed to address the real-world challenge of having only loading scheme parameters but lacking complete experimental data in actual engineering or prediction scenarios. Based on given loading condition parameters, this step strictly follows the multi-stage physical loading logic of "three-dimensional synchronous loading - graded loading - perturbation coupling" in real experiments. The algorithm automatically calculates and generates a complete discrete-time sequence of the three principal stress directions from the start of loading to the end of the perturbation. This mechanism essentially constructs a reproducible stress evolution generation model, mapping limited condition parameters to high-fidelity full-process stress path data, thus eliminating the dependence on existing experimental time-series data during prediction.

[0045] Step S400: Align the time scales of the simulated three-dimensional principal stress sequence with the mechanical response time series data in the database of true triaxial dynamic disturbance failure of deeply buried hard rock. By unifying the sampling density of the simulated three-dimensional principal stress sequence to the same time resolution as the real experimental data, a multi-source training dataset is obtained.

[0046] It should be noted that the time-scale alignment in step S400 is to address the time axis misalignment issue that may exist between the simulated stress sequence generated by the algorithm and the data sequence acquired from real experiments due to inconsistent sampling intervals. This step first extracts the inherent time resolution of the real experimental data from the constructed database, and then adjusts the time step of the simulated three-dimensional principal stress sequence to be completely consistent with this resolution, thereby achieving strict alignment between the two in the time dimension. Finally, the time-resolution aligned simulated stress sequence and the corresponding real three-dimensional principal strain sequence in the database are precisely paired using time as the index, forming a multi-source training dataset with simulated stress as the input and real strain as the supervision label, ensuring the physical consistency of input and output in the time scale during model learning.

[0047] Step S500: Construct a prediction model based on the multi-source training dataset. By learning and fitting the nonlinear mapping relationship between the three-dimensional principal stress sequence and the three-dimensional principal strain sequence under multiple working conditions, a three-dimensional strain prediction model is obtained.

[0048] Understandably, step S500 utilizes a multi-source training dataset, using the time-resolution aligned simulated three-dimensional principal stress sequence and perturbation parameters as input features, and the corresponding three-dimensional principal strain sequence measured in real experiments as the learning target. Through artificial intelligence algorithms, massive amounts of time-series data are trained, and deep learning is employed to fit the nonlinear mapping relationship between the three-dimensional principal stress evolution path and the three-dimensional strain response of the rock mass under different confining pressures and perturbation conditions. This results in a data-driven model capable of predicting strain evolution based on the stress path, i.e., a three-dimensional strain prediction model.

[0049] Step S600: Predict the failure behavior based on the loading parameters of the working condition to be predicted. The prediction result is obtained by inputting the loading parameters of the working condition to be predicted into the three-dimensional strain prediction model.

[0050] It should be noted that when predicting a new working condition, only the loading parameters of that condition need to be input. The system will first invoke the stress sequence reconstruction mechanism in step S300 to generate the corresponding simulated three-dimensional principal stress sequence, and then input this sequence into the three-dimensional strain prediction model trained in step S500. Based on the learned stress-strain mapping relationship, the model automatically deduces and outputs the entire process sequence of the evolution of the principal strains in the three directions of the rock mass over time under that working condition. Thus, it is possible to predict the deformation behavior of rock mass under complex true triaxial dynamic disturbances proactively, based solely on the loading scheme, without conducting actual experiments.

[0051] Further, step S200 includes steps S210 to S230.

[0052] Step S210: Collect raw data based on mechanical response time series data and failure characteristic information. By synchronously recording the full process stress-strain time series of the three principal stress directions, the corresponding fracture occurrence time and the number of fracture disturbances during the test, the raw data of the whole process under each test condition is obtained.

[0053] Step S220: Based on the original data of the whole process and the time of rupture, the effective interval is truncated. The three-dimensional stress-strain time series is truncated in the time domain by taking the time of rupture as the cutoff boundary, and the residual stage data after rupture is removed to obtain the effective three-dimensional stress-strain sequence of each sample.

[0054] Step S230: Based on the effective three-dimensional stress-strain sequence and the corresponding loading condition parameters, rupture occurrence time, and number of rupture disturbances, the samples are structurally integrated. By summarizing the test samples under different confining pressure combinations and dynamic disturbance conditions according to a unified data structure, a true triaxial dynamic disturbance failure database for deeply buried hard rock is obtained.

[0055] Specifically, in the true triaxial dynamic perturbation test, by controlling the intermediate principal stress, minimum principal stress, loading stress path, dynamic perturbation frequency, dynamic perturbation amplitude, perturbation application timing, and perturbation duration, the stress-strain data and corresponding time series of the three principal stress directions are recorded simultaneously. The time of fracture occurrence and the corresponding number of perturbations are also recorded simultaneously, thus obtaining the original data of the entire process under each test condition. Subsequently, the time series is truncated based on the recorded time of failure, and residual stage data after the failure of deeply buried hard rock is removed using this time as the cutoff boundary, resulting in an effective three-dimensional stress-strain sequence for each sample that avoids interference with model learning. Afterwards, the experimental data are constructed into a unified data structure, which specifically includes... The database comprises input parameters, time-series data, output labels, and a sample set. Input parameters include intermediate principal stress, minimum principal stress, stress path, perturbation frequency, and amplitude. Time-series data includes time and stress. Output labels include failure strength and strain in the three principal stress directions. The sample set is a summary of all the above data. Finally, all samples under different confining pressure combinations and dynamic perturbation conditions are aggregated according to this unified structure to construct a multi-factor coupled database containing at least the confining pressure combination, loading stress path, dynamic perturbation parameters, full-process three-dimensional stress, full-process three-dimensional strain, failure time, and number of failure perturbations. This forms a true triaxial dynamic perturbation failure database for deep-buried hard rock that can be directly used for training and prediction of artificial intelligence models. The data contained in each sample within the database are shown in Table 1 below.

[0056] Table 1. Data contained in each sample within the database

[0057]

[0058] in, This is the collection of data that needs to be input into the database; Indicates the intermediate principal stress; Indicates the minimum principal stress; Indicates the frequency of dynamic disturbance; Indicates the amplitude of dynamic disturbance; Represents a time series; Represents a sequence of stress changes over time, with subscripts. The values ​​1, 2, and 3 correspond to the maximum principal stress, intermediate principal stress, and minimum principal stress, respectively. This represents a sequence of strain changes over time along the three principal stress directions. This is the output result that needs to be provided for subsequent predictions; For the sample set; For sample index; The total number of samples; Indicates the first The set of input parameters for each sample; Indicates the first Time series data collected during the individual sample trials; Indicates the first The sequence of strain changes over time in the three principal stress directions measured during a sample test; Indicates the first The set of output labels for each sample.

[0059] Further, step S300 includes steps S310 to S330.

[0060] Step S310: Generate a three-dimensional synchronous loading segment sequence based on the loading rate in the loading condition parameters. By synchronously and linearly increasing the three principal stress directions at a uniform rate until the minimum principal stress setting value, the principal stress time sequence of the three-dimensional synchronous loading stage is obtained.

[0061] Step S320: Generate a graded loading segment sequence based on the principal stress time series of the three-dimensional synchronous loading stage. This is achieved by sequentially increasing the maximum principal stress and intermediate principal stress to the intermediate principal stress set value, and then individually increasing the maximum principal stress to the stress level at the time of disturbance application.

[0062] Step S330: Generate a disturbance coupling segment sequence based on the principal stress time sequence of the graded loading stage and the disturbance frequency, disturbance amplitude and disturbance duration in the loading condition parameters. By superimposing periodic disturbance stress in the form of a sine wave in the middle or minimum principal stress direction until the disturbance duration ends, a simulated three-dimensional principal stress sequence is obtained.

[0063] Specifically, in the intelligent prediction of true triaxial dynamic disturbance failure in deeply buried hard rock, the loading path of the three-dimensional principal stresses and their evolution over time are the core input parameters describing the mechanical behavior of the rock mass. From the initial loading stage, the confining pressure establishment stage, the step-by-step enhancement stage, to the occurrence of failure, the stress path not only determines the stress redistribution characteristics within the rock mass but also directly affects the evolution of cumulative damage and the final instability mode. Traditional true triaxial tests typically obtain stress data from hundreds of thousands to millions of discrete sampling points, which are key input features in failure prediction models. However, in the practical application scenarios addressed by this invention, only loading scheme parameters, including confining pressure combinations, disturbance frequency, disturbance amplitude, and stress path rules, are often available, without the complete experimental time-series stress data. Therefore, if real experimental data is still used as model input, equivalent input conditions cannot be provided during the actual inference stage, leading to insufficient model generalization ability or even model failure. To solve the above problems, this invention proposes a three-dimensional simulation stress generation mechanism driven by experimental schemes. This mechanism is not simply data supplementation, but rather the construction of a reproducible, standardizable stress evolution generation model consistent with the real physical loading process. This model can automatically generate discrete-time series of the three principal stresses from the start of loading to failure, given only the input loading conditions. This simulated stress generation mechanism has the following characteristics:

[0064] (1) Consistency of loading logic:

[0065] Strictly following the loading sequence and path rules in a real triaxial test, the following stages were implemented sequentially: triaxial synchronous linear loading stage; confining pressure establishment and principal stress graded loading stage; stepped stress enhancement stage; disturbance superposition and coupling stage; until the preset maximum loading level was reached.

[0066] In the true triaxial dynamic disturbance test of deeply buried hard rock, the deeply buried hard rock sample mainly undergoes five loading stages, namely:

[0067] Stage I: The three principal stresses move at the same rate (MPa / s) Synchronous loading, after a certain time The magnitudes of the three principal stresses reach ;

[0068] Phase II: Maintain constant, , Continue at a rate Loading (MPa / s) for a specified time , , Size reached ,in, , and These represent the target values ​​for the maximum, intermediate, and minimum principal stresses, respectively.

[0069] Phase III: Maintain , constant, Continue at a rate Loading (MPa / s) for a specified time , achieve That is, the maximum principal stress is applied to the stress level at the time of the disturbance.

[0070] Stage IV: When the magnitude of the maximum principal stress reaches the stress level at the time of disturbance, dynamic disturbance stress can be applied in the direction of the intermediate or minimum principal stress, such as... Figure 3 As shown in the figure This represents the perturbation stress wave in the direction of the minimum principal stress. The applied dynamic perturbation stress is a sinusoidal stress wave, i.e. In the formula, Indicates the intermediate principal stress; Indicates the minimum principal stress; Indicates the frequency of dynamic disturbance; Indicates the amplitude of dynamic disturbance; This indicates the duration after the start of the dynamic disturbance. The disturbance duration is... In the formula The superposition of this sinusoidal dynamic disturbance on static stress is as follows: Figure 6 As shown.

[0071] The magnitudes of the three principal stresses at each loading stage are shown in Table 2.

[0072] Table 2 Loading Stages of Hard Rock Specimens During the Test

[0073]

[0074] A visual representation of the multi-stage loading path described above is as follows: Figure 4 As shown. Based on the stress loading path in the above test process, three principal stresses are generated by the algorithm, including the stress data of the deep-buried hard rock sample from the start of loading to reaching the excavation stress path strength.

[0075] (2) Uniformity of time scale:

[0076] To ensure the comparability of training data and real experimental data in the time dimension, this invention stipulates that the sampling interval of simulated stress data is consistent with the sampling interval of real experiments.

[0077] For example, when the actual experimental sampling interval If the value is 0.001s, then the simulated data generates 100 discrete sampling points per second, making the generated three-directional stress sequence completely consistent with the real data in terms of time resolution.

[0078] (3) Computability and cumulativeity of disturbances:

[0079] By superimposing periodic disturbances during the step-holding phase, the stress sequence includes not only static loading information but also the cumulative effect of dynamic disturbances, thus providing a basis for calculating the number of disturbances for the subsequent fracture time conversion mechanism.

[0080] The workflow and core ideas of simulating stress generation mechanisms are as follows: Figure 7 As shown. Through the above-described simulated stress generation mechanism, this invention achieves automatic mapping from "loading scheme parameters" to "three-dimensional stress full-process data", eliminating the dependence on real experimental time-series data, enabling the failure prediction model to directly predict mechanical properties without prior experiments, and constructing a complete technical closed loop of "scheme input - stress generation - failure prediction".

[0081] Compared with traditional methods that directly train models using real experimental time-series data, the simulated stress generation mechanism of this invention improves the model's generalization ability and engineering applicability under different working conditions, while avoiding data misalignment problems caused by uneven sampling of real data, missing time data, or missing data points. Therefore, this simulated stress generation mechanism is not only a data generation step, but also a key foundational module for realizing an intelligent prediction method for true triaxial dynamic disturbance failure in deeply buried hard rock.

[0082] Further, step S400 includes steps S410 to S430.

[0083] Step S410: Extract the benchmark interval based on the time series step size of each sample in the database of true triaxial dynamic disturbance and failure of deep buried hard rock. By statistically analyzing the interval between adjacent sampling points of the time series of each real test sample in the database and taking a unified value, obtain the unified benchmark sampling interval corresponding to the real test data.

[0084] Step S420: Align the sampling density with the simulated three-dimensional principal stress sequence according to the unified reference sampling interval. By adjusting the time step when generating the simulated three-dimensional principal stress sequence to be consistent with the unified reference sampling interval, a simulated three-dimensional principal stress sequence with time resolution alignment is obtained.

[0085] Step S430: Perform sample-level stitching of the simulated three-dimensional principal stress sequence aligned with the time resolution and the effective three-dimensional strain sequence of each sample in the database. By pairing and combining the simulated three-dimensional principal stress sequence aligned with the time resolution of each sample with the corresponding real three-dimensional principal strain sequence based on the time index, a multi-source training dataset is obtained.

[0086] Specifically, since the simulated three-dimensional principal stress sequence constructed in the aforementioned process is generated according to a preset stress path and a fixed sampling time interval, its time series has strict uniformity and there are no problems of uneven sampling or missing data. However, in real triaxial dynamic disturbance tests, due to the long test duration, high loading pressure, and brief response delays in the experimental equipment's execution of control commands, the actual collected stress, deformation, and time series data often exhibit uneven sampling time intervals, local data loss, or duplicate recordings. Simultaneously, short pauses may occur during loading adjustments or operational control during the experiment. Although these time errors account for a small proportion of the overall experimental duration, because the experimental sampling interval is typically small (e.g., 0.01s), these minute time deviations can still lead to hundreds or even thousands of redundant or missing data points in the data series. Therefore, the simulated stress data and the real experimental data form two different time axis structures in the time dimension. If the simulated stress data is directly trained to correspond with the real experimental time series, the model may develop an incorrect time mapping relationship during the learning process, thereby affecting the stability and accuracy of the damage prediction model.

[0087] To address the aforementioned problems, this invention proposes a fracture time conversion mechanism. This mechanism does not directly use time as a unified scale, but instead introduces the number of dynamic perturbations as an intermediate physical variable to achieve a physically consistent mapping between the actual experimental fracture point and the simulated stress sequence. Since the perturbation frequency and the duration of each stage of perturbation are known parameters in the experimental scheme, the number of dynamic perturbations can be uniquely determined by the perturbation frequency and perturbation time. Compared to the time variable, the number of perturbations can more stably reflect the cumulative effect of perturbations during dynamic loading. Preferably:

[0088] First, find the moment of fracture occurrence in the real triaxial dynamic disturbance test data. Simultaneously record the magnitude of the maximum principal stress corresponding to the moment of fracture. Determine the stress step at which the fracture occurred and record the duration of disturbance within that step. ;

[0089] Subsequently, the duration of a single disturbance step is set to be... The perturbation frequency is The total number of actual rupture disturbances is:

[0090] ;

[0091] In the formula, This represents the total number of actual rupture disturbances; Indicates the number of disturbance steps at the moment of rupture; Indicates a disturbance step; Indicates the duration of a step; The perturbation frequency; This represents the duration of continuous disturbance within the step.

[0092] Next, in the simulated stress generation data sequence, find the one that satisfies... The stress platform number, among which, This represents the simulated stress data, enabling the simulation data to terminate the step positioning.

[0093] Next, the simulated fracture time is calculated, assuming the starting time of entering the fracture stress plateau is... The total time for the simulated data to reach the rupture time is:

[0094] ;

[0095] And it needs to meet the following requirements: ;

[0096] In the formula, This represents the total time it takes for the simulated data to reach the rupture point; Indicates the start time of entering the fracture stress plateau; This represents the duration of the disturbance within the step. This represents the total number of actual rupture disturbances; This represents the total number of simulated rupture disturbances.

[0097] Finally, the remaining rupture time is defined, assuming that the time elapsed since the start of the simulation data generation is 1. Then, the remaining rupture time is defined on the simulation time axis:

[0098] ;

[0099] In the formula, It represents the remaining fracture time defined on the simulation timeline, and also serves as an auxiliary damage assessment indicator for the artificial intelligence model; This represents the total time it takes for the simulated data to reach the rupture point; This indicates that time has elapsed since the simulation started generating the data.

[0100] By introducing a fracture time conversion mechanism, this invention achieves temporal unification between real experimental data and simulated stress data, enabling fracture features to be accurately embedded into the simulated stress path. This ensures that the machine learning model can correctly learn the physical correlation between stress evolution and failure occurrence during training. This method not only effectively avoids data misalignment problems caused by unstable experimental time series but also improves the consistency and reliability of model training data, further enhancing the stability and prediction accuracy of the true triaxial dynamic disturbance failure prediction model for deeply buried hard rock. The aforementioned time scale alignment is mainly used to unify data sampling resolution, while the fracture time conversion mechanism is used to establish the physical correspondence between real fracture behavior and simulated stress path. Both mechanisms are applied to data preprocessing and the construction of auxiliary failure evaluation indicators, respectively.

[0101] Further, step S500 includes steps S510 to S530.

[0102] Step S510: Define the model input and output based on the multi-source training dataset. Define the input feature vector of each time step as a set containing the stress values ​​of the three principal stress directions, the frequency of dynamic disturbance, the amplitude of dynamic disturbance, and the cumulative number of disturbances. Define the output target vector as the failure strength and the set of axial strains of the three principal stress directions at the same time step. This will give you the input-output mapping structure of the time series sample.

[0103] Step S520: Construct the model based on the input-output mapping structure. By constructing a function that represents the nonlinear mapping relationship between stress path information from the initial loading time to the current time and three-dimensional strain, and initializing the model parameters, the initial three-dimensional strain prediction model is obtained.

[0104] Step S530: Based on the multi-source training dataset and the initial three-dimensional strain prediction model, perform model training and parameter optimization. Optimize model parameters by constructing a loss function using three-dimensional strain data measured in real experiments in the database as supervision labels, and obtain the trained three-dimensional strain prediction model.

[0105] Specifically, after completing the construction of a true triaxial dynamic disturbance test database for deeply buried hard rock, simulating the stress generation mechanism, and the fracture time conversion mechanism, this invention further constructs an artificial intelligence-based data-driven three-dimensional strain prediction model to establish a nonlinear mapping relationship between the loading stress path, dynamic disturbance conditions, and the three-dimensional strain evolution of the rock mass, thereby realizing the prediction of the three-dimensional deformation process of deeply buried hard rock under complex stress paths and dynamic disturbance conditions. Preferably, the process is as follows:

[0106] First, the input feature sequence is constructed. Under a uniform time sampling interval, the stress state of deeply buried hard rock under true triaxial dynamic disturbance conditions is represented as a time-series input feature vector. For any time step t, the input feature vector is defined as:

[0107] ;

[0108] In the formula, This represents the input feature vector; , and For the three principal stress directions in time The stress value at that moment; For the frequency of dynamic disturbance; The amplitude of the dynamic disturbance; Deadline The cumulative number of disturbances, when a certain dynamic disturbance frequency is selected in a single experiment, can be simplified to the following calculation method: .

[0109] In time series form, the input features at all time steps constitute the input matrix:

[0110] ;

[0111] In the formula, Represents the input feature matrix; , and Indicates the first, second, and third The input feature vector at each time step; The input matrix represents the time series length. It reflects the stress evolution and disturbance accumulation information of the rock mass throughout the loading and dynamic disturbance process.

[0112] Then, the output target variables are defined. The output of the three-dimensional strain prediction model is the strain evolution sequence of the rock mass in the three principal stress directions, expressed as follows:

[0113] ;

[0114] In the formula, This represents the strain evolution sequence of the rock mass in the three principal stress directions; , and These represent the axial strain of the rock mass in the three principal stress directions. For the entire time series, the output data can be represented as: ,in, This represents the output feature matrix. , and Indicates the first, second, and third Output feature vectors at each time step.

[0115] Next, a stress-strain evolution mapping prediction model is established. Since rock masses exhibit significant path dependence during dynamic disturbance loading—meaning the current strain depends not only on the current stress state but also on historical loading paths—an artificial intelligence-based time-series prediction model is constructed. This model performs feature learning on the input time series, and a typical full stress-strain response curve is shown below. Figure 5 As shown. Let the strain prediction function be:

[0116] ;

[0117] In the formula, This represents the strain evolution sequence of the rock mass in the three principal stress directions; For artificial intelligence prediction functions; These are model parameters; This represents the stress path information from the initial loading moment to the current moment. A mapping relationship between stress path and three-dimensional strain is established using a large amount of experimental data from the training database.

[0118] Subsequently, stress change features are extracted. During model training, key physical quantities reflecting rock mass deformation characteristics are extracted from the input sequence, mainly including: stress increment, stress change rate, multiaxial stress coupling relationship, and dynamic disturbance parameters. Through comprehensive learning of the above multidimensional features, the deformation evolution law of rock mass under complex stress path and dynamic disturbance conditions is identified.

[0119] Finally, the three-dimensional strain data measured by real experiments in the database are used as supervision labels to train the prediction model and construct a loss function. By continuously optimizing the model parameters, the error between the predicted strain sequence and the real experimental strain sequence is minimized, and the trained three-dimensional strain prediction model is finally obtained.

[0120] Further, step S600 includes steps S610 to S630.

[0121] Step S610: Generate a simulated three-dimensional principal stress sequence based on the loading condition parameters of the working condition to be predicted. Based on the confining pressure conditions, loading stress path and dynamic disturbance parameters, simulate and generate a three-dimensional principal stress time series from the start to the end of loading to obtain the simulation input sequence of the working condition to be predicted.

[0122] Step S620: Based on the simulated input sequence of the working condition to be predicted and the three-dimensional strain prediction model, perform three-dimensional strain evolution process prediction processing. By inputting the simulated input sequence into the three-dimensional strain prediction model, the nonlinear mapping relationship between the three-dimensional principal stress evolution process and the three-dimensional strain response of the rock mass established by the model is used to calculate and obtain the predicted three-dimensional principal strain evolution sequence.

[0123] Step S630: Extract failure behavior information based on the predicted three-dimensional principal strain evolution sequence. By analyzing the entire strain evolution process from the three-dimensional principal strain evolution sequence, the prediction result is obtained.

[0124] Specifically, step S610, based on the loading parameters of the load condition to be predicted (including confining pressure conditions, loading stress paths, and dynamic disturbance parameters), calls the simulated stress generation mechanism. Strictly following the multi-stage physical logic of "triaxial synchronous linear loading - confining pressure establishment and principal stress graded loading - step-by-step stress enhancement - disturbance superposition coupling" in true triaxial tests, a three-dimensional principal stress discrete time sequence with uniform time resolution is generated step-by-step, thus obtaining the simulated input sequence of the load condition to be predicted. Next, in step S620, this simulated three-dimensional principal stress time sequence (containing specific stress paths and disturbance histories) is input into the trained three-dimensional strain prediction model. Based on the complex nonlinear mapping relationship between the three-dimensional principal stress evolution process and the three-dimensional strain response of the rock mass, learned from multi-source training datasets during the training phase, the model performs calculations and deductions, directly outputting the complete sequence of strain evolution over time in the corresponding three principal stress directions, i.e., the predicted three-dimensional principal strain evolution sequence. Finally, step S630 analyzes the predicted strain sequence. By analyzing the entire virtual strain evolution process (e.g., identifying strain acceleration inflection points, peak values, etc.), core information characterizing rock mass failure behavior, such as failure intensity and failure time (or remaining time to fracture), is extracted. This yields the final prediction result of the rock mass mechanical behavior and failure risk under the new working condition. The remaining time to fracture (TTR) prediction curve and early warning effect of a specific experiment are shown below. Figure 8 As shown.

[0125] Example 2:

[0126] like Figure 2 As shown, this embodiment provides a true triaxial hard rock failure prediction system for deep engineering dynamic disturbance. The system includes:

[0127] The acquisition module 901 is used to acquire basic information, which includes mechanical response time series data, failure characteristic information and corresponding loading condition parameters of the true triaxial dynamic disturbance test of deep buried hard rock.

[0128] Module 902 is used to construct a database based on basic information. By structurally integrating test data under multiple confining pressure and disturbance conditions, a database of true triaxial dynamic disturbance damage in deeply buried hard rock is obtained.

[0129] The reconstruction module 903 is used to reconstruct the stress sequence based on the loading condition parameters. Based on the multi-stage loading logic, it generates discrete time sequences of the three principal stress directions from the start of loading to the end of disturbance step by step to obtain the simulated three-dimensional principal stress sequence.

[0130] Alignment module 904 is used to align the time scale of the simulated three-dimensional principal stress sequence with the mechanical response time series data in the database of true triaxial dynamic disturbance failure of deep buried hard rock. By unifying the sampling density of the simulated three-dimensional principal stress sequence to the same time resolution as the real experimental data, a multi-source training dataset is obtained.

[0131] Modeling module 905 is used to build a prediction model based on a multi-source training dataset. It obtains a three-dimensional strain prediction model by learning and fitting the nonlinear mapping relationship between the three-dimensional principal stress sequence and the three-dimensional principal strain sequence under multiple working conditions.

[0132] The prediction module 906 is used to predict the failure behavior based on the loading parameters of the working condition to be predicted. The prediction result is obtained by inputting the loading parameters of the working condition to be predicted into the three-dimensional strain prediction model.

[0133] In one specific embodiment of this application, the construction module 902 includes:

[0134] The first building unit is used to collect raw data based on mechanical response time series data and failure characteristic information. By synchronously recording the full process stress-strain time series of the three principal stress directions and the corresponding fracture occurrence time and fracture disturbance number during the test, the raw data of the whole process under each test condition is obtained.

[0135] The second building unit is used to extract the effective interval based on the original data of the whole process and the time of rupture. By using the time of rupture as the cutoff boundary to truncate the three-dimensional stress-strain time series in the time domain, the residual stage data after rupture is removed, and the effective three-dimensional stress-strain sequence of each sample is obtained.

[0136] The third building unit is used to structurally integrate samples based on the effective three-dimensional stress-strain sequence and the corresponding loading condition parameters, fracture occurrence time, and fracture disturbance number. By summarizing the test samples under different confining pressure combinations and dynamic disturbance conditions according to a unified data structure, a true triaxial dynamic disturbance failure database for deeply buried hard rock is obtained.

[0137] In one specific embodiment of this application, the reconstruction module 903 includes:

[0138] The first reconstruction unit is used to generate a three-dimensional synchronous loading segment sequence based on the loading rate in the loading condition parameters. By synchronously and linearly increasing the three principal stress directions at a uniform rate until the minimum principal stress setting value is reached, the principal stress time sequence of the three-dimensional synchronous loading stage is obtained.

[0139] The second reconstruction unit is used to generate a graded loading segment sequence based on the principal stress time series of the three-dimensional synchronous loading stage. By sequentially increasing the maximum principal stress and the intermediate principal stress to the intermediate principal stress set value, and increasing the maximum principal stress alone to the stress level when the disturbance is applied, the principal stress time series of the graded loading stage is obtained.

[0140] The third reconstruction unit is used to generate a disturbance coupling segment sequence based on the principal stress time sequence of the graded loading stage and the disturbance frequency, disturbance amplitude and disturbance duration in the loading condition parameters. By superimposing periodic disturbance stress in the form of a sine wave in the middle or minimum principal stress direction until the disturbance duration ends, a simulated three-dimensional principal stress sequence is obtained.

[0141] In one specific embodiment of this application, the alignment module 904 includes:

[0142] The first alignment unit is used to extract the benchmark interval based on the time series step size of each sample in the database of true triaxial dynamic disturbance and damage of deep buried hard rock. By statistically analyzing the interval between adjacent sampling points of the time series of each real test sample in the database and taking a unified value, the unified benchmark sampling interval corresponding to the real test data is obtained.

[0143] The second alignment unit is used to align the sampling density with the simulated three-dimensional principal stress sequence according to the unified reference sampling interval. By adjusting the time step when generating the simulated three-dimensional principal stress sequence to be consistent with the unified reference sampling interval, a time-resolution aligned simulated three-dimensional principal stress sequence is obtained.

[0144] The third alignment unit is used to perform sample-level stitching of the simulated three-dimensional principal stress sequence aligned according to the time resolution and the effective three-dimensional strain sequence of each sample in the database. By pairing and combining the simulated three-dimensional principal stress sequence aligned with the time resolution of each sample with the corresponding real three-dimensional principal strain sequence based on the time index, a multi-source training dataset is obtained.

[0145] In one specific embodiment of this application, the modeling module 905 includes:

[0146] The first modeling unit is used to define the model input and output based on the multi-source training dataset. It defines the input feature vector of each time step as a set containing the stress values ​​of the three principal stress directions, the frequency of dynamic disturbance, the amplitude of dynamic disturbance, and the cumulative number of disturbances. The output target vector is defined as the failure strength and the set of axial strains in the three principal stress directions at the same time step, thus obtaining the input-output mapping structure of the time series sample.

[0147] The second modeling unit is used to build the model according to the input-output mapping structure. It constructs a function that represents the nonlinear mapping relationship between the stress path information from the initial loading time to the current time and the three-dimensional strain, and initializes the model parameters to obtain the initial three-dimensional strain prediction model.

[0148] The third modeling unit is used to train the model and optimize the parameters based on the multi-source training dataset and the initial three-dimensional strain prediction model. It optimizes the model parameters by constructing a loss function using the three-dimensional strain data measured by real experiments in the database as supervision labels, and obtains the trained three-dimensional strain prediction model.

[0149] In one specific embodiment of this application, the prediction module 906 includes:

[0150] The first prediction unit is used to generate a simulated three-dimensional principal stress sequence based on the loading condition parameters of the working condition to be predicted. Based on the confining pressure conditions, loading stress path and dynamic disturbance parameters, it simulates and generates a three-dimensional principal stress time series from the start to the end of loading, thus obtaining the simulation input sequence of the working condition to be predicted.

[0151] The second prediction unit is used to predict the three-dimensional strain evolution process based on the simulated input sequence of the working condition to be predicted and the three-dimensional strain prediction model. By inputting the simulated input sequence into the three-dimensional strain prediction model, the nonlinear mapping relationship between the three-dimensional principal stress evolution process and the three-dimensional strain response of the rock mass established by the model is used to calculate and obtain the predicted three-dimensional principal strain evolution sequence.

[0152] The third prediction unit is used to extract information on destructive behavior based on the predicted three-dimensional principal strain evolution sequence. By analyzing the entire strain evolution process from the three-dimensional principal strain evolution sequence, the prediction results are obtained.

[0153] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting failure of hard rock under dynamic disturbance in deep engineering, characterized in that, include: Acquire basic information, including mechanical response time-series data, failure characteristic information, and corresponding loading condition parameters of the true triaxial dynamic disturbance test of deep-buried hard rock. Based on the aforementioned basic information, a database is constructed. By structurally integrating test data under multiple confining pressure and disturbance conditions, a database of true triaxial dynamic disturbance damage in deeply buried hard rock is obtained. The stress sequence is reconstructed based on the loading condition parameters. Based on the multi-stage loading logic, the discrete time sequence of the three principal stress directions from the start of loading to the end of the disturbance is generated step by step to obtain the simulated three-dimensional principal stress sequence. By aligning the simulated three-dimensional principal stress sequence with the mechanical response time series data in the deep-buried hard rock true triaxial dynamic disturbance failure database using time scales, and by unifying the sampling density of the simulated three-dimensional principal stress sequence to a time resolution consistent with the real experimental data, a multi-source training dataset is obtained. A prediction model is constructed based on the multi-source training dataset. The three-dimensional strain prediction model is obtained by learning and fitting the nonlinear mapping relationship between the three-dimensional principal stress sequence and the three-dimensional principal strain sequence under multiple working conditions. The failure behavior is predicted based on the loading parameters of the working condition to be predicted. The prediction result is obtained by inputting the loading parameters of the working condition to be predicted into the three-dimensional strain prediction model.

2. The method for predicting true triaxial hard rock failure caused by dynamic disturbance in deep engineering according to claim 1, characterized in that, The database is constructed based on the aforementioned basic information, including: Based on the mechanical response time series data and the failure characteristic information, raw data is collected. By synchronously recording the full process stress-strain time series of the three principal stress directions, the corresponding fracture occurrence time and the number of fracture disturbances during the test, the full process raw data under each test condition is obtained. Based on the original data of the entire process and the time of rupture, an effective interval is extracted. The three-dimensional stress-strain time series is truncated in the time domain by taking the time of rupture as the cutoff boundary, and the residual stage data after rupture is removed to obtain the effective three-dimensional stress-strain sequence of each sample. Based on the effective three-dimensional stress-strain sequence and the corresponding loading condition parameters, rupture occurrence time, and number of rupture disturbances, the samples are structurally integrated. By summarizing the test samples under different confining pressure combinations and dynamic disturbance conditions according to a unified data structure, a true triaxial dynamic disturbance failure database for deeply buried hard rock is obtained.

3. The method for predicting true triaxial hard rock failure caused by dynamic disturbance in deep engineering according to claim 1, characterized in that, Based on the loading condition parameters, stress sequence reconstruction is performed, including: The three-dimensional synchronous loading segment sequence is generated based on the loading rate in the loading condition parameters. The principal stress time sequence of the three-dimensional synchronous loading stage is obtained by synchronously and linearly increasing the three principal stress directions at a uniform rate until the minimum principal stress setting value. Based on the principal stress time series of the three-dimensional synchronous loading stage, a graded loading segment sequence is generated. The principal stress time series of the graded loading stage is obtained by sequentially increasing the maximum principal stress and the intermediate principal stress to the intermediate principal stress set value, and increasing the maximum principal stress alone to the stress level when the disturbance is applied. Based on the principal stress time series of the graded loading stage and the disturbance frequency, disturbance amplitude and disturbance duration in the loading condition parameters, a disturbance coupling segment sequence is generated. By superimposing periodic disturbance stress in the form of a sine wave in the middle or minimum principal stress direction until the disturbance duration ends, a simulated three-dimensional principal stress sequence is obtained.

4. The method for predicting true triaxial hard rock failure caused by dynamic disturbance in deep engineering according to claim 1, characterized in that, The time scale is aligned between the simulated three-dimensional principal stress sequence and the mechanical response time series data in the deep-buried hard rock true triaxial dynamic disturbance failure database, including: Based on the time series step size of each sample in the deep-buried hard rock true triaxial dynamic disturbance failure database, the benchmark interval is extracted. By statistically analyzing the interval between adjacent sampling points of the time series of each real test sample in the database and taking a unified value, the unified benchmark sampling interval corresponding to the real test data is obtained. The sampling density is aligned with the unified reference sampling interval and the simulated three-dimensional principal stress sequence. By adjusting the time step when generating the simulated three-dimensional principal stress sequence to be consistent with the unified reference sampling interval, a time resolution aligned simulated three-dimensional principal stress sequence is obtained. The simulated three-dimensional principal stress sequence aligned with the time resolution is concatenated with the effective three-dimensional strain sequence of each sample in the database at the sample level. By pairing and combining the simulated three-dimensional principal stress sequence aligned with the time resolution of each sample with the corresponding real three-dimensional principal strain sequence based on the time index, a multi-source training dataset is obtained.

5. The method for predicting true triaxial hard rock failure caused by dynamic disturbance in deep engineering according to claim 1, characterized in that, Building a prediction model based on the multi-source training dataset includes: The model input and output are defined based on the multi-source training dataset. The input feature vector of each time step is defined as a set containing the stress values ​​of the three principal stress directions, the frequency of dynamic disturbance, the amplitude of dynamic disturbance, and the cumulative number of disturbances. The output target vector is defined as the failure strength and the set of axial strains of the three principal stress directions at the same time step, thus obtaining the input-output mapping structure of the time series sample. The model is constructed based on the input-output mapping structure. By constructing a function that represents the nonlinear mapping relationship between stress path information from the initial loading time to the current time and three-dimensional strain, and initializing the model parameters, an initial three-dimensional strain prediction model is obtained. The model is trained and its parameters are optimized based on the multi-source training dataset and the initial three-dimensional strain prediction model. The model parameters are optimized by constructing a loss function using three-dimensional strain data measured by real experiments in the database as supervision labels, and the trained three-dimensional strain prediction model is obtained.

6. A true triaxial hard rock failure prediction system for deep engineering dynamic disturbance, characterized in that, include: The acquisition module is used to acquire basic information, which includes mechanical response time-series data, failure characteristic information and corresponding loading condition parameters of the true triaxial dynamic disturbance test of deep buried hard rock. The construction module is used to construct a database based on the basic information. By structurally integrating test data under multiple confining pressure and disturbance conditions, a database of true triaxial dynamic disturbance damage in deeply buried hard rock is obtained. The reconstruction module is used to reconstruct the stress sequence according to the loading condition parameters. Based on the multi-stage loading logic, it generates discrete time sequences of the three principal stress directions from the start of loading to the end of disturbance step by step to obtain the simulated three-dimensional principal stress sequence. The alignment module is used to align the time scale of the simulated three-dimensional principal stress sequence with the mechanical response time series data in the deep-buried hard rock true triaxial dynamic disturbance failure database. By unifying the sampling density of the simulated three-dimensional principal stress sequence to the same time resolution as the real experimental data, a multi-source training dataset is obtained. The modeling module is used to construct a prediction model based on the multi-source training dataset. By learning and fitting the nonlinear mapping relationship between the three-dimensional principal stress sequence and the three-dimensional principal strain sequence under multiple working conditions, a three-dimensional strain prediction model is obtained. The prediction module is used to predict the destructive behavior based on the loading parameters of the working condition to be predicted. The prediction result is obtained by inputting the loading parameters of the working condition to be predicted into the three-dimensional strain prediction model.

7. The deep engineering dynamic disturbance true triaxial hard rock failure prediction system according to claim 6, characterized in that, The building module includes: The first construction unit is used to collect raw data based on the mechanical response time series data and the failure characteristic information. By synchronously recording the full process stress-strain time series of the three principal stress directions and the corresponding fracture occurrence time and fracture disturbance number during the test, the raw data of the whole process under each test condition is obtained. The second construction unit is used to extract the effective interval based on the original data of the whole process and the time of rupture. By using the time of rupture as the cutoff boundary to truncate the three-dimensional stress-strain time series in the time domain, the residual stage data after rupture is removed, and the effective three-dimensional stress-strain sequence of each sample is obtained. The third building unit is used to integrate the sample structure according to the effective three-dimensional stress-strain sequence and the corresponding loading condition parameters, fracture occurrence time and fracture disturbance number. By summarizing the test samples under different confining pressure combinations and dynamic disturbance conditions according to a unified data structure, a true triaxial dynamic disturbance failure database for deep buried hard rock is obtained.

8. The deep engineering dynamic disturbance true triaxial hard rock failure prediction system according to claim 6, characterized in that, The reconstruction module includes: The first reconstruction unit is used to generate a three-dimensional synchronous loading segment sequence based on the loading rate in the loading condition parameters. By synchronously and linearly increasing the three principal stress directions at a uniform rate until the minimum principal stress setting value is reached, the principal stress time sequence of the three-dimensional synchronous loading stage is obtained. The second reconstruction unit is used to generate a graded loading segment sequence based on the principal stress time sequence of the three-dimensional synchronous loading stage. By sequentially increasing the maximum principal stress and the intermediate principal stress to the intermediate principal stress set value, and increasing the maximum principal stress alone to the stress level when the disturbance is applied, the principal stress time sequence of the graded loading stage is obtained. The third reconstruction unit is used to generate a disturbance coupling segment sequence based on the principal stress time sequence of the graded loading stage and the disturbance frequency, disturbance amplitude and disturbance duration in the loading condition parameters. By superimposing periodic disturbance stress in the form of a sine wave in the middle or minimum principal stress direction until the disturbance duration ends, a simulated three-dimensional principal stress sequence is obtained.

9. The deep engineering dynamic disturbance true triaxial hard rock failure prediction system according to claim 6, characterized in that, The alignment module includes: The first alignment unit is used to extract the benchmark interval based on the time series step size of each sample in the deep-buried hard rock true triaxial dynamic disturbance failure database. By statistically analyzing the interval between adjacent sampling points of the time series of each real test sample in the database and taking a unified value, the unified benchmark sampling interval corresponding to the real test data is obtained. The second alignment unit is used to align the sampling density with the simulated three-dimensional principal stress sequence according to the unified reference sampling interval. By adjusting the time step when generating the simulated three-dimensional principal stress sequence to be consistent with the unified reference sampling interval, a time-resolution aligned simulated three-dimensional principal stress sequence is obtained. The third alignment unit is used to perform sample-level splicing of the simulated three-dimensional principal stress sequence aligned according to the time resolution and the effective three-dimensional strain sequence of each sample in the database. By pairing and combining the simulated three-dimensional principal stress sequence aligned according to the time resolution with the corresponding real three-dimensional principal strain sequence of each sample based on the time index, a multi-source training dataset is obtained.

10. The deep engineering dynamic disturbance true triaxial hard rock failure prediction system according to claim 6, characterized in that, The modeling module includes: The first modeling unit is used to define the model input and output based on the multi-source training dataset. It defines the input feature vector of each time step as a set containing the stress values ​​of the three principal stress directions, the frequency of dynamic disturbance, the amplitude of dynamic disturbance, and the cumulative number of disturbances, and defines the output target vector as the failure strength and the set of axial strains of the three principal stress directions at the same time step, thus obtaining the input-output mapping structure of the time series sample. The second modeling unit is used to construct a model based on the input-output mapping structure. It constructs a function that represents the nonlinear mapping relationship between stress path information from the initial loading time to the current time and three-dimensional strain, and initializes the model parameters to obtain an initial three-dimensional strain prediction model. The third modeling unit is used to train the model and optimize the parameters based on the multi-source training dataset and the initial three-dimensional strain prediction model. It optimizes the model parameters by constructing a loss function using three-dimensional strain data measured by real experiments in the database as supervision labels, and obtains the trained three-dimensional strain prediction model.

Images