Intelligent picking method and system for arrival of microseismic monitoring in rock mass engineering

Abstract

The present invention discloses an intelligent picking method and system for microseismic monitoring of rock engineering, including the following steps: data source → method proposal → analysis and comparison results → verification → conclusion; the method proposal includes a framework overview, a two-stage P-wave and S-wave arrival time picking method and an image regression method; the two-stage P-wave and S-wave arrival time picking method includes a target detection method, a pixel time transformation method and a one-dimensional convolutional neural network; the verification stage includes the influence of the convolution kernel and the fully connected layer, a multi-index comprehensive evaluation, error analysis and application. The beneficial effect of the present invention is that the method can reduce the number of negative microseismic signal samples and improve the accuracy of S-wave arrival time picking. Finally, the influence of the model structure on the two-stage method is verified, and the multi-index comprehensive evaluation method is used to avoid the one-sided evaluation of the model by a single parameter.

Description

Technical Field

[0001] The invention relates to rock engineering microseismic monitoring, in particular to an intelligent picking method and system for rock engineering microseismic monitoring arrival, belonging to the technical field of intelligent monitoring of rock engineering disasters. Background Art

[0002] Microseismic (MS) monitoring technology uses microfracture signals generated when rock masses are subjected to loads to study fracture or failure processes. This technology has become a basic method for safety monitoring research and management in fields such as slopes, coal mines, and hydropower station projects. In actual rock engineering, the processing of microseismic monitoring data includes the following three main steps: signal recognition, wave arrival time picking, and source location. Usually, data processing is performed by microseismic experts or technicians. However, manual processing is usually slow and can result in problems such as missed processing and processing errors. Some studies have achieved automation of these three steps, but wave arrival time picking still requires more accurate and applicable automatic picking algorithms, because the accuracy of wave arrival time picking seriously affects the accuracy of source location.

[0003] Traditional single evaluation indicators are difficult to comprehensively evaluate the model. However, model performance is multifaceted, including model accuracy indicators such as MAPE (mean absolute percentage error) and PCT (percentage of correct time points), and model speed indicators such as model size, speed, and computing power. A single evaluation indicator analyzes the one-sidedness of the model and is difficult to comprehensively evaluate the application ability of the model. That is, the use of a comprehensive multi-indicator evaluation method is conducive to parameter optimization and model comparison. Therefore, we proposed an intelligent picking method and system for microseismic monitoring of rock engineering. Summary of the invention

[0004] The present invention achieves the above-mentioned purpose through the following technical solutions, including an intelligent picking method and system for microseismic monitoring of rock engineering, including a data source, characterized in that it includes the following steps:

[0005] Data source → method proposal → analysis and comparison of results → verification → conclusion;

[0006] The method proposed includes a framework overview, a two-stage P-wave and S-wave arrival time picking method, and an image regression method;

[0007] The two-stage P-wave and S-wave arrival time picking method includes a target detection method, a pixel time transformation method and a one-dimensional convolutional neural network;

[0008] The verification phase includes the impact of convolution kernels and fully connected layers, multi-indicator comprehensive evaluation, error analysis and application.

[0009] Preferably, the pixel time conversion method:

[0010] The region of interest (ROI) of the microseismic signal is obtained by using the target detection method, but the time-amplitude data of the microseismic signal in the region of interest cannot be directly obtained. min and X max , then find X exactly min and X max The time region corresponding to the time-amplitude data [T min , T max ];

[0011] The relationship between pixels and time is constructed by the equation shown below:

[0012]

[0013] Where W is the number of pixels in the length direction of the image; T is the data sampling time; X min is the starting pixel value of the detection box; X max is the end pixel value of the detection box; T min is the starting point of time; T max It is the end of time;

[0014] The actual image pixel area used in the present invention is 7850 pixels wide and 870 pixels high. The sampling time of each original microseismic signal is 400ms. If the start and end pixels of the target detection feature area are 800 pixels and 1600 pixels, the corresponding time range can be calculated by the equation as 40.76ms-81.53ms. Finally, the amplitude data within the time range is extracted and used for P-wave and S-wave arrival time picking.

[0015] Preferably, the image regression method comprises:

[0016] The P-wave and S-wave arrival time picking network is reasonably designed by using a two-stage method. In order to verify the reliability of this method, an image regression method (one-stage picking method) is constructed for comparison.

[0017] The network structure of the image regression method is an end-to-end P-wave and S-wave arrival time picking structure; ResNet 50 and AlexNet are used to extract feature areas; then, the two regression heads are connected separately, and the regression head consists of a convolutional layer and a fully connected layer; finally, the first arrival time of the P wave and S wave is directly output through the fully connected layer.

[0018] Preferably, in the verification stage: the two-stage picking method is a basic unit consisting of convolution, batch normalization and ReLU layers, and the three units are superimposed to form a feature extraction structure;

[0019] Finally, the fully connected layer is used to perform regression prediction of the P-wave and S-wave arrival times. After comparison, it is determined that the two-stage arrival time picking method has the best prediction effect. The next step is to study the model structure design. The goal is to design a lightweight and high-precision P-wave and S-wave arrival time picking model through network improvement.

[0020] Preferably, the multi-index comprehensive evaluation is:

[0021] In order to evaluate the model more comprehensively, the ranking method was improved and a new multi-index comprehensive evaluation method was proposed; the specific steps are as follows:

[0022] Step 1: Calculate the evaluation indicators of MAPE, FLOPs and parameter quantity of the test set;

[0023]

[0024] Step 2: Normalize using the formula and obtain MAPE', FLOPs' and number of parameters';

[0025] Step 3: Add the MAPE', FLOPs', and number of parameters' of the test set to obtain a comprehensive evaluation index and rank the models.

[0026] Preferably, the error analysis:

[0027] Through the multi-index comprehensive evaluation method, the formula calculates the average error of the P-wave and S-wave arrival time picking; finally, the average error is evaluated by PCT (including PCT-3, PCT-5 and PCT-10); for example, PCT-3 is the average error between the predicted time and the real time is less than 3 milliseconds, and the prediction is considered accurate; the calculated PCT-3, PCT-5 and PCT-10 values ​​are 50.11%, 73.26% and 93.20%, respectively;

[0028]

[0029] in is the average error; P P and P S are the predicted values ​​of the arrival time of the P wave and the S wave respectively; T P and T S are the true values ​​of the arrival times of the P and S waves, respectively;

[0030] Secondly, the noise resistance of the model was analyzed.

[0031] Preferably, three different types of microseismic signals are selected after prediction, namely, microseismic signals without noise interference, current microseismic signals, and noise microseismic signals. When there is no noise interference, the two-stage method can predict the P wave and S wave very well, and the solid line and the dotted line almost coincide (where the solid line is the true value and the dotted line is the predicted value); because the arrival time characteristics of the P wave and the S wave of this waveform are very obvious;

[0032] The current signal fluctuates regularly on the entire microseismic signal; at this time, the prediction result of the model is slightly worse than that of the noise-free microseismic signal; the prediction effect of the S-wave arrival time of the current microseismic is better than that of the P-wave arrival time;

[0033] Noise microseismic: The environmental noise at the construction site is high, resulting in a large number of noise microseismic signals in the microseismic system. This part of the signal is also very important. Noise microseismic is mainly affected by environmental noise, and does not have the same fluctuation pattern as current microseismic, which increases the difficulty of picking up the arrival time. It can be observed that under the influence of strong noise, both P waves and S waves can obtain good prediction results, indicating that this two-stage method has a certain degree of noise resistance.

[0034] Noisy microseismic signals contain deceptive high-quality noise that can be easily confused with microseismic signals. At this time, target detection is needed to accurately locate the ROI (region of interest) of the microseismic signal to ensure the accuracy of arrival time picking.

[0035] In addition, the prediction of P and S waves based on waveform characteristics found that the interference signals in these two microseismic signals include not only noise signals but also current signals; this part of the signal is a mixed signal of microseismic, current and noise, and the arrival characteristics of P and S waves are not obvious, which requires adding more microseismic signals with current and noise to the original data set; the network can have a deeper understanding of the characteristics of these waveforms, thereby improving the accuracy of arrival time picking.

[0036] Preferably, the intelligent picking system is as follows:

[0037] Firstly, a two-stage method for picking the arrival time of P and S waves is proposed, which includes waveform ROI extraction and P and S wave arrival time prediction. Then, the picking accuracy of the one-stage and two-stage methods is compared, where the one-stage method includes the image regression method (method one) and the MLP model that processes microseismic signals using the amplitude data method (method two). The two-stage method performs better in picking the arrival time of P and S waves because it reduces the impact of negative samples on the accuracy.

[0038] Then, in order to obtain a lightweight and high-precision two-stage model, the effects of five different sizes of convolution kernels (5-5-5, 11-11-11, 31-31, 71-71, and 101-31-11) on the model were verified; in addition, a greater weight was given to the S value in the loss function, which improved the prediction accuracy of the S value;

[0039] Furthermore, the model accuracy and model size cannot be optimized at the same time, and the number of parameters is mainly concentrated in the first fully connected layer; then the global maximum pooling layer is used to replace the fully connected layer, and it is found that the number of model parameters is significantly reduced; finally, in order to select the most accurate and lightweight structure, a multi-index comprehensive evaluation method is proposed;

[0040] In addition, in the prediction results of microseismic signals without noise interference, current microseismic signals, and noise microseismic signals, it is found that the two-stage method has a certain noise resistance because it can distinguish deceptive high-quality noise and obtain good prediction results for both P waves and S waves under the influence of strong noise.

[0041] The beneficial effects of the present invention are as follows: In the present invention, a complex microseismic data set is firstly extracted from the microseismic monitoring system of the Han-Wei River Diversion Project, and the data set contains many microseismic waveforms with severe noise. Then, according to the characteristics of the microseismic signal, a new two-stage arrival time picking method is proposed (target detection method: firstly, the ROI of the microseismic waveform image is given, the amplitude data of the ROI is extracted, and the P-wave and S-wave arrival times are predicted using a one-dimensional convolutional neural network (1D CNN) regression model. This method can reduce the number of negative microseismic signal samples and improve the accuracy of S-wave arrival time picking. Finally, the influence of the model structure on the two-stage method is verified, and the multi-index comprehensive evaluation method is used to avoid the one-sided evaluation of the model by a single parameter. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the process of the present invention;

[0043] Figure 2 Some example diagrams of microseismic signals of different levels according to the present invention;

[0044] Figure 3 It is the technical roadmap of the two-stage method of the present invention;

[0045] Figure 4 It is a schematic diagram of YOLO of the present invention;

[0046] Figure 5 It is a schematic diagram of pixel time transformation of the present invention;

[0047] Figure 6 The convolution working principle diagram of the present invention;

[0048] Figure 7The structural diagram of 1D CNN of the present invention;

[0049] Figure 8 The network structure diagram of the image regression method of the present invention;

[0050] Fig. 9 MAPE value diagram of the three methods of the present invention;

[0051] Fig.10 The test results of the models 1-5 of the present invention are shown in FIG.

[0052] Fig.11 A diagram showing the ratio of the number of parameters and triggers on the first fully connected layer of the present invention;

[0053] Fig.12 The test results of Models 6-8 of the present invention are shown in FIG.

[0054] Fig.13 The prediction result diagram of three signals of noise-free microseismic, current microseismic and noise microseismic in the present invention;

[0055] Fig.14 The test result diagram of the current, noise and microseismic mixed signal of the present invention;

[0056] Fig.15 The present invention uses different arrival time picking methods to determine the location map. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] The embodiment of the present invention discloses an intelligent picking method and system for microseismic monitoring arrival of rock mass engineering.

[0059] Including data sources: including the following steps:

[0060] Data source → method proposal → analysis and comparison of results → verification → conclusion;

[0061] The proposed method includes framework overview, two-stage P-wave and S-wave arrival time picking method and image regression method;

[0062] The two-stage P-wave and S-wave arrival time picking method includes a target detection method, a pixel time transformation method, and a one-dimensional convolutional neural network;

[0063] The verification includes the impact of convolution kernels and fully connected layers, multi-indicator comprehensive evaluation, error analysis and application.

[0064] Framework Overview: In order to extract the first arrival times of P and S waves, a new two-stage extraction method is proposed according to the characteristics of microseismic signals.

[0065] Firstly, the data set is preprocessed, including image data and amplitude data. Then, the data is analyzed using a two-stage selection method, which consists of two main parts:

[0066] In the first part, the waveform image data is input into the model, and the fluctuation area of ​​the waveform is extracted according to the target detection method, and the approximate time range of the first arrival of the P wave and the S wave is determined. Then, the amplitude data within the ROI (region of interest) is extracted (the ROI of the microseismic signal is the main fluctuation area of ​​the waveform);

[0067] The second part is a P-wave and S-wave arrival time picking regression method based on one-dimensional CNN, which directly outputs the actual arrival time of P-wave and S-wave.

[0068] Target detection methods:

[0069] The You Only Look Once (YOLO) network unifies target classification and localization into a regression problem. It divides the feature map of the input image into S×S grid cells and predicts the bounding box, location, and category of the target contained in all grid cells at once. The predicted confidence score indicates the accuracy of the model's prediction of these boxes.

[0070] The loss function of YOLO is shown in the formula as follows:

[0071]

[0072] Among them, coordError is the loss of coordinate position; iouError is the loss of lost intersection; classError is the loss of classification;

[0073] Finally, YOLO applies non-maximum suppression (NMS) to remove all bounding boxes that do not contain any objects or contain the same objects as other bounding boxes; by selecting a threshold, NMS removes all overlapping bounding boxes (the IOU values ​​of these bounding boxes are higher than the threshold) and finally outputs the ROI (region of interest) of the microseismic signal.

[0074] Pixel time transformation method:

[0075] The region of interest (ROI) of the microseismic signal is obtained by using the target detection method, but the time-amplitude data of the microseismic signal in the region of interest cannot be directly obtained. min and X max , then find X exactly min and X max The time region corresponding to the time-amplitude data [T min , T max ];

[0076] The relationship between pixels and time is constructed by the equation shown below:

[0077]

[0078] Where W is the number of pixels in the length direction of the image; T is the data sampling time; X min is the starting pixel value of the detection box; X max is the end pixel value of the detection box; T min is the starting point of time; T max It is the end of time;

[0079] The actual image pixel area used in the present invention is 7850 pixels wide and 870 pixels high. The sampling time of each original microseismic signal is 400ms. If the start and end pixels of the target detection feature area are 800 pixels and 1600 pixels, the corresponding time range can be calculated by the equation as 40.76ms-81.53ms. Finally, the amplitude data within the time range is extracted and used for P-wave and S-wave arrival time picking.

[0080] One-dimensional Convolutional Neural Network:

[0081] CNN is usually divided into the following two forms: 2D CNN and 1D CNN; 2D CNN is widely used in image recognition, while 1D CNN is usually used to process sequence data. It is worth noting that there are significant differences in network design between 1D CNN and 2D CNN; small convolution kernels are usually used in 2D CNN;

[0082] Assuming that the number of input and output channels of the network is C, two layers of 3×3 convolution kernels (with 18C2 parameters) have the same experience as one layer of 5×5 convolution kernels (with 25C2 parameters). However, in 1D CNN, when the size of the two-layer convolution kernel is 3, its parameters are 6C2; when the size of the convolution kernel is 5, its parameters are 5C2, and both have the same experience.

[0083] The network structure proposed in the present invention consists of multiple superimposed blocks, and the number of blocks is adjusted according to actual conditions; each block consists of a convolution layer, a batch normalization layer and a ReLU layer; each block is set by the stride parameter of the convolution, and the output size is adjusted. After multiple convolution feature extraction and nonlinear transformation processes, a dropout layer is set in the fully connected layer to prevent the network from overfitting; finally, the first arrival time of the P wave and the S wave is directly output through the output layer.

[0084] Image regression methods include:

[0085] The P-wave and S-wave arrival time picking network is reasonably designed by using a two-stage method. In order to verify the reliability of this method, an image regression method (one-stage picking method) is constructed for comparison.

[0086] The network structure of the image regression method is an end-to-end P-wave and S-wave arrival time picking structure; ResNet 50 and AlexNet are used to extract feature areas; then, the two regression heads are connected separately, and the regression head consists of a convolutional layer and a fully connected layer; finally, the first arrival time of the P wave and S wave is directly output through the fully connected layer.

[0087] Analysis and comparison results:

[0088] In order to compare the one-stage and two-stage methods, the original dataset is first divided into training, test and validation sets, with the numbers of 3166, 1780 and 1161 respectively; then, three P-wave and S-wave arrival time picking methods are selected as follows:

[0089] Method 1: Use image regression method to process the waveform image of microseismic signal (400ms); Method 2: Use multi-layer perceptron (MLP) model to process the amplitude data of microseismic signal (400ms), where Method 1 and Method 2 are one-stage methods; Method 3: Use two-stage method to process waveform image and amplitude data (60ms) (in order to compare with Method 2, first use MLP model instead of 1D CNN in the two-stage method);

[0090] During the training process, the network uses the stochastic gradient descent (SGD) optimization algorithm, and the momentum is set to 0.9. Finally, the MAPE evaluation indicator is used to compare the accuracy of the model;

[0091]

[0092] where y i True value, is the predicted value;

[0093] Method I has the worst recognition accuracy (maximum MAPE value), which shows that the direct image regression method is difficult to meet the requirements of high-precision prediction. This is because the core area of ​​the microseismic signal is an area with concentrated and similar energy, and the pixel values ​​in this area are basically the same; due to the changes in the arrival time of the P wave and the S wave, it will not cause a large change in the energy concentration area; therefore, the arrival time of the P wave and the S wave is directly predicted using pixel features, and the picking accuracy is poor;

[0094] Although the prediction accuracy of method II is higher than that of method I, the difference between the MAPE values ​​of the P wave and the S wave arrival predicted by method II is larger; the results show that in the prediction process of method II, the P value is easier to predict than the S value; in method II, a neural network is directly established, 4001 amplitude data are input, and 4001 features are generated; however, only about 15% of these features are valuable, and the other 85% of the amplitude values ​​are about 0; this 85% of the data does not contribute positively to the results, but because too many useless features are input, they often interfere with the network and affect the prediction accuracy of the model;

[0095] The best method is method III, and the MAPE values ​​of P and S are also very close. Method III simulates the P-wave and S-wave arrival time picking process of microseismic experts, that is, first observe the ROI (region of interest) of the microseismic signal, and then judge and pick the P-wave and S-wave arrival time by combining the amplitudes in the ROI. This method combines the advantages of easy image feature extraction and high-precision feature picking using amplitude data.

[0096] Although the joint target detection and MLP method (method III) has higher picking accuracy than methods I and II, the MLP network has many nodes and a large model, which is difficult to meet the requirements of field applications; compared with MLP, the CNN architecture is less affected by the "curse of dimensionality"; if the dimension of the input vector is high, CNN will be able to achieve performance comparable to MLP, but with much fewer trainable parameters; CNN considers spatial correlation and learns spatially local input patterns, so it is more suitable for continuous input data; CNN architecture is more stable than MLP for microseismic signals with some noise.

[0097] Verification stage: The two-stage picking method is a basic unit consisting of convolution, batch normalization and ReLU layers. The three units are superimposed to form a feature extraction structure;

[0098] Finally, the fully connected layer is used to perform regression prediction of the P-wave and S-wave arrival times. After comparison, it is determined that the two-stage arrival time picking method has the best prediction effect. The next step is to study the model structure design. The goal is to design a lightweight and high-precision P-wave and S-wave arrival time picking model through network improvement.

[0099] The impact of convolution kernels and fully connected layers:

[0100] The computing conditions at the microseismic monitoring site are limited. If the model is evaluated only in terms of accuracy, it will easily make it more difficult to deploy the model in the field. Therefore, the size of the model needs to be considered comprehensively. Regarding the model size, the evaluation indicators selected are the number of model parameters and the number of floating-point operations (FLOPs) performed by the model per second.

[0101] In order to compare the effects of different convolution kernel sizes on the prediction results of P-wave and S-wave arrival times, models 1-5 are convolution kernels 5-5-5, 11-11-11, 31-31-31, 71-71-71 and 101-31-11, respectively. Unlike the MLP network, models 1-5 have a higher prediction ability for S values ​​than P values. Since the S value is given a greater weight in the loss function, the model pays more attention to the prediction of the S value. Among these models, model 2 has the highest accuracy, but the model size is not optimal; model 4 has the best size, but the accuracy is not the highest. Therefore, it is difficult to optimize both accuracy and model size by adjusting the convolution kernel alone. The number of parameters and triggers of models 1-5 is mainly concentrated in the first fully connected layer. The average parameters account for 90.18% of the model parameters. If the model needs to be further compressed, the prediction part of the model needs to be improved, especially the first fully connected layer. Models 6-8 were further designed. Compared with models 1-5, the fully connected layer in models 6-8 was replaced by the global maximum pooling layer. From the test results, we can see that the number of model parameters has been significantly reduced. Model 6 and Model 8 have fewer parameters than Model 7.

[0102] In summary, the effects of convolution kernel size and fully connected layers on model accuracy and size were verified. Changing the size of the convolution kernel can improve the accuracy of the model and increase the number of parameters (models 1-5). Considering the impact of the first fully connected layer of the model on the overall model size, it was further optimized (models 6-8). However, it is difficult to select the most accurate and lightest structure from these eight models. For example, if only the model size is considered, model 8 is the best. If only the accuracy is considered, model 2 is the best, but its number of parameters is 60 times that of model 8. Therefore, considering only a single indicator cannot truly evaluate the comprehensive performance of the model, and it is necessary to conduct a comprehensive evaluation of multiple indicators.

[0103] Multi-index comprehensive evaluation:

[0104] In order to evaluate the model more comprehensively, the ranking method was improved and a new multi-index comprehensive evaluation method was proposed. The specific steps are as follows:

[0105] Step 1: Calculate the evaluation indicators of MAPE, FLOPs, and parameter quantity of the test set.

[0106]

[0107] Step 2: Normalize and obtain MAPE', FLOPs' and Number of Parameters' using the formula.

[0108] Step 3: Add the MAPE', FLOPs' and number of parameters' of the test set to get the comprehensive evaluation index and rank the models.

[0109] Models 1-8 combine object detection and 1D CNN, and Model-9 is a combination of object detection and MLP network. The experimental results are processed using a multi-index comprehensive evaluation method and the final scores are ranked. Model-8 is the best model after comprehensive evaluation in terms of accuracy and brightness. Model 9 has the worst evaluation results, further indicating that 1D CNN is superior to MLP network in the task of picking P-wave and S-wave arrival times.

[0110] Error analysis:

[0111] Through the multi-index comprehensive evaluation method, model 8 was determined to be the optimal model. Then, the test results of model 8 were analyzed; the formula calculated the average error of the P-wave and S-wave arrival time picking. Finally, the average error was evaluated by PCT (including PCT-3, PCT-5 and PCT-10). For example, PCT-3 is the average error between the predicted time and the true time is less than 3 milliseconds, and the prediction is considered accurate. The calculated PCT-3, PCT-5 and PCT-10 values ​​are 50.11%, 73.26% and 93.20%, respectively.

[0112]

[0113] in is the average error; P P and P S are the predicted values ​​of the arrival time of the P wave and the S wave respectively; T P and T S are the true values ​​of the arrival times of the P and S waves, respectively;

[0114] Secondly, the noise resistance of the model was analyzed. Three different types of microseismic signals were selected after prediction, namely, microseismic signals without noise interference, current microseismic signals, and noise microseismic signals. 14(a) and 14(b), it can be observed that when there is no noise interference, the two-stage method can predict the P wave and S wave very well, and the solid line and the dotted line almost coincide (where the solid line is the true value and the dotted line is the predicted value). Because the arrival time characteristics of the P wave and S wave of this waveform are very obvious, it is also easy for microseismic experts to obtain a high arrival time picking accuracy. In addition, it is also easy to detect the ROI (region of interest) of the microseismic signal during target detection. 14(c) and 14(d) are the prediction results of current microseismic signals, and the current signal fluctuates regularly on the entire microseismic signal. At this time, the prediction results of the model are slightly worse than those of the noise-free microseismic signal, and the current microseismic S wave arrival time prediction is better than the P wave arrival time prediction, which may be caused by the influence of the electrical signal fluctuation. 14(e) to 14(h) are noisy microseismic signals. The environmental noise at the construction site is high, resulting in a large number of noisy microseismic signals in the microseismic system. These signals are also very important. Noise microseismic signals are mainly affected by environmental noise and do not have the same fluctuation pattern as current microseismic signals, which increases the difficulty of arrival time picking. It can be observed that under the influence of strong noise, both P-waves and S-waves can obtain good prediction results, indicating that this two-stage method has certain noise resistance. There is deceptive high-quality noise in the noisy microseismic signal in 14(h), which is easily confused with the microseismic signal. At this time, target detection is needed to accurately locate the ROI (region of interest) of the microseismic signal to ensure the accuracy of arrival time picking.

[0115] However, under the influence of interference signals, some microseismic signals are not perfectly predicted. In the prediction of P-wave and S-wave in 14(a) and 14(b), it is found that the interference signals in these two microseismic signals not only have noise signals but also electrical signals based on the waveform characteristics. This shows that the current model is not good at predicting such microseismic signals. This requires adding more microseismic signals with electricity and noise to the original data set. Therefore, the network can have a deeper understanding of the characteristics of these waveforms, thereby improving the accuracy of arrival time picking.

[0116] Limited computer conditions at construction sites and a large number of microseismic signals and environmental noise result in poor application of traditional arrival time picking methods on site. The two-stage method is a lightweight, high-precision arrival time picking method with certain noise resistance, which has obvious advantages in practical applications. A typical field application case is used to further illustrate the practicality of this method.

[0117] A small-scale rock burst occurred near the pile number K39+975 of the Han-Wei River Diversion Project. The specific location of the rock burst was the tunnel arch. The main rock type of the surrounding rock was granite, and the falling rocks were in the form of flakes and blocks. This event was discovered in the microseismic monitoring system. First, the default method STA / LTA of the microseismic monitoring system was used for arrival time picking (each ball in the figure represents a microseismic event, and the size represents the energy).

[0118] It can be observed that the positioning error caused by the systematic time-picking process is unacceptable and has far exceeded the tunnel. Then, the microseismic expert regenerated the arrival time picking of the event and reprocessed it using the two-stage method. The event location after the microseismic expert picked the arrival time is almost consistent with the true location of the rockburst.

[0119] The event location after the two-stage method's arrival time picking is also very close to the actual location of the rockburst. This shows that the arrival time picking accuracy of the two-stage method is close to that of microseismic experts and is much better than traditional methods. This method is highly practical and can improve the efficiency of microseismic signal processing.

[0120] Due to the noisy environment at the rock engineering construction site, the microseismic signals collected by the accelerometer are accompanied by a large number of interference signals (current microseismic and noise microseismic). Too many negative samples in the microseismic signals will also seriously affect the accuracy of the arrival time prediction. Some studies only select P waves, and the source location accuracy will be significantly improved after adding S waves. In addition, due to the limitations of the computer and other conditions at the construction site, we should not only pay attention to the picking accuracy of the algorithm, but also pay attention to the size of the model.

[0121] To solve these problems, a two-stage P-wave and S-wave arrival time picking method is first proposed, which consists of the following two parts: waveform ROI (region of interest) extraction and P-wave and S-wave arrival time prediction. Then, the picking accuracy of the one-stage and two-stage methods is compared, where the one-stage method includes the image regression method (method one) and the MLP model that processes microseismic signals using the amplitude data method (method two).

[0122] The intelligent picking system is as follows: the two-stage method performs better in picking the arrival time of P and S waves because it reduces the impact of negative samples on accuracy; in order to obtain a lightweight and high-precision two-stage model, the effects of five different sizes of convolution kernels (5-5-5, 11-11-11, 31-31, 71-71, and 101-31-11) on the model (models 1-5) are verified; in addition, a larger weight is given to the S value in the loss function, which improves the prediction accuracy of the S value;

[0123] However, the model accuracy and model size cannot be optimized at the same time, and the number of parameters is mainly concentrated in the first fully connected layer. Then the global maximum pooling layer is used to replace the fully connected layer (models 6-8), and it is found that the number of parameters of the model is significantly reduced. Finally, in order to select the most accurate and lightest structure, a multi-index comprehensive evaluation method is proposed, and model 8 is determined to be the best model among the nine models.

[0124] In addition, in the prediction results of microseismic signals without obvious noise interference, current microseismic signals, and noise microseismic signals, it is found that the two-stage method has a certain noise resistance because it can distinguish deceptive high-quality noise, and obtain good prediction results for both P-waves and S-waves under the influence of strong noise. In summary, the two-stage method is an effective, high-precision, and practical new method for picking the arrival time of P-waves and S-waves, which lays the necessary foundation for providing real-time warnings during rock engineering construction.

[0125] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. An intelligent picking method for microseismic monitoring of rock mass engineering, including a data source, is characterized by: The steps include: Data source → method proposal → analysis and comparison of results → verification phase → conclusion; The method proposed includes a framework overview, a two-stage P-wave and S-wave arrival time picking method, and an image regression method; The two-stage P-wave and S-wave arrival time picking method includes a target detection method, a pixel time transformation method and a one-dimensional convolutional neural network; The verification phase includes the impact of convolution kernels and fully connected layers, multi-index comprehensive evaluation, error analysis and application; The specific steps of the multi-index comprehensive evaluation are as follows: Step 1: Calculate the evaluation indicators of MAPE, FLOPs and parameter quantity of the test set; Step 2: Normalize using the formula and obtain MAPE', FLOPs' and number of parameters'; Step 3: Add the MAPE', FLOPs' and parameter count' of the test set to obtain a comprehensive evaluation index and rank the models; The pixel time conversion method: By obtaining the pixel value X of the microseismic image in the area of ​​interest min and X max , then find X exactly min and X max The time region corresponding to the time-amplitude data [T min , T max ]; The relationship between pixels and time is constructed by the equation shown below: Where W is the number of pixels in the length direction of the image; T is the data sampling time; X min is the starting pixel of the detection box; X max is the end pixel value of the detection box; T min It is the starting point of time; T max It is the end of time; Image regression methods include: The network structure of the image regression method is an end-to-end P-wave and S-wave arrival time picking structure; ResNet50 and AlexNet are used to extract feature regions; then, two regression heads are connected respectively, and the regression head consists of a convolutional layer and a fully connected layer; finally, the first arrival time of the P-wave and S-wave is directly output through the fully connected layer; The error analysis: Through the multi-index comprehensive evaluation method, the formula calculates the average error of the P-wave and S-wave arrival time picking; finally, the average error is evaluated by PCT-3, PCT-5 and PCT-10; PCT-3 is the average error between the predicted time and the real time is less than 3 milliseconds, and the prediction is considered accurate; the calculated PCT-3, PCT-5 and PCT-10 values ​​are 50.11%, 73.26% and 93.20% respectively; in is the average error; P P and P S are the predicted values ​​of the arrival time of the P wave and the S wave respectively; T P and T S are the true values ​​of the arrival times of the P and S waves, respectively; Secondly, the noise resistance of the model was analyzed.

2. The intelligent picking method for rock mass engineering microseismic monitoring arrival time according to claim 1 is characterized by: The verification stage: The two-stage P-wave and S-wave arrival time picking method is a basic unit composed of convolution, batch normalization and ReLU layers, and the three units are superimposed to form a feature extraction structure; Finally, the fully connected layer is used to perform regression prediction of the P-wave and S-wave arrival times. After comparison, it is determined that the two-stage P-wave and S-wave arrival time picking method has the best prediction effect. The next step is to study its model structure design.

3. The intelligent picking method for rock mass engineering microseismic monitoring arrival time according to claim 1 is characterized by: Three different types of microseismic signals were selected after prediction, namely, microseismic signals without noise interference, current microseismic signals, and noise microseismic signals.

Images