Method for predicting displacement of surrounding rock of high-stress large underground cavern group

Abstract

The invention relates to the technical field of underground cavern surrounding rock displacement prediction, in particular to a high-stress large underground cavern group surrounding rock displacement prediction method. According to the technical scheme, the method comprises the following steps that data collection and arrangement are carried out, and various types of sensors are arranged at different parts of a high-stress large underground cavern group to collect data; the method comprises the following steps: performing preliminary screening, cleaning and normalization processing on collected data, performing feature engineering and data enhancement work, including time-frequency domain feature extraction, wavelet transform-based data decomposition and reconstruction and a data enhancement strategy, and constructing a hybrid neural network model fusing a convolutional neural network and a gating circulation unit. Through multi-source accurate data acquisition and processing, depth feature engineering, hybrid neural network construction training and scientific evaluation and early warning, the accuracy, reliability and timeliness of high-stress large underground cavern group surrounding rock displacement prediction are remarkably improved, powerful support is provided for engineering safety guarantee, and risks and losses are effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of underground cavity surrounding rock displacement prediction, and particularly to a method for predicting the displacement of surrounding rock of a large underground cavity group under high stress. Background Art

[0002] In the prediction of surrounding rock displacement of large underground cavity groups, traditional numerical simulation methods are difficult to accurately depict the complex nonlinear mechanical behavior of surrounding rock and the displacement response law under the coupling action of multiple factors, resulting in a large deviation between the prediction results and the actual engineering situation. For example, when considering the influence of multiple factors such as the heterogeneity of rocks, the random distribution of joints and fissures, the dynamic stress changes during the excavation process, and the coupling of seepage field and stress field, the models established by traditional methods are often too simplified to accurately reflect the true mechanical change process of surrounding rock, thus affecting the accurate evaluation of the stability of surrounding rock of the cavity group and the scientific nature of engineering decisions.

[0003] In large underground cavity group projects (such as large hydropower station underground powerhouses, underground nuclear waste repositories, etc.), the surrounding rock displacement is a key indicator to measure the stability of the cavity. With the continuous expansion of the project scale and the increasing complexity of geological conditions, the accuracy requirements for surrounding rock displacement prediction are getting higher and higher. Traditional numerical simulation methods based on finite element, discrete element, etc. can simulate the mechanical behavior of surrounding rock to a certain extent, but they have limitations in dealing with complex nonlinear problems and multi-factor coupling. For example, when the finite element method simulates the nonlinear constitutive relationship of rock materials, a simplified elastoplastic model is often used, which is difficult to accurately describe the true mechanical properties of rocks under high stress and large deformation; when the discrete element method simulates large-scale jointed rock masses, the calculation efficiency is low and the sensitivity to joint parameters is high, and it is difficult to determine the parameters.

[0004] In recent years, the rapid development of artificial intelligence technology has provided a new way to solve the problem of surrounding rock displacement prediction of large underground cavity groups. Under the traditional monitoring technology, the obtained surrounding rock displacement data are often massive and complex time-series data. For these data, it is difficult to efficiently extract the deep-seated feature information and potential displacement change patterns only by relying on conventional mathematical statistical analysis. For example, in the face of the displacement data fluctuations under different geological conditions, construction processes, and external environmental influences, it is difficult to accurately identify which data features are the key signals indicating the impending instability of the surrounding rock and which are the general changes caused by normal construction disturbances.

[0005] This low-efficiency feature extraction and pattern recognition situation greatly reduces the accuracy and timeliness in using monitoring data for surrounding rock displacement prediction and stability assessment. Therefore, the present application proposes a method for predicting the displacement of surrounding rock of a large underground cavity group under high stress. Summary of the Invention

[0006] The object of the present invention is to address the problem of low efficiency in feature extraction and pattern recognition of underground cavity surrounding rock displacement monitoring data in the background art, and to propose a prediction method for the surrounding rock displacement of large underground cavity groups under high stress.

[0007] Technical solution of the present invention: A prediction method for the surrounding rock displacement of large underground cavity groups under high stress, the following method steps:

[0008] Step 1: Conduct data collection and collation, and arrange various types of sensors at different parts of the large underground cavity group under high stress to collect data;

[0009] Step 2: Conduct preliminary screening, cleaning, and normalization processing on the collected data;

[0010] Step 3: Carry out feature engineering and data augmentation work, including time-frequency domain feature extraction, data decomposition and reconstruction based on wavelet transform, and data augmentation strategies;

[0011] Step 4: Construct a hybrid neural network model integrating a convolutional neural network and a gated recurrent unit, and train and optimize the model;

[0012] Step 5: Evaluate the model using model evaluation metrics, and perform displacement prediction and stability warning on the data to be predicted based on the trained model.

[0013] Optionally, in the data collection and collation, the multi-source data collection and deployment are specifically as follows:

[0014] Arrange sensors at the crown, side walls, floor, and areas with complex geological structures of the large underground cavity group under high stress, where the measurement accuracy of the high-precision laser displacement meter reaches ±0.01 mm, and the sampling frequency is once every 30 minutes; the measurement range of the micro-electromechanical system accelerometer is ±2g, and the sampling frequency is 100 Hz; the accuracy of the fiber Bragg grating strain sensor is ±1 με, and the sampling frequency is once every 1 hour.

[0015] Optionally, in the data collection and collation, the preliminary screening and cleaning of the data include: using the Grubbs criterion to identify and remove outliers in the displacement data, and calculating the mean and standard deviation s, Grubbs statistic:

[0016]

[0017] When G > G 临界值 , remove x i , G 临界值 is obtained from the Grubbs criterion table according to the number n of measurement data and the significance level α = 0.05; use the Savitzky-Golay filtering algorithm to smooth the displacement data, the filtering window m = 5 - 11, and the polynomial fitting degree k = 2 - 4.

[0018] Optionally, in the data collection and sorting step, Min-Max normalization method is used for data normalization, and the formula is:

[0019]

[0020] Map the data to the interval [0, 1], where x 归一化 is the normalized data, x is the original data, x min and x max are the minimum and maximum values of this type of data respectively.

[0021] Optionally, in the feature engineering and data augmentation, the time-frequency domain feature extraction includes: calculating the mean, variance, standard deviation, skewness, kurtosis statistical features and autocorrelation function of the displacement data in the time domain:

[0022]

[0023] where N is the data length, k is the lag order, and the lag order k = 1 - 10; the frequency domain features use the fast Fourier transform to obtain the displacement data spectrum and calculate the power spectral density:

[0024]

[0025] where X(f) is the Fourier transform of x(n).

[0026] Optionally, in the feature engineering and data augmentation, the data decomposition and reconstruction based on wavelet transform uses db4 wavelet to perform multi-scale decomposition on the displacement data, and threshold processing is performed on the high-frequency wavelet coefficients. The threshold σ is the noise standard deviation, N is the data length, and the soft threshold method is used. When |w| > T, w 新 = sgn(w)(|w| - T), where w is the original wavelet coefficient, w 新 is the processed wavelet coefficient, and sgn(w) is the sign function.

[0027] Optionally, in the feature engineering and data augmentation, the data augmentation strategies include: data translation randomly translates the displacement data in the time series, and the translation range is ±10 sampling points; data scaling randomly scales the displacement data, and the scaling factor is between 0.9 - 1.1; data superposition superimposes the displacement data in different time periods, and the weight range is 0.5 - 1.5.

[0028] Optionally, in the model construction and training, the constructed hybrid neural network model is as follows: The CNN part includes two convolutional layers and two pooling layers. The first convolutional layer has a convolutional kernel size of 3×3, 16 convolutional kernels, a ReLU activation function, a stride of 1, and the first pooling layer is a max pooling layer with a pooling kernel size of 2×2 and a stride of 2; the second convolutional layer has a convolutional kernel size of 3×3, 32 convolutional kernels, a ReLU activation function, a stride of 1, and the second pooling layer has a pooling kernel size of 2×2 and a stride of 2; the GRU part includes two GRU layers, each GRU layer has 64 hidden units, and finally a fully connected layer is connected, with 1 neuron and a linear activation function for outputting the predicted value of the surrounding rock displacement.

[0029] Optionally, in the model construction and training, the model training and optimization include: The data set is divided into a training set, a validation set, and a test set according to 70%:20%:10%; the loss function combines the mean square error and the Huber loss. The mean square error:

[0030]

[0031] where n is the number of samples, y i is the actual surrounding rock displacement value, is the predicted value of the surrounding rock displacement output by the prediction model;

[0032] The Huber loss function:

[0033]

[0034] δ = 0.1; The Adam optimization algorithm is used, with an initial learning rate of 0.001, decaying to 0.9 times the original every 10 training epochs, for a total of 200 training epochs; The early stopping method is used, and training stops when the validation set loss function does not decrease for 5 consecutive training epochs.

[0035] Optionally, in the model evaluation and prediction steps: The model evaluation index uses the root mean square error:

[0036]

[0037] The mean absolute error:

[0038]

[0039] The coefficient of determination:

[0040]

[0041] where, is the mean value of the actual surrounding rock displacement. For displacement prediction, the processed data to be predicted is input into the trained model, and the future displacement is predicted according to the time step of the input data. For example, if the data of the past 24 hours is input, the displacement within the next 6 hours can be predicted, and the prediction time step is one prediction value per 1 hour. For stability warning, when the predicted displacement change rate exceeds 0.2 mm / h and the duration exceeds 3 hours, or the predicted cumulative displacement exceeds 10 mm, a warning signal is sent, and the warning signal is notified to relevant personnel through SMS, email or the sound and light alarm of the monitoring system.

[0042] Compared with the prior art, the present application includes at least one of the following beneficial technical effects:

[0043] The present invention collects data from different dimensions through multi-type high-precision sensors, and through scientific screening, cleaning and normalization, provides a reliable and comparable data basis for subsequent analysis, improves data quality, and reduces error interference. By means of time-frequency domain feature extraction, wavelet transform, etc., the internal laws of the data are comprehensively analyzed, the details of displacement changes and potential patterns are accurately captured, and normal and dangerous signals are effectively distinguished, providing a key basis for accurate prediction.

[0044] Furthermore, through a hybrid neural network that combines CNN and GRU, the learning capabilities of spatial and temporal features are integrated to adapt to data complexity; combined with a carefully designed training optimization strategy, the robustness, generalization and training efficiency of the model are enhanced, and the prediction accuracy and stability are improved. The performance of the model is evaluated through multiple indicators, providing a clear direction for improvement; based on the displacement prediction and stability warning with scientific thresholds, timely basis for engineering decision-making is provided, the construction and operation safety of the large underground cavern group is guaranteed, risks and losses are reduced, and the overall efficiency and reliability of the project are improved.

[0045] Through multi-source accurate data acquisition and processing, deep feature engineering, construction and training of a hybrid neural network, and scientific evaluation and warning, the present invention significantly improves the accuracy, reliability and timeliness of the surrounding rock displacement prediction of large underground cavern groups under high stress, provides strong support for engineering safety guarantee, and effectively reduces risks and losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of a method for predicting the surrounding rock displacement of a large underground cavern group under high stress. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] The technical solutions of the present invention will be further described below in conjunction with the drawings and specific embodiments.

[0048] Embodiment 1

[0049] As Figure 1 shown, a method for predicting the surrounding rock displacement of a large underground cavern group under high stress proposed by the present invention is described in detail below for each part of the process.

[0050] I. Data Acquisition and Arrangement

[0051] (1) Multi-source Data Acquisition Deployment

[0052] At different parts of large underground cavern groups under high stress, including the vault, side walls, floor, and areas with complex geological structures, various types of sensors are arranged:

[0053] High-precision laser displacement meter: Its measurement accuracy can reach ±0.01 mm, which is used to accurately measure the linear displacement of surrounding rock. The sampling frequency is set to once every 30 minutes to obtain sufficiently dense displacement data points.

[0054] Microelectromechanical System (MEMS) accelerometer: It can measure the vibration acceleration of surrounding rock, thereby indirectly reflecting the dynamic change characteristics of surrounding rock. Its measurement range is ±2g (g is the acceleration due to gravity), and the sampling frequency is 100 Hz to capture small and rapid vibration signals.

[0055] Fiber Bragg grating strain sensor: It can accurately measure the strain of surrounding rock with an accuracy of ±1 με (microstrain), and the sampling frequency is once every 1 hour. By using the strain data to assist in analyzing the stress state and displacement trend of surrounding rock. By arranging a variety of high-precision sensors (such as high-precision laser displacement meters, Microelectromechanical System accelerometers, Fiber Bragg grating strain sensors) at different key parts, information about the surrounding rock can be obtained from multiple dimensions. The laser displacement meter accurately measures the linear displacement, which can directly reflect the position change of the surrounding rock; the accelerometer captures the vibration acceleration, which helps to understand the response of the surrounding rock under dynamic stress and is of great significance for identifying potential instability factors such as vibrations caused by small rock fractures; the strain sensor provides data from the perspective of stress deformation. These multi-source data complement each other, providing rich materials for comprehensively and accurately analyzing the state of the surrounding rock, and can more accurately depict the complex mechanical behavior of the surrounding rock compared with a single type of sensor.

[0056] Reasonably setting the measurement accuracy and sampling frequency of the sensors not only ensures the accuracy of the data but also takes into account the rationality of the data volume. For example, the ±0.01 mm accuracy of the high-precision laser displacement meter can capture small displacement changes, and the sampling frequency of once every 30 minutes can track the displacement trend at a certain time resolution, without missing important information due to too sparse sampling or causing data redundancy and excessive processing burden due to too dense sampling.

[0057] (2) Preliminary Data Screening and Cleaning

[0058] Outlier Processing:

[0059] Use Grubbs' test to identify and remove outliers in the displacement data. For a set of measurement data x1, x2, …, xn , first calculate its mean value and standard deviation s. Grubbs statistic:

[0060]

[0061] When G > G 临界值 , reject x i , G 临界值 is obtained from the Grubbs criterion table according to the number of measurement data n and the significance level α = 0.05;

[0062] Data smoothing:

[0063] Apply the Savitzky-Golay filtering algorithm to smooth the displacement data. This algorithm uses polynomial fitting to locally smooth the data. For a filtering window of length m, take m = 5 - 11, select according to the data fluctuation situation, use a k-th order polynomial fitting, take k = 2 - 4, and determine the polynomial coefficients by the least squares method to obtain the smoothed displacement data sequence. Using the Grubbs criterion to identify and reject outliers can effectively remove obvious error data caused by factors such as sensor failures and sudden interferences, and improve the reliability of the data. Compared with the traditional simple threshold method for judging outliers, the Grubbs criterion is based on the statistical distribution characteristics of the data, more scientifically determines the abnormal data points, avoids misjudging normal but slightly fluctuating data, thus retaining more real and effective data information and providing a more accurate basis for subsequent analysis.

[0064] The Savitzky-Golay filtering algorithm smooths the displacement data. While removing data noise and fluctuations, it can better retain the original characteristics and trends of the data. Compared with simple methods such as moving average filtering, it uses polynomial fitting to smooth local data, has better adaptability to non-linearly changing data, makes the processed data better reflect the true law of surrounding rock displacement, and reduces the interference of noise on subsequent feature extraction and model training.

[0065] (3) Data normalization

[0066] Use the Min-Max normalization method to map all the collected data to the interval [0, 1]. The formula is:

[0067]

[0068] where, x 归一化 is the normalized data, x is the original data, x max and x minThey are the minimum and maximum values of this type of data respectively. This step helps to improve the efficiency and accuracy of subsequent data processing and model training. Min-Max normalization maps the data to the interval [0,1], enabling data of different types and magnitudes to be compared and processed on the same scale. In subsequent model training, this helps to improve the convergence speed and accuracy of the model. For example, for displacement data and strain data, their original numerical ranges may vary greatly. After normalization, the model can learn the relationships between different data features more evenly during training, avoiding over- or under-learning of certain features due to data magnitude differences, thus enhancing the generalization ability and prediction performance of the model.

[0069] II. Feature Engineering and Data Augmentation

[0070] (I) Time-Frequency Domain Feature Extraction

[0071] Time domain features:

[0072] Calculate the mean, variance, standard deviation, skewness, and kurtosis statistical features of the displacement data. These features can reflect the central tendency, dispersion degree, and distribution pattern of the displacement data in the time domain, helping to initially understand the overall characteristics of the surrounding rock displacement.

[0073] Extract the autocorrelation function (ACF) feature of the displacement data. The autocorrelation function

[0074]

[0075] where N is the data length and k is the lag order, taking k = 1 - 10. Through the autocorrelation function, the correlation of the displacement data at different lag times can be analyzed to mine potential periodic or trend information.

[0076] Frequency domain features:

[0077] Use the fast Fourier transform (FFT) to transform the displacement data into the frequency domain to obtain the spectrum of the displacement data. Calculate the power spectral density (PSD) of the spectrum. The power spectral density Among them, X(f) is the Fourier transform of x(n). The main frequency components in the displacement data can be identified through the power spectral density. The vibrations at certain specific frequencies are related to the natural frequencies of the surrounding rock or the external excitation frequencies, thus providing a basis for judging the stability of the surrounding rock. Time-domain features calculate various statistical quantities (mean, variance, standard deviation, skewness, kurtosis, etc.) and the autocorrelation function. These statistical quantities can describe the central tendency, dispersion degree, distribution pattern of the displacement data and its correlation in the time series from different perspectives. For example, the mean reflects the average level of the displacement, the variance reflects the fluctuation degree of the displacement, and the autocorrelation function can reveal whether there is periodicity or lag correlation in the displacement data. These features help to deeply understand the internal laws of the surrounding rock displacement and provide more bases for judging the stability of the surrounding rock.

[0078] Frequency-domain features obtain the frequency spectrum through the fast Fourier transform and calculate the power spectral density, which can identify the main frequency components in the displacement data. Different geological conditions, construction processes and external environments may cause different responses of the surrounding rock displacement at different frequencies. For example, the natural frequencies of certain geological structures may be obvious in specific frequency bands. When the construction vibration frequency approaches this frequency band, resonance may be triggered, leading to the instability of the surrounding rock. Through frequency-domain analysis, these potential risks can be discovered in advance, providing guidance for taking corresponding preventive measures.

[0079] (2) Data decomposition and reconstruction based on wavelet transform

[0080] The wavelet transform is used to perform multi-scale decomposition on the displacement data. An appropriate wavelet basis function (such as the db4 wavelet) is selected to decompose the displacement data into wavelet coefficients in different frequency bands. For the obtained high-frequency wavelet coefficients, a threshold processing method is used to remove noise interference. The threshold where σ is the noise standard deviation, which can be estimated through statistical analysis of the high-frequency coefficients, N is the data length, and the soft threshold method is used to process the coefficients, that is, when |w|>T, w 新 = sgn(w)(|w|-T), where w is the original wavelet coefficient, w 新 is the processed wavelet coefficient, and sgn(w) is the sign function. Then the processed wavelet coefficients are reconstructed to obtain the denoised displacement data. While retaining the characteristics of the original signal, this data reduces the influence of noise, which helps to more accurately extract the displacement change pattern. The wavelet transform performs multi-scale decomposition on the displacement data, decomposing the data into wavelet coefficients in different frequency bands. This is similar to decomposing a complex signal into detailed information at different levels, and can more carefully analyze the change characteristics of the displacement data at different frequencies and time scales. Different from the traditional Fourier transform that can only analyze the overall frequency components, the wavelet transform has the characteristics of time-frequency localization and has better processing ability for non-stationary displacement data.

[0081] Threshold processing is applied to the high-frequency wavelet coefficients to remove noise interference. The soft threshold method can better retain the edge and detail information of the signal while removing noise. The denoised displacement data obtained after reconstruction has a higher signal-to-noise ratio, can more accurately reflect the true changes in the surrounding rock displacement, reduces the misguidance of noise to subsequent model training and prediction, and improves the model's ability to capture and learn effective signals.

[0082] (III) Data Augmentation Strategy

[0083] To increase the diversity of training data and improve the generalization ability of the model, the following data augmentation methods are adopted:

[0084] Data Translation: Randomly translate the displacement data in the time series, with the translation range being ±10 sampling points, to generate a new displacement data sequence.

[0085] Data Scaling: Randomly scale the displacement data, with the scaling factor between 0.9 - 1.1, to obtain displacement data with different amplitude variations.

[0086] Data Superposition: Superpose the displacement data of different time periods. When superposing, the data can be appropriately weighted according to needs, with the weight range being 0.5 - 1.5, to simulate the displacement combination under different working conditions. Data augmentation methods such as data translation, scaling, and superposition increase the diversity of training data. Data translation simulates the small offsets of displacement data in the time series, similar to the time synchronization errors or small working condition changes that may occur in actual monitoring; data scaling changes the amplitude of the displacement data, which can correspond to external loads or geological condition changes of different intensities; data superposition combines the displacement characteristics of different time periods, simulating the displacement situation under the interaction of complex and variable construction processes and geological environments. Through these data augmentation means, the model can learn more displacement patterns in different scenarios, improve the generalization ability of the model, and enable it to make more accurate predictions when facing various complex and variable situations in actual engineering.

[0087] III. Model Construction and Training

[0088] (I) Construct a Hybrid Neural Network Model

[0089] Construct a hybrid neural network model that integrates a Convolutional Neural Network (CNN) and a Gated Recurrent Unit (GRU):

[0090] CNN Part:

[0091] The CNN contains two convolutional layers and two pooling layers. The convolutional kernel size of the first convolutional layer is 3×3, the number of convolutional kernels is 16, the ReLU activation function is adopted, and the stride is 1; the first pooling layer uses max pooling, the pooling kernel size is 2×2, and the stride is 2. The convolutional kernel size of the second convolutional layer is 3×3, the number of convolutional kernels is 32, the ReLU activation function is also adopted, and the stride is 1; the pooling kernel size of the second pooling layer is 2×2, and the stride is 2. The CNN part is mainly used to extract the spatial local features of displacement data and convert the temporal features of displacement data into a feature map suitable for GRU processing.

[0092] GRU part:

[0093] The GRU network contains two GRU layers, and the number of hidden units in each GRU layer is 64. The GRU layer performs temporal modeling on the feature map output by the CNN, learns the long-term dependence relationship of displacement data in the time series, can effectively process the surrounding rock displacement data with temporality, and capture the trends and patterns of displacement changes. Finally, a fully connected layer is connected, the number of neurons in the fully connected layer is 1, and the linear activation function is adopted to output the predicted value of the surrounding rock displacement. The hybrid neural network model that fuses CNN and GRU gives full play to the advantages of both. The convolutional layer and pooling layer of the CNN can automatically extract the spatial local features of displacement data. For example, it can identify the change patterns, trends in local regions of displacement data, and the spatial correlation relationships between different sensor data, and convert the original temporal displacement data into a feature map with spatial information. The GRU layer is good at processing temporal data, performs temporal modeling on the feature map output by the CNN, and can effectively learn the long-term dependence relationship of displacement data in the time series, capturing features such as the trends and periodicity of displacement changes. This structure combines the spatial feature extraction ability of the CNN and the temporal modeling ability of the GRU, and is more adaptable to the spatio-temporal complexity of surrounding rock displacement data than a single neural network structure, improving the model's learning and prediction ability for displacement change patterns.

[0094] (2) Model training and optimization

[0095] Dataset division:

[0096] The processed data is divided into a training set, a validation set, and a test set according to the ratio of 70%:20%:10%. Ensure that the training set data is rich enough to train the model, the validation set is used to adjust the model hyperparameters, and the test set is used to evaluate the final performance of the model.

[0097] Loss function selection:

[0098] The loss function that combines the mean squared error (MSE) and the Huber loss is adopted. Mean squared error:

[0099]

[0100] Among them, n is the number of samples, and y i is the actual surrounding rock displacement value, is the predicted value of the surrounding rock displacement output by the prediction model, which is used to measure the average error between the predicted value and the actual value. Huber loss function:

[0101]

[0102] Taking δ = 0.1, this function adopts the mean square error form when the error is small and the linear form when the error is large, which can reduce the influence of outliers on the loss calculation and make the model training more stable.

[0103] Optimization algorithm:

[0104] The model is trained using the Adaptive Moment Estimation (Adam) optimization algorithm. The Adam optimization algorithm combines the advantages of the momentum method and Root Mean Square Propagation (RMSProp), can automatically adjust the learning rate, and accelerate the model convergence speed. The initial learning rate is set to 0.001, and a learning rate decay strategy is adopted during training. After every 10 training epochs, the learning rate decays to 0.9 times the original value, and the total number of training epochs is 200.

[0105] Early Stopping:

[0106] During the training process, the early stopping method is used to prevent overfitting. When the loss function on the validation set does not decrease for 5 consecutive training epochs, the training is stopped, and the model parameters at this time are saved as the optimal model parameters.

[0107] A reasonable dataset division (70%:20%:10% divided into the training set, validation set, and test set) ensures that while the model is fully trained, the validation set can be used for hyperparameter tuning and preventing overfitting, and the test set provides an objective evaluation of the final performance of the model.

[0108] The loss function combining the mean square error and the Huber loss uses the mean square error to ensure the accuracy of the model when the error is small, and adopts the Huber loss to reduce the influence of outliers when the error is large, making the model training more stable. Compared with simply using the mean square error, this combined loss function can better handle the possible abnormal situations in the actual data and improve the robustness of the model.

[0109] The Adam optimization algorithm automatically adjusts the learning rate and combines the learning rate decay strategy. It can converge quickly in the initial stage of training, and gradually reduces the learning rate as the training progresses, avoiding the model skipping the optimal solution due to too large a learning rate or falling into a local optimal solution due to too small a learning rate in the later stage of training, improving the training efficiency and accuracy of the model.

[0110] The application of the early stopping method effectively prevents model overfitting. When the loss function on the validation set no longer decreases for five consecutive training epochs, the training is stopped, avoiding the decline in the generalization ability of the model due to overlearning on the training set and ensuring good performance of the model on both the training set and the test set.

[0111] IV. Model Evaluation and Prediction

[0112] (I) Model Evaluation Metrics

[0113] The root mean square error RMSE, mean absolute error MAE, and coefficient of determination R 2 are used as model evaluation metrics:

[0114] Root Mean Square Error It measures the average deviation degree between the predicted value and the actual value. The smaller the RMSE value, the higher the accuracy of the model prediction.

[0115] Mean Absolute Error This metric reflects the average absolute deviation between the predicted value and the actual value, is relatively insensitive to outliers, and can more intuitively reflect the error size of the model prediction.

[0116] Coefficient of Determination where is the mean value of the actual surrounding rock displacement values. The closer it is to 1, the better the fitting effect of the model on the data, that is, the higher the proportion of the data variance that the model can explain, and the better the prediction performance.

[0117] The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) are used as evaluation metrics to comprehensively evaluate the model performance from different perspectives. RMSE is sensitive to the size of the error, can reflect the average deviation degree between the predicted value and the actual value, and reflects the accuracy of the model prediction; MAE is relatively insensitive to outliers and more intuitively reflects the average absolute deviation of the model prediction, which can assist in judging the prediction effect of the model within the normal data range; R 2 measures the goodness of fit of the model to the data, that is, the proportion of the data variance that the model can explain, and reflects the degree to which the model grasps the overall law of the data. Through the comprehensive evaluation of these three metrics, the performance of the model can be understood more comprehensively and accurately, providing a clear direction for the improvement and optimization of the model.

[0118] (II) Prediction and Early Warning

[0119] Displacement Prediction:

[0120] After the displacement data to be predicted is processed through the above data collection and collation, feature engineering, and data augmentation steps, it is input into the trained hybrid neural network model, and the model outputs the corresponding predicted values of the surrounding rock displacement. The time step of the predicted value is related to the time step of the input data. For example, if the input data is the displacement data of the past 24 hours, the sampling frequency is once every 30 minutes, and there are a total of 48 data points, then the displacement change situation within the next 6 hours can be predicted, and the prediction time step is one predicted value per hour.

[0121] Stability warning:

[0122] Set the warning threshold according to the predicted displacement value and displacement change rate. When the predicted displacement change rate exceeds 0.2 mm / h and lasts for more than 3 hours, or the predicted cumulative displacement exceeds 10 mm, a warning signal for the instability of the surrounding rock is issued. The warning signal can be timely notified to relevant engineering personnel through means such as text messages, emails, or the sound and light alarms of the monitoring system, so as to take corresponding reinforcement or construction adjustment measures to ensure the safe construction and operation of the underground cavern group.

[0123] Displacement prediction based on the time step of the input data, such as predicting the displacement within the next 6 hours based on the data of the past 24 hours and the prediction time step being one predicted value per hour, can provide timely and targeted displacement change information for engineering construction and operation. Engineering personnel can arrange the construction progress in advance, take corresponding support measures, or adjust the operation strategy according to these prediction results to ensure the safety of the underground cavern group.

[0124] Setting a reasonable stability warning threshold (such as issuing a warning signal when the displacement change rate exceeds 0.2 mm / h and lasts for more than 3 hours, or the predicted cumulative displacement exceeds 10 mm) can timely detect the risk of surrounding rock instability. Notify relevant personnel through various warning methods (text messages, emails, or the sound and light alarms of the monitoring system), ensure that a rapid response can be made when the surrounding rock may be in a dangerous situation, take emergency reinforcement or personnel evacuation and other measures, effectively avoid the occurrence of safety accidents, and reduce engineering risks and losses.

[0125] In this implementation, data is collected from different dimensions through multi-type high-precision sensors, and after scientific screening, cleaning, and normalization, it provides a reliable and comparable data basis for subsequent analysis, improves data quality, and reduces error interference. Means such as time-frequency domain feature extraction and wavelet transform are used to comprehensively analyze the internal laws of the data, accurately capture the details of displacement changes and potential patterns, effectively distinguish normal and dangerous signals, and provide a key basis for accurate prediction.

[0126] It should be noted that by integrating the spatial and temporal feature learning capabilities through a hybrid neural network that combines CNN and GRU, it can adapt to data complexity. Coupled with a carefully designed training optimization strategy (reasonable loss function, optimization algorithm, early stopping method), it enhances the robustness, generalization ability and training efficiency of the model, and improves the prediction accuracy and stability.

[0127] By evaluating the model performance through multiple indicators, it provides a clear direction for improvement; based on the displacement prediction and stability warning with scientific thresholds, it provides a basis for engineering decisions in a timely manner, ensuring the safety of the construction and operation of the underground cavern group, reducing risks and losses, and enhancing the overall efficiency and reliability of the project.

[0128] The above specific embodiments are only several alternative embodiments of the present invention. Based on the technical solution of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A prediction method for the surrounding rock displacement of a large underground cavern group under high stress, characterized in that It includes the following method steps: Step 1, data collection and arrangement. Deploy various types of sensors at different parts of the large underground cavern group under high stress to collect data; Step 2, conduct preliminary screening, cleaning and normalization processing on the collected data; Step 3, adopt feature engineering and data augmentation processing, which includes time-frequency domain feature extraction, data decomposition and reconstruction based on wavelet transform, and data augmentation strategies; Step 4, construct a hybrid neural network model integrating a convolutional neural network and a gated recurrent unit, and train and optimize this model; Step 5, evaluate the model using model evaluation metrics, and conduct displacement prediction and early warning on the data to be predicted based on the trained model.

2. A prediction method for the surrounding rock displacement of a large underground cavern group under high stress according to claim 1, characterized in that, In the data collection and arrangement, the multi-source data collection and deployment are specifically as follows: Deploy sensors at the crown, side walls, floor and areas with complex geological structures of the large underground cavern group under high stress. Among them, the measurement accuracy of the high-precision laser displacement meter reaches ±0.01 mm, and the sampling frequency is once every 30 minutes; the measurement range of the micro-electromechanical system accelerometer is ±2g, and the sampling frequency is 100Hz; the accuracy of the fiber Bragg grating strain sensor is ±1με, and the sampling frequency is once every 1 hour.

3. A prediction method for surrounding rock displacement of a large underground cavern group under high stress according to claim 2, characterized in that, In the data collection and collation, the preliminary screening and cleaning of data include: identifying and removing outliers in displacement data using the Grubbs criterion, and calculating the mean and the standard deviation s, the Grubbs statistic: When G > G 临界值 exclude x i , G 临界值 is obtained from the Grubbs' criterion table according to the number of measurement data n and the significance level α = 0.05; the Savitzky-Golay filtering algorithm is used to smooth the displacement data, the filtering window m = 5 - 11, and the polynomial fitting degree k = 2 - 4.

4. A prediction method for surrounding rock displacement of a large underground cavern group under high stress according to claim 3, characterized in that, In the data collection and arrangement step, the Min-Max normalization method is adopted for data normalization processing, and the formula is: Map the data to the interval [0, 1], where x 归一化 is the normalized data, x is the original data, x min and x max are the minimum and maximum values of this type of data respectively.

5. The prediction method for surrounding rock displacement of a high-stress large underground cavern group according to claim 1, characterized in that, In the feature engineering and data augmentation, the time-frequency domain feature extraction includes: calculating the mean, variance, standard deviation, skewness, kurtosis statistical features and autocorrelation function of the displacement data in the time domain: where N is the data length, k is the lag order, and the lag order k = 1-10; the frequency domain feature uses the fast Fourier transform to obtain the displacement data spectrum and calculate the power spectral density: where X(f) is the Fourier transform of x(n).

6. A prediction method for surrounding rock displacement of a large underground cavern group under high stress according to claim 5, characterized in that, In the feature engineering and data augmentation, for the data decomposition and reconstruction based on wavelet transform, the db4 wavelet is used to perform multi-scale decomposition on the displacement data, and threshold processing is applied to the high-frequency wavelet coefficients. The threshold where σ is the standard deviation of the noise and N is the data length. The soft thresholding method is adopted. When |w| > T, w 新 = sgn(w)(|w| - T), where w is the original wavelet coefficient, and w 新 is the processed wavelet coefficient, and sgn(w) is the sign function.

7. A prediction method for surrounding rock displacement of a large underground cavern group under high stress according to claim 6, characterized in that, In the feature engineering and data augmentation, the data augmentation strategies include: Data translation randomly translates the displacement data in the time series, and the translation range is ±10 sampling points; Data scaling randomly scales the displacement data, and the scaling factor is between 0.9 and 1.1; Data superposition superimposes the displacement data in different time periods, and the weight range is 0.5-1.

5.

8. A prediction method for surrounding rock displacement of a large underground cavern group under high stress according to claim 1, characterized in that In the model construction and training, the constructed hybrid neural network model is: The CNN part includes two convolutional layers and two pooling layers. The first convolutional layer has a convolutional kernel size of 3×3, 16 convolutional kernels, a ReLU activation function, a stride of 1, the first pooling layer uses max pooling, the pooling kernel size is 2×2, and the stride is 2; the second convolutional layer has a convolutional kernel size of 3×3, 32 convolutional kernels, a ReLU activation function, a stride of 1, the second pooling layer has a pooling kernel size of 2×2, and the stride is 2; the GRU part includes two GRU layers, each GRU layer has 64 hidden units, and finally connects to a fully connected layer, with 1 neuron, and a linear activation function is used to output the predicted value of the surrounding rock displacement.

9. A prediction method for surrounding rock displacement of a large underground cavern group under high stress according to claim 8, characterized in that, In the model construction and training, the model training and optimization include: The data set is divided into a training set, a validation set and a test set according to 70%:20%:10%; the loss function combines the mean square error and the Huber loss, and the mean square error: where n is the number of samples, and y i is the actual surrounding rock displacement value, and is the predicted surrounding rock displacement value output by the prediction model; Huber loss function: δ = 0.1; The Adam optimization algorithm is adopted, with an initial learning rate of 0.001, which decays to 0.9 times the original value every 10 training epochs, for a total of 200 training epochs; The early stopping method is adopted, and the training stops when the validation set loss function does not decrease for 5 consecutive training epochs.

10. A prediction method for surrounding rock displacement of a high-stress large underground cavern group according to claim 1, characterized in that In the model evaluation and prediction steps: The model evaluation index adopts the root mean square error: Mean absolute error: Coefficient of determination: Among them, is the average value of the actual surrounding rock displacement. For displacement prediction, the processed data to be predicted is input into the trained model, and the future displacement is predicted according to the time step of the input data. For example, if the data of the past 24 hours is input, the displacement within the next 6 hours can be predicted, and the prediction time step is one prediction value per 1 hour. The early warning issues a warning signal when the predicted displacement change rate exceeds 0.2 mm / h and lasts for more than 3 hours, or when the predicted cumulative displacement exceeds 10 mm.

Images