Methods for Monitoring and Predicting the Impact Range of Blasting Vibration in Long Tunnels in Environmentally Sensitive Areas

Abstract

This invention provides a method for monitoring and predicting the impact range of blasting vibration in long tunnels in environmentally sensitive areas. It relates to the field of tunnel engineering blasting safety monitoring and vibration control technology. The method includes constructing a three-dimensional dynamic monitoring network, collecting multi-source data based on the network and preprocessing it to obtain a blasting vibration feature matrix, building a prediction model based on an improved LSTM model, training the prediction model based on a preset blasting vibration feature matrix and its corresponding blasting parameters and geological conditions, and finally inputting the blasting vibration feature matrix to be detected, along with its corresponding blasting parameters and geological conditions, into the trained prediction model to obtain the prediction result. This invention significantly improves the accuracy and real-time performance of blasting vibration monitoring and prediction, effectively solving the vibration control problem in the blasting construction of long tunnels in environmentally sensitive areas. It features high prediction accuracy, strong real-time performance, strong adaptability, and high data utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering blasting safety monitoring and vibration control technology, and in particular to a method for monitoring blasting vibration and predicting the impact range of long tunnels in environmentally sensitive areas. Background Technology

[0002] Tunnel blasting is one of the main methods of tunnel excavation, but the resulting vibrations can adversely affect the surrounding environment (such as buildings, cultural relics, and residential areas). Currently, tunnel blasting vibration monitoring mainly relies on traditional empirical formula methods (Sadowski formula, Hopkinson bar model), which have the following shortcomings:

[0003] 1. Low prediction accuracy: Traditional methods do not consider the dynamic coupling effect between geological conditions and blasting parameters, resulting in large prediction errors; they cannot meet the needs of environmentally sensitive areas for millimeter-level vibration control.

[0004] 2. Poor real-time performance: Existing methods are difficult to predict the vibration impact range in real time; and there is a lack of real-time graded early warning mechanism for sensitive targets.

[0005] 3. Weak adaptability: Traditional models rely on historical data and have poor adaptability to new geological sections; they cannot dynamically adjust model parameters to adapt to complex geological conditions.

[0006] 4. Insufficient data utilization: Existing methods do not make full use of multi-source data (such as vibration, acoustic emission, and stress data); there is a lack of in-depth mining and fusion analysis of monitoring data.

[0007] Therefore, it is essential to design a method for monitoring blasting vibration and predicting the impact range of long tunnels in environmentally sensitive areas. Summary of the Invention

[0008] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide a method for monitoring blasting vibration and predicting the impact range of long tunnels in environmentally sensitive areas.

[0009] To achieve the above objectives, the present invention provides the following solution:

[0010] This invention provides a method for monitoring blasting vibration and predicting the impact range of long tunnels in environmentally sensitive areas, including:

[0011] Step 1: Construct a three-dimensional dynamic monitoring network;

[0012] Step 2: Collect multi-source data based on a three-dimensional dynamic monitoring network and preprocess it to obtain the blasting vibration characteristic matrix;

[0013] Step 3: Build a prediction model based on the improved LSTM model. Train the prediction model based on the preset blasting vibration feature matrix and its corresponding blasting parameters and geological conditions to obtain the trained prediction model.

[0014] Step 4: Input the blasting vibration feature matrix to be detected, along with its corresponding blasting parameters and geological conditions, into the trained prediction model to obtain the prediction results.

[0015] Preferably, in step 1, constructing a three-dimensional dynamic monitoring network specifically involves:

[0016] The monitoring terminals are deployed at intervals along the tunnel axis by a drone swarm. Dual redundant monitoring terminals are set up around sensitive targets. Real-time data transmission is achieved based on LoRa and 5G dual-mode communication. The monitoring terminals are equipped with vibration sensors, acoustic emission sensors and stress sensors.

[0017] Preferably, in step 2, multi-source data is collected based on a three-dimensional dynamic monitoring network and preprocessed to obtain the blasting vibration characteristic matrix, specifically as follows:

[0018] Step 201: Collect the original vibration signal based on the vibration sensor, collect the original acoustic emission signal based on the acoustic emission sensor, and collect the original stress signal based on the stress sensor;

[0019] Step 202: Perform noise reduction processing on the acquired original vibration signal, original acoustic emission signal and original stress signal;

[0020] Step 203: Perform data normalization and time alignment on the vibration signal, acoustic emission signal and stress signal after signal denoising;

[0021] Step 204: Extract the characteristics of the processed vibration signal, acoustic emission signal and stress signal, and construct the blasting vibration characteristic matrix.

[0022] Preferably, in step 202, the acquired original vibration signal, original acoustic emission signal, and original stress signal are subjected to signal noise reduction processing, specifically as follows:

[0023] The original vibration signal, original acoustic emission signal, and original stress signal are denoised using an improved wavelet threshold denoising algorithm. The specific process is as follows:

[0024] The original vibration signal, original acoustic emission signal and original stress signal are discretized, and the wavelet is decomposed into 4 levels to obtain the approximate components and detail components of each level, as well as the amplitude of the wavelet detail signal of that level.

[0025] Calculate the noise standard deviation and adaptive threshold of each layer based on the detailed signal coefficients of each layer;

[0026] The signal coefficients of each layer are processed using a threshold quantization function to complete threshold noise reduction at each scale.

[0027] Perform inverse wavelet transform on the components processed at each layer to obtain the denoised components, thus completing wavelet reconstruction.

[0028] Preferably, in step 204, the features of the processed vibration signal, acoustic emission signal, and stress signal are extracted to construct a blasting vibration feature matrix, specifically as follows:

[0029] For vibration signals, extract their vibration characteristics, including peak particle vibration velocity, dominant frequency, and vibration duration;

[0030] For acoustic emission signals, their acoustic emission characteristics are extracted, including acoustic emission event counts, acoustic emission energy, and acoustic emission amplitude;

[0031] For stress signals, extract their stress characteristics, including peak stress, stress change rate, and stress recovery time;

[0032] Based on vibration characteristics, acoustic emission characteristics, and stress characteristics, the blasting vibration characteristic matrix F is constructed as follows:

[0033] F = [PPV, dominant frequency, duration, acoustic emission event count, acoustic emission energy, acoustic emission amplitude, peak stress, stress change rate, stress recovery time].

[0034] Preferably, in step 3, a prediction model is built based on the improved LSTM model. The prediction model is trained based on the preset blasting vibration feature matrix and its corresponding blasting parameters and geological conditions to obtain the trained prediction model, specifically as follows:

[0035] A prediction model is built based on the LSTM-SA model, which consists of a Self-Attention computation module, an LSTM module, and a fully connected neural network.

[0036] During its training process, the SGDR learning rate adjustment algorithm is used for training.

[0037] Preferably, the blasting parameters include charge quantity, resistance line, blasting method and delay time, and the geological conditions include rock mass grade, fracture density, rock mass wave velocity and groundwater conditions.

[0038] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0039] This invention provides a method for monitoring blasting vibration and predicting the impact range of long tunnels in environmentally sensitive areas. The method includes: constructing a three-dimensional dynamic monitoring network; collecting multi-source data based on the three-dimensional dynamic monitoring network and preprocessing it to obtain a blasting vibration feature matrix; building a prediction model based on an improved LSTM model; training the prediction model based on a preset blasting vibration feature matrix and its corresponding blasting parameters and geological conditions to obtain a trained prediction model; and inputting the blasting vibration feature matrix to be detected, along with its corresponding blasting parameters and geological conditions, into the trained prediction model to obtain the prediction result. This invention has the following advantages:

[0040] 1. Improved prediction accuracy: By using multi-source data fusion and combining multi-modal data such as vibration, acoustic emission, and stress, a blasting vibration feature matrix is ​​constructed. Through data fusion, the input quality of the prediction model is improved, the model can be dynamically corrected, and the error is significantly reduced. It can still maintain high prediction accuracy under complex geological conditions.

[0041] 2. Real-time monitoring and early warning are achieved. Real-time data acquisition is possible. LoRa+5G dual-mode communication is used to achieve real-time transmission of monitoring data. The vibration impact range prediction results can be updated every 5 seconds. A hierarchical early warning mechanism is adopted: a three-level early warning system is established, which can trigger early warning signals in a timely manner.

[0042] 3. Improved data utilization efficiency: Multimodal data fusion was adopted to integrate multi-source data such as vibration, acoustic emission, and stress, and a blasting vibration feature matrix was constructed. By extracting key features, the value of the data was deeply mined, and the data quality was improved through noise reduction processing.

[0043] 4. It can protect environmentally sensitive targets, accurately predict the range of vibration impact, avoid damage to sensitive targets, is easy to implement, reduces construction costs, and can reduce unnecessary vibration reduction measures. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0046] Figure 2 A schematic diagram illustrating the improved wavelet thresholding noise reduction process;

[0047] Figure 3 This is a schematic diagram of the LSTM-SA model structure. Detailed Implementation

[0048] 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.

[0049] The purpose of this invention is to provide a method for monitoring blasting vibration and predicting the impact range of long tunnels in environmentally sensitive areas. This method significantly improves the accuracy and real-time performance of blasting vibration monitoring and prediction, effectively solves the vibration control problem in the blasting construction of long tunnels in environmentally sensitive areas, and features high prediction accuracy, strong real-time performance, strong adaptability, and high data utilization efficiency.

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a method for monitoring blasting vibration and predicting the impact range of long tunnels in environmentally sensitive areas, including:

[0052] Step 1: Construct a three-dimensional dynamic monitoring network;

[0053] Step 2: Collect multi-source data based on a three-dimensional dynamic monitoring network and preprocess it to obtain the blasting vibration characteristic matrix;

[0054] Step 3: Build a prediction model based on the improved LSTM model. Train the prediction model based on the preset blasting vibration feature matrix and its corresponding blasting parameters and geological conditions to obtain the trained prediction model.

[0055] Step 4: Input the blasting vibration feature matrix to be detected, along with its corresponding blasting parameters and geological conditions, into the trained prediction model to obtain the prediction results.

[0056] In step 1, a three-dimensional dynamic monitoring network is constructed, specifically as follows:

[0057] The monitoring terminals are deployed at intervals along the tunnel axis by a drone swarm. Dual redundant monitoring terminals are set up around sensitive targets. Real-time data transmission is achieved based on LoRa and 5G dual-mode communication. The monitoring terminals are equipped with vibration sensors, acoustic emission sensors and stress sensors.

[0058] In step 2, multi-source data is collected based on a three-dimensional dynamic monitoring network and preprocessed to obtain the blasting vibration characteristic matrix, specifically:

[0059] Step 201: Collect the original vibration signal based on the vibration sensor, collect the original acoustic emission signal based on the acoustic emission sensor, and collect the original stress signal based on the stress sensor;

[0060] Step 202: Perform noise reduction processing on the acquired original vibration signal, original acoustic emission signal and original stress signal;

[0061] Step 203: Perform data normalization and time alignment on the vibration signal, acoustic emission signal and stress signal after signal denoising;

[0062] Step 204: Extract the characteristics of the processed vibration signal, acoustic emission signal and stress signal, and construct the blasting vibration characteristic matrix.

[0063] In step 202, the acquired original vibration signal, original acoustic emission signal, and original stress signal are subjected to signal noise reduction processing, specifically as follows:

[0064] This paper describes a method for denoising the original vibration signal, original acoustic emission signal, and original stress signal based on an improved wavelet threshold denoising algorithm. The improved wavelet threshold denoising algorithm is first introduced below.

[0065] Wavelet thresholding is a time-scale analysis and denoising method for signals. It features multi-resolution analysis and the ability to represent local signal characteristics in both the time and frequency domains. It is a time-frequency localization analysis method with a fixed window size but adjustable shape, where both the time and frequency windows can be changed. That is, it has lower time resolution and higher frequency resolution in the low-frequency part, and higher time resolution and lower frequency resolution in the high-frequency part. It is very suitable for analyzing non-stationary signals and extracting local signal characteristics. The wavelet coefficients generated by the signal contain important information. After wavelet decomposition, the wavelet coefficients of the signal are larger, while those of the noise are smaller. The wavelet coefficients of the noise should be smaller than those of the signal. By selecting an appropriate threshold, wavelet coefficients larger than the threshold are considered useful signals and should be retained, while those smaller than the threshold are noise and set to zero. Finally, reconstruction is used to achieve the purpose of denoising.

[0066] The principle behind it will be explained as follows:

[0067] A typical one-dimensional noisy signal can be represented as f(t) = s(t) + n(t), where s(t) is the useful signal and n(t) follows the expression N(0,σ). 2In most cases, the signal is concentrated in the low-frequency part, while the noise is distributed in the high-frequency part. Therefore, after wavelet decomposition, the wavelet transform coefficients of the microseismic signal are greater than those of the noise. In the processing, a suitable threshold is generally selected. When the wavelet coefficients are less than this value, it indicates that the wavelet is generated by noise; conversely, it is mainly caused by the signal. Threshold denoising is performed on the sub-band coefficients with distributed noise. The high-frequency components after denoising are then reconstructed by inverse wavelet transform with the low-frequency components concentrated in the signal. This effectively removes noise interference unrelated to the microseismic signal. Therefore, the basic steps of wavelet threshold denoising are:

[0068] 1. Decomposition: Select the wavelet and determine the number of wavelet decomposition levels;

[0069] 2. Thresholding stage: Selecting a threshold and threshold function to perform nonlinear processing on the wavelet coefficients;

[0070] 3. Reconstruction: The signal is reconstructed using the new wavelet coefficients to obtain the filtered signal;

[0071] The selection of the wavelet threshold is specifically as follows:

[0072] The four commonly used thresholds in wavelet thresholding analysis include fixed-form thresholding, adaptive thresholding, heuristic thresholding, and minimax thresholding.

[0073] For a fixed threshold, the threshold value in this method is a constant, and its formula is as follows:

[0074]

[0075] In the formula, σ is the standard deviation of the noise signal, and N is the signal length;

[0076] The threshold is always fixed, but the maximum magnitude of the noise signal decreases as the number of decomposition layers increases. If the same threshold is applied to different decomposition layers, too much useful signal will be filtered out in the low-frequency coefficients, while some noise will be retained in the high-frequency coefficients.

[0077] Adaptive thresholding based on Stein unbiased likelihood estimation:

[0078] Take the absolute value of each element in the signal s(i), arrange them in order, and then square each element to obtain a new signal sequence:

[0079] f(k) = [sort(|s|)] 2 (k = 0, 1, ..., N-1)

[0080] If we take the square root of the kth element of the threshold f(k), that is...

[0081]

[0082] The risk generated by this threshold is

[0083]

[0084] Based on the obtained risk curve Rish(k), the value corresponding to its minimum risk point is denoted as k. min Therefore, the adaptive threshold is defined as:

[0085]

[0086] The computational load increases during the value selection process, the operation time is long, and the noise reduction effect is generally poor.

[0087] Heuristic thresholding, a method that compromises between fixed and adaptive thresholds, is determined by the following formula:

[0088]

[0089] In the formula: N is the signal length, and S is the number of coefficients;

[0090] 4. Minimize the maximum threshold, and its value is determined by the following formula:

[0091]

[0092] In the formula, σ = median(|x|) / 0.6745;

[0093] This method also belongs to the fixed threshold method. It can minimize the mean square error under the worst conditions. However, if the value becomes extreme, it can easily cause the loss of useful signal information.

[0094] The selection of the threshold is another important step in wavelet thresholding. If the threshold is too large, some of the original signal will be mistaken for noise and some useful signals will be filtered out. If the threshold is too small, the wavelet coefficients will contain too much noise, which will lead to signal distortion and a decrease in the noise reduction effect. Therefore, the selection of the threshold will also affect the quality of filtering.

[0095] Therefore, in the noise reduction algorithm proposed in this invention, after threshold selection experiments, the following improved threshold determination method is adopted:

[0096]

[0097] In the formula, j is the decomposition scale. It can be seen that the noise wavelet coefficients decrease as the threshold increases with the increase of the decomposition scale, so they can more effectively distinguish between noise and signal, and retain the signal to a greater extent while removing noise.

[0098] The choice of wavelet threshold function is as follows:

[0099] Threshold denoising involves performing wavelet decomposition on the signal to obtain the high-frequency coefficients and the lowest low-frequency coefficients at each level. Generally, the high-frequency coefficients represent the original signal, while the low-frequency coefficients represent noise. Therefore, a threshold function is needed to filter out the noisy low-frequency coefficients and remove the noise coefficients. The traditional hard and soft threshold functions proposed by Donoho, as well as commonly used threshold selection functions, are as follows:

[0100]

[0101] The traditional soft thresholding method estimates y s (x) is a continuous function, but there is a fixed constant between the absolute value and the true value (when |x|≥λ). This difference will cause the value to be taken as an error. It is not necessarily the best to completely remove this deviation in the hard thresholding process. Correspondingly, the function is continuous at ±λ but cannot be differentiated, and there is a step phenomenon. The corresponding noise components in x still need to be removed.

[0102] Considering the above problems, an improved threshold function is designed. This threshold function is not only continuous with the soft threshold function, but also the deviation between y(x) and |x| gradually decreases as |x| increases. The new threshold function proposed in this invention is as follows:

[0103]

[0104] In the formula, j is the decomposition scale;

[0105] When |x|→λ,y p (x)→0, indicating that this function is similar to the soft thresholding function, and like the soft thresholding function, it is continuous without a step phenomenon. However, when |x|→∞, y p (x)→x, at this time the new threshold function is similar to the hard threshold function, which solves the problem that the soft threshold function always has a certain deviation in the design. In addition, it can be found that the improved threshold function will change with the decomposition scale j, and the adaptability to different decomposition levels will be enhanced. There are no uncertain parameters in the function, thus improving the stability of signal denoising.

[0106] In summary, as Figure 2 As shown, the specific noise reduction process of this invention is as follows:

[0107] Input: original vibration signal, original acoustic emission signal, and original stress signal x(t), t = 1, 2, ..., N;

[0108] Output: Noise-reduced signal y(t), t=1,2,...,N;

[0109] 1. Discretize the original vibration signal, original acoustic emission signal, and original stress signal x(t), and perform a 4-level decomposition using wavelets to obtain the approximate components {CA} for each level of decomposition. j,k} and detail components {CD j,k} and the amplitude A of the wavelet detail signal at that layer j ;

[0110]

[0111] 2. Calculate the noise standard deviation σ of each layer based on the detail signal coefficients of that layer. j and adaptive threshold thr j ;

[0112]

[0113] 3. Use threshold quantization functions to process the signal coefficients of each layer to complete threshold noise reduction at each scale;

[0114] 4. Perform inverse wavelet transform on the components processed at each layer to obtain the denoised components and complete the wavelet reconstruction.

[0115]

[0116] The algorithm process has ended.

[0117] In step 204, the features of the processed vibration signal, acoustic emission signal, and stress signal are extracted to construct the blasting vibration feature matrix, specifically as follows:

[0118] For vibration signals, their vibration characteristics are extracted, including:

[0119] Peak particle velocity (PPV): Records the maximum vibration velocity caused by the blast (unit: mm / s);

[0120] Dominant frequency: The dominant frequency of the vibration signal extracted by Fourier transform (unit: Hz);

[0121] Vibration duration: The time from the start of the vibration signal to its decay to 10% of the peak value (unit: ms);

[0122] For acoustic emission signals, extract their acoustic emission features, including:

[0123] Acoustic emission event count: Records the total number of acoustic emission events generated by rock mass fracturing during blasting;

[0124] Acoustic emission energy: Calculates the total energy of the acoustic emission signal (unit: aJ);

[0125] Acoustic emission amplitude: Extracts the maximum amplitude of the acoustic emission signal (unit: dB);

[0126] For stress signals, extract their stress characteristics, including:

[0127] Peak stress: Record the maximum stress value caused by the blast (unit: MPa);

[0128] Stress change rate: Calculates the rate of change of stress over time (unit: MPa / s);

[0129] Stress recovery time: The time it takes for stress to recover from its peak value to a steady state (unit: seconds);

[0130] Based on vibration characteristics, acoustic emission characteristics, and stress characteristics, the blasting vibration characteristic matrix F is constructed as follows:

[0131] F = [PPV, dominant frequency, duration, acoustic emission event count, acoustic emission energy, acoustic emission amplitude, peak stress, stress change rate, stress recovery time].

[0132] In step 3, a prediction model is built based on the improved LSTM model. The prediction model is trained based on the preset blasting vibration feature matrix and its corresponding blasting parameters and geological conditions to obtain the trained prediction model, specifically:

[0133] A prediction model is built based on the LSTM-SA model, which consists of a Self-Attention computation module, an LSTM module, and a fully connected neural network.

[0134] The LSTM-SA model consists of a self-attention computation module, an LSTM module, and a fully connected neural network. The model structure is as follows: Figure 3 As shown, the LSTM-SA model first calculates the self-attention weight vector of the data at each time step in the input sequence by applying a self-attention mechanism. In this step, the correlation between the data at each time step and the data at all other time steps is calculated, and a series of weight coefficients are generated to form the attention weight vector. In order to combine LSTM with the self-attention mechanism and alleviate the memory decay problem of LSTM, the LSTM-SA model concatenates the original input at the current time step with the attention weight vector at the corresponding time step and uses it as the input of the LSTM module.

[0135] By connecting self-attention weights to the original input, the LSTM-SA model can leverage the advantages of the self-attention mechanism to dynamically adjust the importance of the input data. This approach helps the model better capture the correlations and dependencies at different times in the input sequence, thereby improving the model's expressive power and predictive performance.

[0136] In this model, the self-attention weight calculation module uses the multilayer perceptron method to calculate the relevance weights. The LSTM module consists of two stacked LSTM layers with 64 and 32 neurons respectively. The fully connected network consists of two fully connected layers with 32 and 1 neurons respectively. Its main function is to integrate the output of the LSTM module and map it to health indicators.

[0137] Regarding the choice of activation function, the ReLU function is used in both the self-attention weight calculation module and the LSTM module, while the Sigmoid function is used in the last layer of the fully connected network to restrict the output value to the range of [0, 1].

[0138] During its training process, the SGDR learning rate adjustment algorithm is used for training, specifically as follows:

[0139] During neural network training, many hyperparameters need to be adjusted to achieve the best convergence. The learning rate is one of the most significant hyperparameters affecting network training. In neural network training, the learning rate represents the step size by which the model moves in the direction of minimum error. In each iteration, the neural network calculates the predicted value based on the input data, calculates the loss error with the label value, and then backpropagates the error to adjust the network weights. This process is repeated iteratively to gradually bring the network model to the global optimum. The learning rate is the parameter for adjusting the weights in each iteration, usually represented by n. The formula for updating the network weights is:

[0140]

[0141] An appropriate learning rate can help a network converge to the optimal solution quickly. However, if the learning rate is set too large or too small, the results will be contrary to expectations. If the learning rate is too large, the gradient updates will oscillate around the optimal solution but fail to reach it. As a result, the network is difficult to converge and the prediction accuracy fluctuates. If the learning rate is too small, although the model updates the gradient in the direction of the optimal solution, the update speed is slow. As a result, the network converges slowly. On the other hand, a learning rate that is too small may also cause the model to fall into a local optimum, i.e., a saddle point, and fail to reach the global optimum.

[0142] Therefore, the learning rate setting determines the ease of training the network model and directly affects the model accuracy. In this invention, an algorithm for dynamically adjusting the learning rate is used. Commonly used strategies for dynamically adjusting the learning rate include:

[0143] Step decay: After the network has been trained iteratively a specified number of times, the learning rate is multiplied by a preset coefficient;

[0144] Exponential decay: During the iterative training of a network, the learning rate decreases exponentially.

[0145] Cyclic learning rate: A method of cyclically adjusting the learning rate, gradually increasing it from a minimum to a maximum value, and repeating this process to change the learning rate periodically;

[0146] Cosine annealing learning rate: During training, the learning rate is reduced from its maximum value to its minimum value in the form of a cosine function. The rate of decrease is slow at first, then accelerates, and finally decreases slowly again. This decreasing pattern can bring good results to network training.

[0147] The objective function of a neural network is a convex function. Therefore, during training, when the model searches for the optimal solution in the direction of minimizing error, it may get stuck in a local optimum. Increasing the learning rate can help the model escape from the local optimum. This method is called hot-restart stochastic gradient descent (SGDR).

[0148] This invention applies the SGDR learning rate adjustment algorithm to the training of LSTM-SA models. Its advantage is that it can suddenly increase the learning rate in each cycle of learning rate change to force the model to jump out of local optima. However, at the beginning of each cycle, the learning rate will suddenly increase to the preset maximum value, which will cause the model to oscillate during training and is not conducive to the smooth convergence of the model. Therefore, this invention transforms the sudden increase of the initial learning rate to the maximum value in each cycle into a linear increase to the maximum value, which reduces the problem of model oscillation.

[0149] As can be seen from the above analysis, in the hot restart cosine annealing algorithm, the learning rate changes periodically within the range of maximum and minimum values. However, the algorithm does not provide a method for determining the boundary value. Therefore, this invention first performs multiple pre-training operations on the training set to find the optimal boundary value. During the pre-training process, the initial value of the learning rate is set to a very small value. In each iteration, the learning rate is increased exponentially to a larger value. During this process, the value of the learning rate and the value of the loss error are recorded. The learning rate at which the loss error begins to decrease is taken as the lower limit of SGDR, and the learning rate at which the loss error begins to increase is taken as the upper limit of SGDR. This method can ensure that the periodic change of the learning rate is always within the range that minimizes the loss error, making the training process smoother and accelerating the convergence speed of the model.

[0150] The inputs to the prediction model are blasting parameters, geological conditions, and blasting vibration characteristic matrix;

[0151] The output of the prediction model is:

[0152] Vibration impact range: The area where the predicted vibration velocity exceeds the threshold (unit: m);

[0153] Vibration intensity distribution: Predicted vibration velocity (PPV) values ​​at different locations (unit: mm / s).

[0154] The blasting parameters include:

[0155] Charge (Q): The amount of explosive used in a single blast (unit: kg);

[0156] Resistance line (W): The minimum distance from the center of the charge to the free surface (unit: m);

[0157] Blasting methods: such as hole-by-hole detonation, row-by-row detonation, etc.

[0158] Delay time: Inter-segment delay time (unit: ms);

[0159] The geological conditions include:

[0160] Rock mass grade (R): Classified according to RQD (Rock Quality Index) or rock mass integrity coefficient;

[0161] Fracture density (D): The number of fractures per unit volume (unit: fractures / m²) 3 );

[0162] Rock mass wave velocity (Vp): The propagation speed of longitudinal waves in rock mass (unit: m / s);

[0163] Groundwater conditions: such as dry, wet, or saturated.

[0164] In step 4, the blasting vibration feature matrix to be detected, along with its corresponding blasting parameters and geological conditions, are input into the trained prediction model to obtain the prediction results, specifically:

[0165] The blasting vibration characteristic matrix to be detected, along with its corresponding blasting parameters and geological conditions, is input into the trained prediction model to obtain prediction results. Risk levels are then classified based on these prediction results, for example:

[0166] Low risk: PPV < 50% of the limit;

[0167] Medium risk: 50% ≤ PPV < 85% of the limit;

[0168] High risk: PPV ≥ 85% of the limit;

[0169] It can also provide early warning information, such as outputting a level-three early warning signal:

[0170] Yellow alert: Predicted vibration velocity reaches 70% of the limit;

[0171] Orange alert: Predicted vibration velocity reaches 85% of the limit;

[0172] Red alert: Predicted vibration velocity reaches 100% of the limit.

[0173] This invention provides an embodiment for monitoring and predicting blasting vibration in a high-speed railway tunnel project. The tunnel is 12.5 km long, with a rock mass grade of Class III (moderately intact) and a fracture density of 12 fractures / m. 3 Groundwater conditions: moist; Sensitive target: There is an ancient Ming Dynasty building 80m away from the tunnel axis (permissible vibration velocity ≤3mm / s);

[0174] When deploying the sensors, a main control node is set up every 50m, four dual-redundant monitoring terminals are set up around the ancient building, and sensors are set up around the main control node. In the end, 120 vibration sensors, 80 acoustic emission sensors and 60 stress sensors are deployed.

[0175] The superiority of the present invention is determined by comparing the traditional method (Sadowski formula) with the method described in this invention. The comparison results are shown in Table 1.

[0176] Table 1 Comparison Results

[0177]

[0178]

[0179] As shown in Table 1, the prediction error of the method of the present invention is significantly lower than that of the traditional method, and it can accurately predict the range of vibration influence and trigger early warning in a timely manner.

[0180] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0181] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for monitoring blasting vibration and predicting the impact range of long tunnels in environmentally sensitive areas, characterized in that, include: Step 1: Construct a three-dimensional dynamic monitoring network; Step 2: Collect multi-source data based on a three-dimensional dynamic monitoring network and preprocess it to obtain the blasting vibration characteristic matrix; Specifically: Step 201: Collect the original vibration signal based on the vibration sensor, collect the original acoustic emission signal based on the acoustic emission sensor, and collect the original stress signal based on the stress sensor; Step 202: Perform noise reduction processing on the acquired original vibration signal, original acoustic emission signal and original stress signal; Step 203: Perform data normalization and time alignment on the vibration signal, acoustic emission signal and stress signal after signal denoising; Step 204: Extract the characteristics of the processed vibration signal, acoustic emission signal and stress signal, and construct the blasting vibration characteristic matrix; In step 202, the acquired original vibration signal, original acoustic emission signal, and original stress signal are subjected to signal noise reduction processing, specifically as follows: The original vibration signal, original acoustic emission signal, and original stress signal are denoised using an improved wavelet threshold denoising algorithm. The specific process is as follows: The original vibration signal, original acoustic emission signal and original stress signal are discretized, and the wavelet is decomposed into 4 levels to obtain the approximate components and detail components of each level, as well as the amplitude of the wavelet detail signal of that level. Calculate the noise standard deviation and adaptive threshold of each layer based on the detailed signal coefficients of each layer; The signal coefficients of each layer are processed using a threshold quantization function to complete threshold noise reduction at each scale. Perform inverse wavelet transform on the components processed at each layer to obtain the denoised components and complete the wavelet reconstruction. In step 204, the features of the processed vibration signal, acoustic emission signal, and stress signal are extracted to construct the blasting vibration feature matrix, specifically as follows: For vibration signals, extract their vibration characteristics, including peak particle vibration velocity, dominant frequency, and vibration duration; For acoustic emission signals, their acoustic emission characteristics are extracted, including acoustic emission event counts, acoustic emission energy, and acoustic emission amplitude; For stress signals, extract their stress characteristics, including peak stress, stress change rate, and stress recovery time; Based on vibration characteristics, acoustic emission characteristics, and stress characteristics, the blasting vibration characteristic matrix F is constructed as follows: F = [PPV, dominant frequency, duration, acoustic emission event count, acoustic emission energy, acoustic emission amplitude, peak stress, stress change rate, stress recovery time]; Step 3: Build a prediction model based on the improved LSTM model. Train the prediction model based on the preset blasting vibration feature matrix and its corresponding blasting parameters and geological conditions to obtain the trained prediction model. Step 4: Input the blasting vibration feature matrix to be detected, along with its corresponding blasting parameters and geological conditions, into the trained prediction model to obtain the prediction results.

2. The method according to claim 1, characterized in that, In step 1, a three-dimensional dynamic monitoring network is constructed, specifically as follows: The monitoring terminals are deployed at intervals along the tunnel axis by a drone swarm. Dual redundant monitoring terminals are set up around sensitive targets. Real-time data transmission is achieved based on LoRa and 5G dual-mode communication. The monitoring terminals are equipped with vibration sensors, acoustic emission sensors and stress sensors.

3. The method according to claim 2, characterized in that, In step 3, a prediction model is built based on the improved LSTM model. The prediction model is trained based on the preset blasting vibration feature matrix and its corresponding blasting parameters and geological conditions to obtain the trained prediction model, specifically: A prediction model is built based on the LSTM-SA model, which consists of a Self-Attention computation module, an LSTM module, and a fully connected neural network. During its training process, the SGDR learning rate adjustment algorithm is used for training.

4. The method according to claim 3, characterized in that, The blasting parameters include charge quantity, resistance line, blasting method and delay time, and the geological conditions include rock mass grade, fracture density, rock mass wave velocity and groundwater conditions.

Images