Precise forecasting and early warning system for instability failure time of deep well rock mass

Abstract

The invention discloses a deep well rock mass instability failure time accurate forecasting and early warning system, and relates to the technical field of deep well rock mass failure early warning, and the deep well rock mass instability failure time accurate forecasting and early warning system comprises the following steps: collecting a plurality of deep well rock samples to carry out a uniaxial compression experiment, and collecting acoustic emission signals in real time based on an acoustic emission sensor to obtain original acoustic emission signal data; carrying out self-adaptive preprocessing, and carrying out data classification and integration to obtain rock sample acoustic emission parameter data; constructing an original damage forecasting model, and performing model training and verification optimization based on the rock sample acoustic emission parameter data to obtain a rock damage forecasting model; collecting an acoustic emission signal of the deep well rock mass, and predicting the instability failure time of the deep well rock mass in real time based on the rock failure forecasting model; the method is used for solving the problems that when an existing deep well rock mass damage early warning technology carries out early warning on rock mass instability damage through acoustic emission monitoring, single or few characteristic parameters are manually processed and analyzed, and the early warning reliability and timeliness are not high.

Description

Technical Field

[0001] This invention relates to the field of deep well rock mass failure early warning technology, specifically a system for accurate prediction and early warning of the time of deep well rock mass instability and failure. Background Technology

[0002] Deep well rock mass failure early warning technology refers to a comprehensive technical system that integrates monitoring, analysis, modeling and decision-making technologies in deep well environments to identify early signs of rock mass failure such as instability, collapse, spalling and roof fall caused by factors such as stress concentration, geological tectonic activity and excavation disturbance, thereby predicting failure risks in advance and issuing early warning signals to avoid or mitigate engineering disasters.

[0003] Existing deep-well rock mass failure early warning technologies often rely on acoustic emission monitoring to predict rock mass instability and failure. This is because the deep-well environment contains numerous non-rock mass failure-related interference signals, making it difficult to effectively distinguish between valid and interfering signals. This leads to high false alarm rates or the suppression of valid signals, affecting the accuracy of the early warning. Furthermore, traditional acoustic emission analysis often depends on single or a few characteristic parameters, such as peak amplitude, ring count, rise time, and energy. Data analysis and early warning decisions are highly dependent on operator experience, with parameter interpretation often limited to "threshold judgments." However, deep-well rock mass... Destruction is a complex process involving multiple stages and mechanisms, with significant differences in acoustic emission signal characteristics at different stages. Traditional methods struggle to capture these multi-dimensional characteristics, leading to delayed early warnings. For instance, patent application CN112444564A discloses a rock fracture early warning method based on statistical analysis of acoustic emission signals. This method uses a single ring count to warn of rock fracture, resulting in limited accuracy and delayed warnings. Therefore, existing deep-well rock mass failure early warning technologies rely on manual processing and analysis of single or a few characteristic parameters when using acoustic emission monitoring to warn of rock mass instability, leading to low reliability and timeliness of early warnings. Summary of the Invention

[0004] This invention aims to at least partially solve one of the technical problems in the prior art. It involves collecting various deep-well rock samples for uniaxial compression experiments and acquiring acoustic emission signals in real time using acoustic emission sensors to obtain raw acoustic emission signal data. The raw acoustic emission signal data undergoes adaptive preprocessing and data classification and integration to obtain rock sample acoustic emission parameter data. A raw failure prediction model is constructed, and based on the rock sample acoustic emission parameter data, the model is trained, validated, and optimized to obtain a rock failure prediction model. Acoustic emission signals from deep-well rock masses are collected, and the instability and failure time of the deep-well rock mass is predicted in real time based on the rock failure prediction model. This addresses the problem that existing deep-well rock mass failure early warning technologies rely on manual processing and analysis of single or a few characteristic parameters when using acoustic emission monitoring to predict rock mass instability and failure, resulting in low reliability and timeliness of the early warning.

[0005] To achieve the above objectives, this application provides a precise prediction and early warning system for the instability and failure time of deep well rock masses, including an experimental simulation module, a data processing module, a model construction module, and a rock mass early warning module;

[0006] The experimental simulation module is used to collect various deep well rock samples for uniaxial compression experiments, and to collect acoustic emission signals in real time based on acoustic emission sensors to obtain raw acoustic emission signal data.

[0007] The data processing module includes a preprocessing unit and a classification and integration unit. The preprocessing unit is used to perform adaptive preprocessing on the original acoustic emission signal data, and the classification and integration unit is used to classify and integrate the data to obtain rock sample acoustic emission parameter data.

[0008] The model building module is used to build the original failure prediction model, and to train and validate the model based on the acoustic emission parameter data of the rock sample to obtain the rock failure prediction model.

[0009] The rock mass early warning module is used to collect acoustic emission signals of deep well rock masses and to predict the instability and failure time of deep well rock masses in real time based on the rock failure prediction model.

[0010] Furthermore, the test simulation module is configured with test simulation strategies, which include:

[0011] Rock samples are collected from deep wells and processed into standard specimens according to experimental specifications. These specimens are designated as deep well rock samples, and any type of deep well rock sample is designated as the first rock sample.

[0012] The acoustic emission sensor was tightly attached to the surface of the first rock sample, and a uniaxial compression test was performed on the first rock sample until the first rock sample was destroyed.

[0013] Furthermore, the experimental simulation strategy also includes:

[0014] During the uniaxial compression experiment, the stress magnitude of the first rock sample was monitored and collected in real time, and the collection time was recorded as the stress change data of the first rock sample. At the same time, the acoustic emission signal of the first rock sample was monitored and recorded using an acoustic emission sensor, which was recorded as the original acoustic emission data of the first rock sample.

[0015] The maximum stress in the stress variation data of the first rock sample is marked as the ultimate stress of the first rock sample, and the sampling time corresponding to the ultimate stress is recorded as the rock sample failure point of the first rock sample.

[0016] Repeatedly acquire stress variation data and raw acoustic emission data for all deep well rock samples.

[0017] Furthermore, the preprocessing unit is configured with a preprocessing strategy, which includes:

[0018] For the acoustic emission signal of the first rock sample, denoted as the first acoustic emission signal, calculate the average value of all data points of the first acoustic emission signal, and subtract the average value of all data points from each data point in the first acoustic emission signal to obtain the second acoustic emission signal.

[0019] The second acoustic emission signal is decomposed into k2 scales by selecting a Morlet wavelet with a center frequency of k1. The modulus value of the complex wavelet coefficients of each scale is taken, and the integral of the square of the modulus value is calculated to obtain the energy of the corresponding scale. Here, k1 is the set frequency and k2 is the set number.

[0020] Calculate the proportion of energy at each scale to the total energy across all scales, denoted as the energy percentage. Retain the k3 scales with the highest energy percentages, denoted as the effective scales, where k3 is the number of scales set. <k2。

[0021] Furthermore, the preprocessing strategy also includes:

[0022] For any effective scale, denoted as the first scale, the modulus of the complex wavelet coefficients of the first scale is taken to obtain the envelope sequence corresponding to the first scale, denoted as the first envelope sequence.

[0023] Calculate the mean of the first envelope sequence, denoted as AU. Denote k4*AU as the peak threshold. Denote the points in the first envelope sequence that exceed the peak threshold as valid peak points. Count the number of valid peak points and the total number of points in the first envelope sequence, and denote them as BS and AS0 respectively. Calculate the sparsity factor XS corresponding to the first scale, where XS = BS / AS0.

[0024] Obtain the standard deviation of the first envelope sequence, denoted as BU, and calculate the suppression threshold YZ, where YZ = R0 * BU, R0 = k4 * -10 * (k4 - 1) * XS, where k4 is the maximum set multiple;

[0025] For points in the first envelope sequence that are less than the suppression threshold, the corresponding wavelet coefficients are set to zero; for points in the first envelope sequence that are greater than or equal to the suppression threshold, the corresponding wavelet coefficients are retained; the corrected envelope sequence corresponding to the first scale is then completed.

[0026] Repeatedly acquire the corrected envelope sequences corresponding to all effective scales. For each corrected envelope sequence, perform inverse mapping using the corresponding Morlet wavelet, align all the effective signal components obtained by inverse mapping along the time axis, and superimpose them to synthesize the denoised transmission signal of the first acoustic transmission signal, denoted as the first denoised signal.

[0027] Repeatedly acquire the denoised emission signals corresponding to the acoustic emission signals of all deep well rock samples.

[0028] Furthermore, the classification and integration unit is configured with a classification and integration strategy, which includes:

[0029] For the first denoised signal of the first rock sample, the first time interval is set to t1; based on t1 and with the rock sample failure point as the endpoint, multiple time points are selected from the first denoised signal and recorded as parameter sampling points, and any one of the parameter sampling points is recorded as the first sampling point.

[0030] Obtain the stress magnitude corresponding to the first sampling point, denoted as parameter 1; obtain the amplitude and frequency of the first sampling point, and denot them as parameter 2 and parameter 3 in sequence, respectively.

[0031] Using the first sampling point as the endpoint of the first denoised signal, the corresponding number of AE events, ring count, rise time, energy entropy, RA value, AE event rate, and ring event rate are obtained and recorded as parameters 4-10 in sequence; parameters 1-10 are recorded as the parameter information of the first sampling point.

[0032] Repeatedly acquire parameter information from all parameter sampling points and arrange them in chronological order according to the time corresponding to the parameter sampling points, and record them as rock sample acoustic emission parameter data.

[0033] Furthermore, the model building module is configured with model building strategies, which include:

[0034] Based on long short-term memory networks, an original damage prediction model is constructed, which includes an input layer, a hidden layer, and an output layer.

[0035] For parameters 1-10 of the first denoised signal, normalization is performed according to parameter type, and the results are denoted as standard parameters 1-10. Any one of the standard parameters 2-9 is denoted as standard parameter i. Standard parameter 1 and standard parameter i at the same parameter sampling point are combined to form the corresponding single-parameter feature vector, denoted as DM = {standard parameter 1, standard parameter i}. All single-parameter feature vectors are arranged in the corresponding time order and denoted as single-parameter feature data.

[0036] The single-parameter feature data is divided into a single-parameter training set and a single-parameter test set in a 7:3 ratio. The output result of the output layer is set as the stress magnitude. The original damage prediction model is trained using the single-parameter training set. After completion, the model is tested using the single-parameter test set, and the root mean square error of the model is calculated and denoted as the RMSE corresponding to the standard parameter i.

[0037] Repeatedly obtain the RMSE corresponding to all standard parameters and arrange them in ascending order, which is called the error sequence. Reorder the corresponding standard parameters according to the order of the error sequence and record them as important parameters 1-9.

[0038] Furthermore, model building strategies also include:

[0039] Let the combination of standard parameter 1 and important parameter 1-j be denoted as parameter combination j, where j∈[1,9], and obtain all parameter combinations;

[0040] For parameter combination j, the standard parameter 1 and the important parameter 1-j at the same parameter sampling point are combined into the corresponding parameter feature vector, denoted as XM = {standard parameter 1, important parameter 1, ..., important parameter j}; and all parameter feature vectors are arranged in the corresponding time order, denoted as the j-th combination feature data.

[0041] The feature data of the j-th combination are divided into a combination training set and a combination test set in a 7:3 ratio. The original damage prediction model is trained using the combination training set to obtain the j-th prediction model. The j-th prediction model is then tested using the combination test set, and the model's determination coefficient, mean determination error, root mean square error, and mean absolute percentage error are calculated and denoted as the evaluation parameters of parameter combination j.

[0042] Furthermore, model building strategies also include:

[0043] Repeatedly obtain the evaluation parameters for all parameter combinations, and arrange the average determination error, root mean square error, and mean absolute percentage error of all parameter combinations in descending order, and record them as error sequence 1, error sequence 2, and error sequence 3 respectively. Arrange the determination coefficients of all parameter combinations in ascending order, and record them as error sequence 4.

[0044] Obtain the position index of the evaluation parameter of parameter combination j in the corresponding error sequence 1-4, and sum them up to record the overall score of parameter combination j;

[0045] Repeatedly obtain the overall score of all parameter combinations, and record the parameter combination with the largest overall score as the predicted parameter combination of the first rock sample. Mark the prediction model corresponding to the predicted parameter combination as the rock failure prediction model of the first rock sample, and record the mean absolute percentage error corresponding to the rock failure prediction model as PE.

[0046] Repeatedly obtain the predicted parameter combinations, rock failure prediction models, and PE for all deep rock samples.

[0047] Furthermore, the rock mass early warning module includes a rock mass early warning strategy, which includes:

[0048] The stress magnitude and acoustic emission signal of the deep well rock mass are collected and recorded as the current stress magnitude and current acoustic emission signal, respectively. The current acoustic emission signal is then subjected to adaptive preprocessing to obtain the current denoised signal.

[0049] Multiple time points are uniformly selected from the current denoised signal according to the first time interval, and the parameters corresponding to the combination of prediction parameters are obtained at the time points and recorded as the current parameters. All current parameters and current stress magnitudes are normalized according to parameter type and combined into corresponding parameter feature vectors. Then, according to the type of deep well rock mass, the corresponding rock failure prediction model is input to obtain stress prediction data for future moments.

[0050] Multiply the predicted stress magnitude at each future time point by (1+PE) to obtain the corrected stress magnitude at each future time point. Search the corrected stress magnitudes sequentially according to time order, and record the first corrected stress magnitude that is greater than or equal to the corresponding ultimate stress as the instability failure point. Based on the time point corresponding to the instability failure point, obtain the predicted time of instability failure of the deep well rock mass, and issue a forecast and early warning.

[0051] The beneficial effects of this invention are as follows: This invention collects various deep-well rock samples for uniaxial compression experiments and acquires acoustic emission signals in real time based on acoustic emission sensors to obtain raw acoustic emission signal data; adaptive preprocessing is performed on the raw acoustic emission signal data, and data classification and integration are performed to obtain rock sample acoustic emission parameter data; an original failure prediction model is constructed, and the model is trained, validated, and optimized based on the rock sample acoustic emission parameter data to obtain a rock failure prediction model; acoustic emission signals of deep-well rock masses are collected, and the instability and failure time of deep-well rock masses are predicted in real time based on the rock failure prediction model; multiple characteristic parameters of acoustic emission signals can be automatically processed and analyzed without relying on manual intervention, improving the reliability and timeliness of early warning;

[0052] This invention utilizes Morlet wavelet decomposition and the sparsity factor of the decomposed scale to set an adaptive threshold for denoising at each scale. Its advantage lies in leveraging the physical characteristics of acoustic emission signals to achieve adaptive denoising, automatically adjusting the denoising threshold to ensure denoising effectiveness while improving robustness. Furthermore, by first training the model with single parameters and ranking them by RMSE, and then progressively combining parameters to train the model separately and evaluating them against multiple indicators, the predictive power of each parameter can be quantitatively obtained, avoiding errors from subjective selection. This allows for the selection of the optimal parameter combination for predictive power, avoiding errors caused by invalid or redundant parameters, and improving the reliability of early warnings. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the system of the present invention;

[0054] Figure 2 This is a flowchart illustrating the steps of the method of the present invention;

[0055] Figure 3 This is a flowchart of the preprocessing strategy of the present invention;

[0056] Figure 4 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0057] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1, please refer to Figure 1 As shown, this application provides a precise prediction and early warning system for the instability and failure time of deep well rock mass, including an experimental simulation module, a data processing module, a model construction module, and a rock mass early warning module;

[0059] The experimental simulation module is used to collect various deep well rock samples for uniaxial compression experiments, and to collect acoustic emission signals in real time based on acoustic emission sensors to obtain raw acoustic emission signal data;

[0060] The test simulation module is equipped with a test simulation strategy, which includes: collecting rock samples from deep wells and processing them into standard specimens according to experimental specifications, which are designated as deep well rock samples, and any type of deep well rock sample is designated as the first rock sample; processing natural rocks from different sources into standard specimens with the same geometric dimensions and surface morphology can eliminate test deviations caused by size and shape, ensure uniform stress during loading, and avoid eccentricity. The standard specimens are generally cylindrical in shape.

[0061] The acoustic emission sensor was tightly attached to the surface of the first rock sample, and a uniaxial compression test was performed on the first rock sample until it was destroyed. The tight attachment was to maximize signal transmission efficiency, ensure the authenticity of signal characteristics, reduce energy attenuation, reduce environmental interference, and improve the signal-to-noise ratio.

[0062] During the uniaxial compression experiment, the stress magnitude of the first rock sample was monitored and collected in real time, and the collection time was recorded as the stress change data of the first rock sample. At the same time, the acoustic emission signal of the first rock sample was monitored and recorded using an acoustic emission sensor, which was recorded as the original acoustic emission data of the first rock sample.

[0063] The maximum stress in the stress variation data of the first rock sample is marked as the ultimate stress of the first rock sample, and the acquisition time corresponding to the ultimate stress is recorded as the failure point of the first rock sample; the ultimate stress serves as an identification label for whether subsequent rock mass failure will occur.

[0064] Repeatedly acquire stress variation data and raw acoustic emission data for all deep well rock samples;

[0065] In practical implementation, the uniaxial compression test of rock samples is one of the most basic and commonly used tests in the field of rock mechanics. By applying slowly increasing pressure along the axis of the rock sample, the entire process from stress to deformation and eventual failure is observed to obtain the basic mechanical properties and failure laws of the rock. Its core is to simulate the mechanical response of the rock under uniaxial stress, providing key parameters for geotechnical engineering design, geological disaster early warning, etc. Other destructive mechanics tests can also be selected according to the actual application scenario.

[0066] The data processing module includes a preprocessing unit and a classification and integration unit. The preprocessing unit is used to perform adaptive preprocessing on the raw acoustic emission signal data, and the classification and integration unit is used to classify and integrate the data to obtain the acoustic emission parameter data of the rock sample.

[0067] The preprocessing unit is configured with a preprocessing strategy, which includes: (See details) Figure 3 As shown, for the acoustic emission signal of the first rock sample, denoted as the first acoustic emission signal, the average value of all data points of the first acoustic emission signal is calculated, and the average value of all data points is subtracted from each data point in the first acoustic emission signal to obtain the second acoustic emission signal. The purpose is to remove the DC component of the first acoustic emission signal, that is, the constant offset in the signal. In the deep well environment, the acoustic emission sensor may be affected by temperature drift and cable vibration, resulting in baseline drift or instantaneous large amplitude noise. These interferences are unrelated to microcracks, and removing them can avoid their interference with subsequent decomposition.

[0068] Select a Morlet wavelet with a central frequency of k1 to decompose the second acoustic emission signal into k2 scales. Take the modulus value of the complex wavelet coefficients for each scale, and then calculate the integral of the squared modulus value to obtain the energy corresponding to the scale. Here, k1 is the set frequency and k2 is the set number; in this embodiment, k1 = 200 kHz because the main frequency of microcrack acoustic emission in deep well rock samples is mostly between 100 - 400 kHz, k2 = 3, and k2 is generally 3 - 5. k1 and k2 can be flexibly set according to the actual application scenario;

[0069] Calculate the proportion of the energy of each scale to the total energy of all scales, denoted as the energy proportion. Retain the k3 scales with the largest energy proportion, denoted as the effective scales. k3 is the set number, and k3 < k2; in this embodiment, k3 = 2; Focus on the frequency band where the effective signals are most concentrated and automatically eliminate the low-energy noise scales.

[0070] For any one of the effective scales, denoted as the first scale, take the modulus value of the complex wavelet coefficients of the first scale to obtain the envelope sequence corresponding to the first scale, denoted as the first envelope sequence;

[0071] Calculate the mean value of the first envelope sequence, denoted as AU. Denote k4 * AU as the peak threshold. Denote the points in the first envelope sequence that exceed the peak threshold as effective peak points. Count the number of effective peak points and the total number of points in the first envelope sequence, denoted as BS and AS0 in sequence, and calculate the sparsity factor XS corresponding to the first scale, where XS = BS / AS0; The sparsity factor represents the proportion of effective peak points to the total signal length and quantifies the density of transient pulses in the signal; When XS is small, it means the signal is mainly continuous noise, with few and sparse transient pulses; When XS is large, it means the signal is mainly transient pulses, with frequent microcrack activities and dense pulses;

[0072] XS provides a quantitative basis for subsequent threshold adjustment: avoiding the problems of excessive denoising of sparse signals and insufficient denoising of dense signals by traditional fixed thresholds, and making the denoising strategy adapt to the dynamic changes of the signal.

[0073] Obtain the standard deviation of the first envelope sequence, denoted as BU, and calculate the suppression threshold YZ. Here, YZ = R0 * BU, and R0 = k4 * -10 * (k4 - 1) * XS, where k4 is the set maximum multiple; in this embodiment, k4 = 3 and can be flexibly set; The envelope amplitude of continuous noise is usually low and stable, while the amplitude of weak pulses may be close to that of noise. By dynamically adjusting the threshold through XS, a balance can be found between noise suppression and pulse retention;

[0074] For points in the first envelope sequence that are less than the suppression threshold, the corresponding wavelet coefficients are set to zero; for points in the first envelope sequence that are greater than or equal to the suppression threshold, the corresponding wavelet coefficients are retained; the corrected envelope sequence corresponding to the first scale is completed; that is, the impulse components exceeding the threshold are retained, and the noise below the threshold is suppressed;

[0075] Repeatedly acquire the corrected envelope sequences corresponding to all effective scales. For each corrected envelope sequence, perform inverse mapping using the corresponding Morlet wavelet, align all the effective signal components obtained by inverse mapping along the time axis, and superimpose them to synthesize the denoised transmission signal of the first acoustic transmission signal, denoted as the first denoised signal.

[0076] Repeatedly acquire the denoised emission signals corresponding to the acoustic emission signals of all deep well rock samples.

[0077] The classification and integration unit is configured with a classification and integration strategy, which includes: for the first denoised signal of the first rock sample, setting a first time interval as t1; based on t1 and with the rock sample failure point as the endpoint, selecting multiple time points from the first denoised signal and recording them as parameter sampling points, and recording any one of the parameter sampling points as the first sampling point; in this embodiment, t1 = 0.1s. A smaller t1 provides finer-grained time series information, but increases the sample dimension and computational load; a larger t1 is coarser but more robust and can be flexibly set.

[0078] The stress magnitude corresponding to the first sampling point is obtained and recorded as parameter 1. The amplitude and frequency of the first sampling point are obtained and recorded as parameter 2 and parameter 3 in sequence, respectively, reflecting the intensity of a single event and the type of rupture.

[0079] Using the first sampling point as the endpoint of the first denoised signal, the corresponding number of AE events, ring count, rise time, energy entropy, RA value, AE event rate, and ring event rate are obtained and recorded as parameters 4-10 in sequence. These parameters are used to characterize the event density, waveform morphology, and energy distribution, which usually show significant changes before destruction. Parameters 1-10 are recorded as the parameter information of the first sampling point. The selection of parameters can be changed or increased or decreased according to the actual application scenario.

[0080] Repeatedly acquire parameter information from all parameter sampling points and arrange them in the time sequence corresponding to the parameter sampling points, and record them as rock sample acoustic emission parameter data;

[0081] In practical implementation, the oscillation characteristics of Morlet wavelets match the time-frequency characteristics of acoustic emission transient pulses, effectively capturing the start and end times and amplitude changes of the pulses. The envelope of transient pulses is characterized by steep peaks over a short period of time, while the envelope of continuous noise is characterized by stable fluctuations over a long period of time. After decomposition, the two can be intuitively distinguished in the envelope domain.

[0082] The model building module is used to build the original failure prediction model, and to train and validate the model based on the acoustic emission parameter data of the rock sample to obtain the rock failure prediction model.

[0083] The model building module is configured with a model building strategy, which includes: building an original damage prediction model based on a long short-term memory network. The original damage prediction model includes an input layer, a hidden layer, and an output layer. The long short-term memory network is specifically designed to learn the long-term and short-term dependencies of sequence data, such as the cumulative effect of cracks from initiation to convergence. It is more suitable than traditional static regression for predicting the time of destruction, which is strongly correlated with historical behavior.

[0084] For parameters 1-10 of the first denoised signal, normalization is performed according to parameter type. Different parameters have different dimensions and large differences in numerical range. Normalization can avoid errors caused by excessive numerical differences and excessive increases or decreases during the change process. After completion, they are recorded as standard parameters 1-10. Any one of the standard parameters 2-9 is recorded as standard parameter i. Standard parameter 1 and standard parameter i at the same parameter sampling point are combined to form the corresponding single-parameter feature vector, denoted as DM = {standard parameter 1, standard parameter i}. All single-parameter feature vectors are arranged in the corresponding time order and recorded as single-parameter feature data. The independent predictive ability of each standard parameter is evaluated one by one. By pairing stress with a single standard parameter, training the long short-term memory network and calculating the error, the predictive ability brought by each standard parameter alone can be quantified.

[0085] The single-parameter feature data is divided into a single-parameter training set and a single-parameter test set in a 7:3 ratio. The output result of the output layer is set as the stress magnitude. The original failure prediction model is trained using the single-parameter training set. After training, the model is tested using the single-parameter test set, and the root mean square error (RMSE) of the model is calculated and denoted as the RMSE corresponding to the standard parameter i. The RMSE serves as a measure of the predictive ability of a single standard parameter; the smaller the RMSE, the stronger the correlation between the standard parameter and stress. The ratio of the single-parameter training set to the single-parameter test set can be flexibly set.

[0086] Repeatedly obtain the RMSE corresponding to all standard parameters and arrange them in ascending order, which is called the error sequence. Reorder the corresponding standard parameters according to the order of the error sequence and record them as important parameters 1-9.

[0087] The combination of standard parameter 1 and important parameter 1-j is denoted as parameter combination j, where j∈[1,9]. All parameter combinations are obtained; that is, the combination is gradually accumulated from a single important parameter to multiple important parameters. For example, parameter combination 1 is standard parameter 1 + important parameter 1, parameter combination 2 is standard parameter 1 + important parameter 1 + important parameter 2, and parameter combination 4 is standard parameter 1 + important parameter 1 + important parameter 2 + important parameter 3 + important parameter 4. By gradually accumulating, it can help identify which feature parameters truly bring gains and avoid blindly adding redundant and noisy features.

[0088] For parameter combination j, the standard parameter 1 and the important parameter 1-j at the same parameter sampling point are combined into the corresponding parameter feature vector, denoted as XM = {standard parameter 1, important parameter 1, ..., important parameter j}; and all parameter feature vectors are arranged in the corresponding time order, denoted as the j-th combination feature data.

[0089] The feature data of the j-th combination are divided into a combination training set and a combination test set in a 7:3 ratio. The original damage prediction model is trained using the combination training set to obtain the j-th prediction model. The j-th prediction model is tested using the combination test set, and the model's determination coefficient, mean determination error, root mean square error, and mean absolute percentage error are calculated and denoted as the evaluation parameters of parameter combination j. Starting from the most important feature, the less important features are gradually added, and the marginal improvement of the model performance at each step is observed to find effective and non-redundant parameter combinations.

[0090] Repeatedly obtain the evaluation parameters for all parameter combinations, and arrange the mean decision error, root mean square error, and mean absolute percentage error of all parameter combinations in descending order, and denote them as error sequence 1, error sequence 2, and error sequence 3. Arrange the determination coefficients of all parameter combinations in ascending order, and denote them as error sequence 4. The smaller the mean decision error, root mean square error, and mean absolute percentage error, the better the predictive performance of the model; the larger the determination coefficient, the better the predictive performance of the model.

[0091] The position numbers of the evaluation parameters of parameter combination j in the corresponding error sequences 1-4 are obtained and accumulated to form the overall score of parameter combination j. The positions of the four different ranking results are accumulated to quantify the predictive performance of different parameter combinations, forming a selection criterion to avoid selecting combinations with one extremely good indicator but other poor indicators.

[0092] The overall score of all parameter combinations is repeatedly obtained, and the parameter combination with the largest overall score is recorded as the predicted parameter combination of the first rock sample. The prediction model corresponding to the predicted parameter combination is marked as the rock failure prediction model of the first rock sample. The mean absolute percentage error corresponding to the rock failure prediction model is recorded as PE. PE is used to correct the output results of the model in the future, so as to reduce the risk of missed warnings and improve the timeliness of warnings.

[0093] Repeatedly acquire all deep rock samples, including the predicted parameter combinations, rock failure prediction models, and PE.

[0094] In the specific implementation process, the input layer is used to receive the raw sequence data and convert it into a format that can be processed by the Long Short-Term Memory network. The number of units and layers in the hidden layer can be flexibly adjusted according to the actual application scenario to match the input complexity and sample size. The output layer is used to convert the temporal features output by the hidden layer into results that meet the task objectives.

[0095] The rock mass early warning module is used to collect acoustic emission signals of deep well rock masses and to predict the instability and failure time of deep well rock masses in real time based on the rock failure prediction model.

[0096] The rock mass early warning module includes a rock mass early warning strategy, which includes: collecting the stress magnitude and acoustic emission signal of the deep well rock mass, recording them as the current stress magnitude and current acoustic emission signal respectively; performing adaptive preprocessing on the current acoustic emission signal to obtain the current denoised signal; and synchronizing the stress acquisition and acoustic emission signal acquisition times for subsequent sampling to extract and combine parameters based on time points.

[0097] Multiple time points are uniformly selected from the current denoised signal based on the first time interval, and the parameters corresponding to the predicted parameter combination are obtained at each time point and recorded as the current parameters. All current parameters and current stress magnitudes are normalized according to parameter type and combined into corresponding parameter feature vectors. Then, according to the type of deep well rock mass, the corresponding rock failure prediction model is input to obtain the stress prediction data for future time. That is, the parameter set is exactly the same as the predicted parameter combination. The normalization process maintains the consistency of numerical scale to match the corresponding model and avoids prediction errors caused by scale drift.

[0098] The predicted stress magnitudes at each future time point are multiplied by (1+PE) to obtain the corrected stress magnitudes at each future time point. The corrected stress magnitudes are then retrieved sequentially according to time. The first corrected stress magnitude greater than or equal to the corresponding ultimate stress is recorded as the instability failure point. Based on the time point corresponding to the instability failure point, the predicted time of instability failure of the deep well rock mass is obtained, and a forecast and early warning are issued. Multiplying the stress magnitude by (1+PE) is to compensate for the average percentage underestimation that has occurred in the model in the past, aiming to reduce the risk of missed reports and improve the timeliness of forecasts and early warnings.

[0099] In practical implementation, PE is the average absolute percentage error obtained from the statistical analysis of the model's output results during model testing. When using it, it is necessary to confirm that the PE corresponding to this model is used, rather than the average absolute percentage error of other models.

[0100] Example 2, please refer to Figure 2 As shown, this application provides a method for accurate prediction and early warning of the instability and failure time of deep well rock masses, including the following steps:

[0101] Step S1 involves collecting various deep-well rock samples for uniaxial compression experiments and acquiring acoustic emission signals in real time using an acoustic emission sensor to obtain raw acoustic emission signal data. Step S1 includes the following sub-steps:

[0102] Step S101: Collect rock samples from the deep well and process them into standard specimens according to the experimental specifications. These specimens are recorded as deep well rock samples, and any type of deep well rock sample is recorded as the first rock sample.

[0103] In step S102, the acoustic emission sensor is tightly attached to the surface of the first rock sample, and a uniaxial compression test is performed on the first rock sample until the first rock sample is destroyed.

[0104] Step S103: During the uniaxial compression experiment, the stress magnitude of the first rock sample is monitored and collected in real time, and the collection time is recorded as the stress change data of the first rock sample. At the same time, the acoustic emission signal of the first rock sample is monitored and recorded using an acoustic emission sensor, and recorded as the original acoustic emission data of the first rock sample.

[0105] Step S104: Mark the maximum stress in the stress change data of the first rock sample as the ultimate stress of the first rock sample, and record the acquisition time corresponding to the ultimate stress as the rock sample failure point of the first rock sample.

[0106] Step S105: Repeat the acquisition of stress change data and raw acoustic emission data for all deep well rock samples.

[0107] Step S2 involves adaptive preprocessing of the raw acoustic emission signal data and data classification and integration to obtain rock sample acoustic emission parameter data. Step S2 includes the following sub-steps:

[0108] Step S201: For the acoustic emission signal of the first rock sample, denoted as the first acoustic emission signal, calculate the average value of all data points of the first acoustic emission signal, and subtract the average value of all data points from each data point in the first acoustic emission signal to obtain the second acoustic emission signal.

[0109] Step S202: Select a Morlet wavelet with a center frequency of k1 to decompose the second acoustic emission signal into k2 scales. Take the modulus value of the complex wavelet coefficients of each scale, and then calculate the integral of the square of the modulus value to obtain the energy of the corresponding scale. Here, k1 is the set frequency and k2 is the set number.

[0110] Step S203: Calculate the proportion of energy at each scale to the total energy at all scales, denoted as the energy percentage. Retain the k3 scales with the largest energy percentages, denoted as the effective scales, where k3 is the number of scales set. <k2。

[0111] Step S204: For any effective scale, denoted as the first scale, take the modulus of the complex wavelet coefficients of the first scale to obtain the envelope sequence corresponding to the first scale, denoted as the first envelope sequence;

[0112] Step S205: Calculate the mean of the first envelope sequence, denoted as AU. Denote k4*AU as the peak threshold. Denote the points in the first envelope sequence that exceed the peak threshold as effective peak points. Count the number of effective peak points and the total number of points in the first envelope sequence, and denote them as BS and AS0 respectively in order. Calculate the sparsity factor XS corresponding to the first scale, where XS = BS / AS0.

[0113] Step S206: Obtain the standard deviation of the first envelope sequence, denoted as BU, and calculate the suppression threshold YZ, where YZ = R0 * BU, R0 = k4 * -10 * (k4 - 1) * XS, where k4 is the maximum set multiple;

[0114] Step S207: For points in the first envelope sequence that are less than the suppression threshold, set the corresponding wavelet coefficients to zero; for points in the first envelope sequence that are greater than or equal to the suppression threshold, retain the corresponding wavelet coefficients; the corrected envelope sequence corresponding to the first scale after completion;

[0115] Step S208: Repeatedly obtain the corrected envelope sequence corresponding to all effective scales. For each corrected envelope sequence, perform inverse mapping using the corresponding Morlet wavelet, and align all the effective signal components obtained by inverse mapping according to the time axis. Superimpose and synthesize them to obtain the denoised transmission signal of the first acoustic transmission signal, denoted as the first denoised signal.

[0116] Step S209: Repeatedly acquire the denoised emission signals corresponding to the acoustic emission signals of all deep well rock samples.

[0117] Step S210: For the first denoised signal of the first rock sample, set the first time interval to t1; based on t1 and with the rock sample failure point as the endpoint, select multiple time points from the first denoised signal and record them as parameter sampling points, and record any parameter sampling point as the first sampling point;

[0118] Step S211: Obtain the stress magnitude corresponding to the first sampling point and record it as parameter 1; obtain the amplitude and frequency of the first sampling point and record them as parameter 2 and parameter 3 in sequence.

[0119] Step S212: Taking the first sampling point as the endpoint of the first denoised signal, obtain the corresponding number of AE events, ring count, rise time, energy entropy, RA value, AE event rate and ring event rate, and record them as parameters 4-10 in order; record parameters 1-10 as the parameter information of the first sampling point.

[0120] Step S213: Repeatedly acquire parameter information for all parameter sampling points and arrange them in the time sequence corresponding to the parameter sampling points, and record them as rock sample acoustic emission parameter data.

[0121] Step S3 involves constructing the original failure prediction model and training and optimizing the model based on the acoustic emission parameter data of the rock samples to obtain the rock failure prediction model. Step S3 includes the following sub-steps:

[0122] Step S301: Based on the Long Short-Term Memory Network, construct the original damage prediction model, which includes an input layer, a hidden layer, and an output layer.

[0123] Step S302: For parameters 1-10 of the first denoised signal, normalize them according to their types and record them as standard parameters 1-10. Record any one of the standard parameters 2-9 as standard parameter i. Combine standard parameter 1 and standard parameter i at the same parameter sampling point to form a corresponding single-parameter feature vector, denoted as DM = {standard parameter 1, standard parameter i}. Arrange all single-parameter feature vectors in the corresponding time order and record them as single-parameter feature data.

[0124] Step S303: Divide the single-parameter feature data into a single-parameter training set and a single-parameter test set in a 7:3 ratio. Set the output result of the output layer to the stress magnitude. Use the single-parameter training set to train the original damage prediction model. After completion, use the single-parameter test set to test the model and calculate the root mean square error of the model, denoted as the RMSE corresponding to the standard parameter i.

[0125] Step S304: Repeatedly obtain the RMSE corresponding to all standard parameters, arrange them in ascending order and record them as error sequences. Reorder the corresponding standard parameters according to the error sequence and record them as important parameters 1-9 in sequence.

[0126] Step S305: Denote the combination of standard parameter 1 and important parameter 1-j as parameter combination j, where j∈[1,9], and obtain all parameter combinations;

[0127] Step S306: For parameter combination j, combine the standard parameter 1 and the important parameter 1-j at the same parameter sampling point into the corresponding parameter feature vector, denoted as XM = {standard parameter 1, important parameter 1, ..., important parameter j}; and arrange all parameter feature vectors in the corresponding time order, denoted as the j-th combination feature data;

[0128] Step S307: Divide the feature data of the j-th combination into a combination training set and a combination test set in a 7:3 ratio. Use the combination training set to train the original damage prediction model and obtain the j-th prediction model. Use the combination test set to test the j-th prediction model and calculate the model's determination coefficient, mean determination error, root mean square error, and mean absolute percentage error, which are recorded as the evaluation parameters of parameter combination j.

[0129] Step S308: Repeatedly obtain the evaluation parameters for all parameter combinations, and arrange the average determination error, root mean square error and mean absolute percentage error of all parameter combinations in descending order, and record them as error sequence 1, error sequence 2 and error sequence 3 respectively. Arrange the determination coefficients of all parameter combinations in ascending order, and record them as error sequence 4.

[0130] Step S309: Obtain the position number of the evaluation parameter of parameter combination j in the corresponding error sequence 1-4, and accumulate them to record the overall score of parameter combination j;

[0131] Step S310: Repeatedly obtain the overall score of all parameter combinations, and record the parameter combination with the largest overall score as the predicted parameter combination of the first rock sample, mark the prediction model corresponding to the predicted parameter combination as the rock failure prediction model of the first rock sample, and record the mean absolute percentage error corresponding to the rock failure prediction model as PE.

[0132] Step S311: Repeat the acquisition of prediction parameter combinations, rock failure prediction models, and PE for all deep rock samples.

[0133] Step S4 involves acquiring acoustic emission signals from the deep well rock mass and predicting the instability and failure time of the deep well rock mass in real time based on a rock failure prediction model. Step S4 includes the following sub-steps:

[0134] Step S401: Collect the stress magnitude and acoustic emission signal of the deep well rock mass, and record them as the current stress magnitude and current acoustic emission signal respectively. Perform adaptive preprocessing on the current acoustic emission signal to obtain the current denoised signal.

[0135] Step S402: Select multiple time points uniformly from the current denoised signal according to the first time interval, and obtain the parameters corresponding to the prediction parameter combination at the time point, which are recorded as the current parameters. Normalize all current parameters and current stress magnitude according to parameter type, and combine them into corresponding parameter feature vectors. Then, input the corresponding rock failure prediction model according to the type of deep well rock mass to obtain stress prediction data for future moments.

[0136] Step S403: Multiply the predicted stress magnitude at each future time point by (1+PE) to obtain the corrected stress magnitude at each future time point. Search the corrected stress magnitudes in chronological order and record the first corrected stress magnitude that is greater than or equal to the corresponding ultimate stress as the instability failure point. Based on the time point corresponding to the instability failure point, obtain the predicted time of instability failure of the deep well rock mass and issue a forecast warning.

[0137] Example 3, please refer to Figure 4 As shown, Figure 4 A schematic diagram of an electronic device is provided, which may include a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The memory stores computer-readable instructions, and the processor can call these instructions. When the processor executes a computer-readable instruction, it performs steps similar to those in a method for accurate prediction and early warning of instability and failure time in deep well rock masses, to achieve the following functions: collecting various deep well rock samples for uniaxial compression experiments and acquiring acoustic emission signals in real time based on acoustic emission sensors to obtain raw acoustic emission signal data; performing adaptive preprocessing on the raw acoustic emission signal data and classifying and integrating the data to obtain rock sample acoustic emission parameter data; constructing an original failure prediction model and training and optimizing the model based on the rock sample acoustic emission parameter data to obtain a rock failure prediction model; acquiring acoustic emission signals from deep well rock masses and predicting the instability and failure time of deep well rock masses in real time based on the rock failure prediction model.

[0138] Furthermore, when the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0139] Example 4: This application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it performs the steps of the above-mentioned method for accurate prediction and early warning of instability and failure time of deep well rock mass, to achieve the following functions: collecting various deep well rock samples for uniaxial compression experiments, and acquiring acoustic emission signals in real time based on acoustic emission sensors to obtain raw acoustic emission signal data; performing adaptive preprocessing on the raw acoustic emission signal data, and classifying and integrating the data to obtain rock sample acoustic emission parameter data; constructing an original failure prediction model, and training and verifying the model based on the rock sample acoustic emission parameter data to obtain a rock failure prediction model; acquiring acoustic emission signals of deep well rock mass, and predicting the instability and failure time of deep well rock mass in real time based on the rock failure prediction model.

[0140] Based on the above description of the embodiments, the embodiments of the present invention can be provided as methods, systems, or computer program products. Based on this understanding, the technical solutions described above, or the parts that contribute to the prior art, can be embodied in the form of software products. These computer software products can be stored in computer-readable storage media, such as ROM / RAM, magnetic disks, optical disks, etc., and include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or certain parts of the embodiments.

[0141] In the embodiments provided in this application, it should be understood that the disclosed system or method can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of modules or units is only a logical functional division, and there may be other division methods in actual implementation. Furthermore, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces. The indirect coupling or communication connection between systems, modules, and units may be electrical, mechanical, or other forms.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A precise prediction and early warning system for the instability and failure time of deep well rock masses, characterized in that, It includes an experimental simulation module, a data processing module, a model building module, and a rock mass early warning module; The experimental simulation module is used to collect various deep well rock samples for uniaxial compression experiments, and to collect acoustic emission signals in real time based on acoustic emission sensors to obtain raw acoustic emission signal data. The data processing module includes a preprocessing unit and a classification and integration unit. The preprocessing unit is used to perform adaptive preprocessing on the original acoustic emission signal data, and the classification and integration unit is used to classify and integrate the data to obtain rock sample acoustic emission parameter data. The model building module is used to build the original failure prediction model, and to train and validate the model based on the acoustic emission parameter data of the rock sample to obtain the rock failure prediction model. The rock mass early warning module is used to collect acoustic emission signals of deep well rock masses and to predict the instability and failure time of deep well rock masses in real time based on the rock failure prediction model.

2. The accurate prediction and early warning system for the instability and failure time of deep well rock mass according to claim 1, characterized in that, The test simulation module is configured with test simulation strategies, which include: Rock samples are collected from deep wells and processed into standard specimens according to experimental specifications. These specimens are designated as deep well rock samples, and any type of deep well rock sample is designated as the first rock sample. The acoustic emission sensor was tightly attached to the surface of the first rock sample, and a uniaxial compression test was performed on the first rock sample until the first rock sample was destroyed.

3. The accurate prediction and early warning system for deep well rock mass instability and failure time according to claim 2, characterized in that, Experimental simulation strategies also include: During the uniaxial compression experiment, the stress magnitude of the first rock sample was monitored and collected in real time, and the collection time was recorded as the stress change data of the first rock sample. At the same time, the acoustic emission signal of the first rock sample was monitored and recorded using an acoustic emission sensor, which was recorded as the original acoustic emission data of the first rock sample. The maximum stress in the stress variation data of the first rock sample is marked as the ultimate stress of the first rock sample, and the sampling time corresponding to the ultimate stress is recorded as the rock sample failure point of the first rock sample. Repeatedly acquire stress variation data and raw acoustic emission data for all deep well rock samples.

4. The accurate prediction and early warning system for deep well rock mass instability and failure time according to claim 3, characterized in that, The preprocessing unit is configured with a preprocessing strategy, which includes: For the acoustic emission signal of the first rock sample, denoted as the first acoustic emission signal, calculate the average value of all data points of the first acoustic emission signal, and subtract the average value of all data points from each data point in the first acoustic emission signal to obtain the second acoustic emission signal. The second acoustic emission signal is decomposed into k2 scales by selecting a Morlet wavelet with a center frequency of k1. The modulus value of the complex wavelet coefficients of each scale is taken, and the integral of the square of the modulus value is calculated to obtain the energy of the corresponding scale. Here, k1 is the set frequency and k2 is the set number. Calculate the proportion of energy at each scale to the total energy across all scales, denoted as the energy percentage. Retain the k3 scales with the highest energy percentages, denoted as the effective scales, where k3 is the number of scales set. <k2。 5. The accurate prediction and early warning system for deep well rock mass instability and failure time according to claim 4, characterized in that, Preprocessing strategies also include: For any effective scale, denoted as the first scale, the modulus of the complex wavelet coefficients of the first scale is taken to obtain the envelope sequence corresponding to the first scale, denoted as the first envelope sequence. Calculate the mean of the first envelope sequence, denoted as AU. Denote k4*AU as the peak threshold. Denote the points in the first envelope sequence that exceed the peak threshold as valid peak points. Count the number of valid peak points and the total number of points in the first envelope sequence, and denote them as BS and AS0 respectively. Calculate the sparsity factor XS corresponding to the first scale, where XS = BS / AS0. Obtain the standard deviation of the first envelope sequence, denoted as BU, and calculate the suppression threshold YZ, where YZ = R0 * BU, R0 = k4 * -10 * (k4 - 1) * XS, where k4 is the maximum set multiple; For points in the first envelope sequence that are less than the suppression threshold, the corresponding wavelet coefficients are set to zero; for points in the first envelope sequence that are greater than or equal to the suppression threshold, the corresponding wavelet coefficients are retained; the corrected envelope sequence corresponding to the first scale is then completed. Repeatedly acquire the corrected envelope sequences corresponding to all effective scales. For each corrected envelope sequence, perform inverse mapping using the corresponding Morlet wavelet, align all the effective signal components obtained by inverse mapping along the time axis, and superimpose them to synthesize the denoised transmission signal of the first acoustic transmission signal, denoted as the first denoised signal. Repeatedly acquire the denoised emission signals corresponding to the acoustic emission signals of all deep well rock samples.

6. The accurate prediction and early warning system for the instability and failure time of deep well rock mass according to claim 5, characterized in that, The category integration unit is configured with category integration strategies, which include: For the first denoised signal of the first rock sample, the first time interval is set to t1; based on t1 and with the rock sample failure point as the endpoint, multiple time points are selected from the first denoised signal and recorded as parameter sampling points, and any one of the parameter sampling points is recorded as the first sampling point. Obtain the stress magnitude corresponding to the first sampling point, denoted as parameter 1; obtain the amplitude and frequency of the first sampling point, and denot them as parameter 2 and parameter 3 in sequence, respectively. Using the first sampling point as the endpoint of the first denoised signal, the corresponding number of AE events, ring count, rise time, energy entropy, RA value, AE event rate, and ring event rate are obtained and recorded as parameters 4-10 in sequence; parameters 1-10 are recorded as the parameter information of the first sampling point. Repeatedly acquire parameter information from all parameter sampling points and arrange them in chronological order according to the time corresponding to the parameter sampling points, and record them as rock sample acoustic emission parameter data.

7. The accurate prediction and early warning system for deep well rock mass instability and failure time according to claim 6, characterized in that, The model building module is configured with model building strategies, which include: Based on long short-term memory networks, an original damage prediction model is constructed, which includes an input layer, a hidden layer, and an output layer. For parameters 1-10 of the first denoised signal, normalization is performed according to parameter type, and the results are denoted as standard parameters 1-10. Any one of the standard parameters 2-9 is denoted as standard parameter i. Standard parameter 1 and standard parameter i at the same parameter sampling point are combined to form the corresponding single-parameter feature vector, denoted as DM = {standard parameter 1, standard parameter i}. All single-parameter feature vectors are arranged in the corresponding time order and denoted as single-parameter feature data. The single-parameter feature data is divided into a single-parameter training set and a single-parameter test set in a 7:3 ratio. The output result of the output layer is set as the stress magnitude. The original damage prediction model is trained using the single-parameter training set. After completion, the model is tested using the single-parameter test set, and the root mean square error of the model is calculated and denoted as the RMSE corresponding to the standard parameter i. Repeatedly obtain the RMSE corresponding to all standard parameters and arrange them in ascending order, which is called the error sequence. Reorder the corresponding standard parameters according to the order of the error sequence and record them as important parameters 1-9.

8. The accurate prediction and early warning system for deep well rock mass instability and failure time according to claim 7, characterized in that, Model building strategies also include: Let the combination of standard parameter 1 and important parameter 1-j be denoted as parameter combination j, where j∈[1,9], and obtain all parameter combinations; For parameter combination j, the standard parameter 1 and the important parameter 1-j at the same parameter sampling point are combined into the corresponding parameter feature vector, denoted as XM = {standard parameter 1, important parameter 1, ..., important parameter j}; and all parameter feature vectors are arranged in the corresponding time order, denoted as the j-th combination feature data. The feature data of the j-th combination are divided into a combination training set and a combination test set in a 7:3 ratio. The original damage prediction model is trained using the combination training set to obtain the j-th prediction model. The j-th prediction model is then tested using the combination test set, and the model's determination coefficient, mean determination error, root mean square error, and mean absolute percentage error are calculated and denoted as the evaluation parameters of parameter combination j.

9. The accurate prediction and early warning system for the instability and failure time of deep well rock mass according to claim 8, characterized in that, Model building strategies also include: Repeatedly obtain the evaluation parameters for all parameter combinations, and arrange the average determination error, root mean square error, and mean absolute percentage error of all parameter combinations in descending order, and record them as error sequence 1, error sequence 2, and error sequence 3 respectively. Arrange the determination coefficients of all parameter combinations in ascending order, and record them as error sequence 4. Obtain the position index of the evaluation parameter of parameter combination j in the corresponding error sequence 1-4, and sum them up to record the overall score of parameter combination j; Repeatedly obtain the overall score of all parameter combinations, and record the parameter combination with the largest overall score as the predicted parameter combination of the first rock sample. Mark the prediction model corresponding to the predicted parameter combination as the rock failure prediction model of the first rock sample, and record the mean absolute percentage error corresponding to the rock failure prediction model as PE. Repeatedly obtain the predicted parameter combinations, rock failure prediction models, and PE for all deep rock samples.

10. A precise prediction and early warning system for the instability and failure time of deep well rock masses according to claim 9, characterized in that, The rock mass early warning module includes rock mass early warning strategies, which include: The stress magnitude and acoustic emission signal of the deep well rock mass are collected and recorded as the current stress magnitude and current acoustic emission signal, respectively. The current acoustic emission signal is then subjected to adaptive preprocessing to obtain the current denoised signal. Multiple time points are uniformly selected from the current denoised signal according to the first time interval, and the parameters corresponding to the combination of prediction parameters are obtained at the time points and recorded as the current parameters. All current parameters and current stress magnitudes are normalized according to parameter type and combined into corresponding parameter feature vectors. Then, according to the type of deep well rock mass, the corresponding rock failure prediction model is input to obtain stress prediction data for future moments. Multiply the predicted stress magnitude at each future time point by (1+PE) to obtain the corrected stress magnitude at each future time point. Search the corrected stress magnitudes sequentially according to time order, and record the first corrected stress magnitude that is greater than or equal to the corresponding ultimate stress as the instability failure point. Based on the time point corresponding to the instability failure point, obtain the predicted time of instability failure of the deep well rock mass, and issue a forecast and early warning.

Images