Tunnel micro-seismic waveform identification, arrival time pickup and waveform end prediction integrated analysis method

Abstract

The invention provides a tunnel micro-seismic waveform identification, arrival time pickup and waveform end prediction integrated analysis method, which comprises the following steps: acquiring micro-seismic single-component waveforms to construct a waveform data set, and marking the waveforms; the waveform data and data in the waveform data are preprocessed; constructing a micro-seismic waveform integrated analysis model; the micro-seismic waveform integrated analysis model is optimized; and analyzing waveform data by using the optimized micro-seismic waveform integrated analysis model, and outputting a waveform category, P-wave arrival time, S-wave arrival time and an E-point position. The method is mainly used for structural design and training method optimization of an integrated analysis model for micro-seismic waveform recognition, arrival time pickup and waveform end prediction. According to the method, the model structure and the optimization algorithm are ingeniously designed, the shared features of the three tasks can be extracted at the same time, waveform category recognition, P wave arrival time prediction, S wave arrival time prediction and E point position prediction are completed in one execution process, and therefore the tedious multi-stage processing flow is effectively avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel microseismic monitoring, and particularly to an integrated analysis method for tunnel microseismic waveform recognition, arrival time picking, and waveform end prediction. Background Art

[0002] With the continuous increase in the burial depth of tunnel projects, the stress concentration phenomenon during the excavation process becomes increasingly significant, resulting in a gradual increase in the occurrence frequency of rockburst disasters. The intelligent analysis technology of microseismic data can effectively collect and analyze weak rock fracture signals, and then achieve accurate prediction of the occurrence probability and hazard level of rockbursts. Due to its ability to monitor and evaluate the potential rockburst risks during tunnel excavation in real time, the intelligent analysis technology of microseismic data has become an important tool in the rockburst warning system for deep-buried tunnel projects and is widely used in the safety monitoring and risk warning work in this field.

[0003] In a microseismic monitoring system, the data analysis process mainly relies on the combined action of multiple collaborative working modules, including key links such as waveform recognition, arrival time picking, and picking of the end of the vibration of the fracture signal. The accuracy and complexity of each module directly determine the accuracy and timeliness of the microseismic warning system. The core task of the waveform recognition module is to determine whether the waveform data triggering the sensor belongs to the fracture waveform; the arrival time picking module is responsible for accurately obtaining the arrival times of the shear wave (S wave) and the primary wave (P wave) in the fracture waveform, so as to deduce the occurrence location of the microseismic event through reverse deduction. The prediction module for the end position (E) of the fracture signal vibration mainly predicts the duration of the fracture waveform, which provides an important reference basis for the calculation of microseismic energy. Selecting the end position of the vibration too early may result in missing the tail energy of the signal. Especially during the signal attenuation process, the overall energy may be underestimated. On the contrary, selecting the end position too late may include the attenuated noise or irrelevant signals in the calculation, resulting in overestimated energy. Therefore, reasonably determining the end position of the waveform vibration to ensure that only the effective part of the signal is calculated is crucial for accurately evaluating the energy.

[0004] Existing waveform recognition methods can be implemented through machine learning algorithms. Such methods have a fast recognition speed and certain theoretical support, but they have the problem of low recognition accuracy. Time-of-arrival picking can be implemented through methods such as STA / LTA and autoregressive Akaike information criterion (AR-AIC). Such methods mainly obtain the locations of P-wave and S-wave arrival by mathematically calculating the waveform morphology to obtain the mutation point, so they rely on manual intervention and need to set thresholds and characteristic functions. For example, the invention of "a high-precision microseismic P-wave phase first arrival automatic picking method", patent number 201410242342.4; "a low signal-to-noise ratio acoustic emission signal arrival time picking method", patent number 201811321675.0. With the continuous development of deep learning technology, deep learning models have been gradually introduced into the processing methods of waveform recognition and time-of-arrival picking. Although deep learning models have certain challenges in interpretability, compared with traditional machine learning or mathematical derivation methods, deep learning methods have achieved significant improvements in processing speed and recognition accuracy. For example, relevant patents such as "Multi-layer and multi-level mode automatic classification and identification method of mine microseismic waveforms" (patent number 202210694880.1), "Microseismic event identification method and device" (patent number 201710955082.9) and "Tunnel microseismic waveform arrival time picking method based on U-Net neural network" (patent number 202010478269.6) demonstrate advanced methods based on deep learning, and their application in microseismic monitoring and early warning has shown great potential.

[0005] In the rockburst early warning work of deep tunnel engineering, "fast" and "accurate" are the two key indicators to measure the performance of the early warning system. These two indicators are closely related to the speed and accuracy of data analysis. At present, most of the data analysis modules in the microseismic monitoring system operate independently. This modular design requires frequent calls to different modules during data processing, which affects the overall processing speed. In addition, the independent operation of each module ignores the inherent correlation between tasks and cannot fully utilize the complementary information between modules. Taking waveform recognition as an example, if a waveform is identified as a rupture waveform, it should have obvious P-wave and S-wave arrival time and end position (point E) at the same time. Similarly, if the positions of P-wave, S-wave and point E are not clear, the probability of the waveform being identified as a rupture should be reduced. Therefore, how to effectively integrate the outputs of each module and fully consider the interrelationships between modules has become an important research content to improve the performance of the microseismic monitoring system. Summary of the invention

[0006] According to the technical problems proposed above, an integrated analysis method for tunnel microseismic waveform recognition, arrival time picking, and waveform end prediction is provided. The present invention mainly focuses on the structural design and training method optimization of an integrated analysis model for microseismic waveform recognition, arrival time picking, and waveform end prediction. By ingeniously designing the model structure and optimizing the algorithm, the present invention can simultaneously extract the shared features of the three tasks and complete the prediction of waveform category recognition, P-wave arrival time, S-wave arrival time, and the position of point E in one execution process, thus effectively avoiding the lengthy multi-stage processing flow.

[0007] The technical means adopted by the present invention are as follows:

[0008] An integrated analysis method for tunnel microseismic waveform recognition, arrival time picking, and waveform end prediction, comprising:

[0009] Obtain the microseismic single-component waveform to construct a waveform data set and label the waveform;

[0010] Preprocess the data in the waveform data;

[0011] Construct an integrated analysis model for microseismic waveforms;

[0012] Optimize the integrated analysis model for microseismic waveforms;

[0013] Use the optimized integrated analysis model for microseismic waveforms to analyze the waveform data and output the waveform category, P-wave arrival time, S-wave arrival time, and the position of point E.

[0014] Further, the labeling of the waveform includes waveform type, P-wave arrival time, S-wave arrival time, and the position of point E.

[0015] Further, the preprocessing includes: data clipping, data augmentation, and data denoising;

[0016] The data clipping means clipping the original data into equal-length samples; the data augmentation means horizontally flipping the waveform to increase the number of waveforms; the data denoising means removing the noise outside the broken waveform frequency band through wavelet denoising.

[0017] Further, the integrated analysis model for microseismic waveforms includes a waveform recognition model, an arrival time picking model, and a broken signal vibration end picking model.

[0018] Further, the waveform recognition model extracts features from the input microseismic waveform, flattens the data features into one-dimensional features through global average pooling, and uses a fully connected layer to map the one-dimensional features to the output dimension to achieve the prediction of the waveform category; the feature extraction is expressed as:

[0019] fea = F(x)

[0020] Among them, fea represents the data features learned by the deep learning model, F(x) represents the non-linear mapping function of the deep learning model, with the input being x and the output being the feature fea;

[0021] The waveform category prediction is expressed as:

[0022]

[0023] Among them, The category result predicted through the linear layer L class includes two waveform categories: noise and rupture.

[0024] Furthermore, the arrival time picking model adopts a U-Net structure, extracts feature mutation information through feature subtraction, and outputs the P-wave arrival time, S-wave arrival time, and E-point position after integration through a local attention mechanism:

[0025] p, s, e = D(mask·x)

[0026] Among them, D(mask·x) represents the non-linear mapping function of the arrival time picking model, with the input being mask·x, mask representing the mask information for intercepting x, and x representing the original data input.

[0027] Furthermore, the rupture signal vibration end picking model introduces a vibration attenuation equation as the basis for prediction, performs regression prediction on the attenuation coefficient in the vibration attenuation equation through a fully connected layer neural network, obtains the attenuation process conforming to the rupture waveform, and predicts the end position through calculation:

[0028] y = y0e -kt

[0029] k = f(a)

[0030] Among them, y represents the amplitude after vibration attenuation, y0 represents the vibration amplitude in the initial state, k represents the attenuation coefficient, t represents the sampling point; f(a) represents the mapping function of the deep neural network, with the input being the set of all peaks of the absolute value of the original input waveform, and the output being the attenuation coefficient k.

[0031] Furthermore, the microseismic waveform integrated analysis model is optimized, specifically including:

[0032] For the waveform recognition model and the arrival time picking model, cross-entropy is used as the loss function to calculate the cross-entropy between the predicted value and the true value to optimize the model; for the rupture signal vibration end picking model, the L2 distance between the predicted value and the true value is used as the loss to optimize the model.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] The integrated analysis method for tunnel microseismic waveform recognition, arrival time picking, and waveform end prediction provided by the present invention trains a model through a sample data set with manual annotations. The annotation content includes key information such as waveform category, arrival time points of P-wave and S-wave, and the position of point E. To ensure the unity and consistency of the data, the present invention designs a data preprocessing method, covering processing steps such as data cropping, data augmentation, and wavelet denoising. On this basis, an integrated microseismic analysis model is constructed, integrating multiple originally independent tasks together. Through a single data input, multiple prediction results can be obtained at once, covering waveform category, arrival time points of P-wave and S-wave, and the vibration end point E of the rupture signal. This innovative method effectively eliminates the redundant operations of repeatedly calling the model and reading repeated data in the original system, significantly improving the operation efficiency of the microseismic monitoring system. In addition, the present invention also proposes an integrated model optimization technology, realizing the deep integration of the waveform recognition and arrival time picking models. Specifically, waveform recognition provides a recommended area for arrival time picking, and the arrival time picking result corrects the waveform recognition result, greatly enhancing the robustness and accuracy of data analysis.

[0035] For the above reasons, the present invention can be widely promoted in the field of tunnel microseismic monitoring technology. Description of the Drawings

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0037] Figure 1 It is a flowchart of the integrated analysis method for tunnel microseismic waveform recognition, arrival time picking, and waveform end prediction in the present invention.

[0038] Figure 2 It is a schematic diagram of the collaborative enhancement optimization technology of the microseismic waveform integrated analysis model in the present invention.

[0039] Figure 3 It is a diagram of the adaptive cropping calculation process in the present invention.

[0040] Figure 4 It is a structure diagram of the microseismic waveform integrated analysis model in the present invention.

[0041] Figure 5 It is a structure diagram of the waveform recognition module of the microseismic waveform integrated analysis model in the present invention.

[0042] Figure 6 It is a structure diagram of the arrival time picking module of the microseismic waveform integrated analysis model in the present invention.

[0043] Figure 7 This is the structure diagram of the rupture signal vibration end picking module of the microseismic waveform integrated analysis model in the present invention. Specific implementation manners

[0044] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part rather than all of the embodiments of the present invention. The description of at least one exemplary embodiment below is actually only illustrative and in no way restricts the present invention and its application or use. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0046] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of the described features, steps, operations, devices, components and / or their combinations.

[0047] Unless otherwise specifically stated, the relative arrangements of the components and steps, numerical expressions and values set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be clear that, for the convenience of description, the dimensions of the various parts shown in the drawings are not drawn in actual proportional relationship. Technologies, methods and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods and devices should be regarded as part of the authorization specification. In all the examples shown and discussed here, any specific values should be construed as merely exemplary and not as a limitation. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that: like reference numerals and letters denote like items in the following drawings, and thus, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0048] As shown in 1, the present invention provides an integrated analysis method for tunnel microseismic waveform recognition, arrival time picking and waveform end prediction, including:

[0049] Obtain the microseismic single-component waveform to construct a waveform dataset, and annotate the waveform;

[0050] Specifically, as a preferred embodiment of the present invention, the annotation of the waveform includes waveform type, P-wave arrival time, S-wave arrival time, and the position of point E. The waveform types include rupture and noise. The rupture waveform is annotated as [1,0], and the noise is annotated as [0,1]; the tags for the arrival time are tensors of the same length as the input waveform, where the corresponding arrival time position value is 1 and other values are 0. For noise, it is annotated as an equal-value tensor with all values being the reciprocal of its length. For the waveform end position, its annotation is only the position of the end of the rupture waveform, which is an integer value.

[0051] Preprocess the data in the waveform data;

[0052] Specifically, as a preferred embodiment of the present invention, the preprocessing includes: data clipping, data augmentation, and data denoising;

[0053] The data clipping means clipping the original data into equal-length samples to ensure that the model still maintains strong recognition ability when the position of the rupture main body shifts; the data augmentation means horizontally flipping the waveform, which effectively increases the number of new waveforms without changing any information; the data denoising means removing the noise outside the rupture waveform frequency band through wavelet denoising, statistically analyzing the noise waveforms unique to the site, and then adding typical noise to the clean rupture waveform to obtain a noisy rupture waveform. Using these waveforms to train the model can effectively increase the robustness of the model to analyze noisy rupture waveforms.

[0054] In implementation, the data preprocessing methods are divided into two categories: data preprocessing for training the model and data preprocessing for on-site analysis. The common preprocessing methods for both include data clipping and wavelet denoising; the preprocessing of the training dataset adds data augmentation to improve the generalization ability of the model.

[0055] Construct an integrated microseismic waveform analysis model;

[0056] Specifically, as a preferred embodiment of the present invention, the integrated microseismic waveform analysis model includes a waveform recognition model, an arrival time picking model, and a rupture signal vibration end picking model.

[0057] Specifically, as a preferred embodiment of the present invention, the waveform recognition model extracts features from the input microseismic waveform, flattens the data features into one-dimensional features through global average pooling, and maps the one-dimensional features to the output dimension using a fully connected layer to achieve the prediction of the waveform category; the feature extraction is expressed as:

[0058] fea = F(x)

[0059] Among them, fea represents the data features learned by the deep learning model, F(x) represents the non-linear mapping function of the deep learning model, with the input being x and the output being the feature fea;

[0060] The waveform category prediction is expressed as:

[0061]

[0062] Among them, The category result predicted through the linear layer L class includes two waveform categories: noise and rupture.

[0063] In specific implementation, as a preferred implementation manner of the present invention, the arrival time picking model adopts a U-Net structure, extracts feature mutation information through feature subtraction, and outputs the P-wave arrival time, S-wave arrival time, and E-point position after integration through a local attention mechanism:

[0064] p, s, e = D(mask·x)

[0065] Among them, D(mask·x) represents the non-linear mapping function of the arrival time picking model, with the input being mask·x, mask representing the mask information for intercepting x, and x representing the original data input.

[0066] In specific implementation, as a preferred implementation manner of the present invention, the rupture signal vibration end picking model introduces a vibration attenuation equation as the basis for prediction, performs regression prediction on the attenuation coefficient in the vibration attenuation equation through a fully connected layer neural network, obtains the attenuation process conforming to the rupture waveform, and predicts the end position through calculation:

[0067] y = y0e -kt

[0068] k = f(a)

[0069] Among them, y represents the amplitude after vibration attenuation, y0 represents the vibration amplitude in the initial state, k represents the attenuation coefficient, t represents the sampling point; f(a) represents the mapping function of the deep neural network, with the input being the set of all peaks of the absolute value of the original input waveform, and the output being the attenuation coefficient k.

[0070] Optimize the microseismic waveform integrated analysis model;

[0071] In specific implementation, as a preferred implementation manner of the present invention, optimizing the microseismic waveform integrated analysis model specifically includes:

[0072] For the waveform recognition model and the arrival time picking model, the cross-entropy is used as the loss function to calculate the cross-entropy between the predicted value and the true value, and the model is optimized. For the rupture signal vibration end picking model, the L2 distance between the predicted value and the true value is used as the loss to optimize the model.

[0073] In implementation, the present invention also adopts a collaborative enhancement optimization technique. The feature of waveform recognition is mapped to the main area of the rupture waveform by using a self-supervised training mode, and then the main area is cropped by the adaptive cropping method proposed by the present invention and input into the arrival time picking module to reduce the difficulty of picking the arrival time point by the arrival time picking module. At the same time, the prediction result of the arrival time picking module is fed back to the waveform recognition module after calculating its quality coefficient to enhance the robustness of the waveform recognition module.

[0074] The optimized microseismic waveform integrated analysis model is used to analyze the waveform data, and the waveform category, P-wave arrival time, S-wave arrival time, and E-point position are output.

[0075] Example 1

[0076] In this example, microseismic single-component waveforms are collected as a dataset for model training. Among them, 12,857 rupture data are collected and 12,857 noise data are added to the dataset to ensure data balance during the training process. Then all the rupture waveforms are labeled. Among them, the label of each data includes three attributes: waveform type (rupture or noise), P-wave and S-wave arrival times (the label is a vector of the same length as the input, the corresponding arrival time position value is 1, and other values are 0. For noise, all values are the reciprocal of its length), and the E-point position (the sampling point at the end position of the rupture waveform). The STA / LTA method is used as an aid for labeling, and the method of repeated verification by multiple people is used to ensure the correctness of the data labeling in the sample library.

[0077] In the data preprocessing process, first, all the data are cropped (cropped into sequences of the same length) to ensure the dimensional consistency when used as input. If the waveform length is insufficient before cropping, 0 is padded at both ends. Then the data are normalized to ensure the consistency of data distribution. Then the data are filtered by wavelet to remove the trend fluctuation interference caused by low frequency. For the data used to train the model, the present invention adds a data augmentation method, that is, the common noise signals in the engineering are statistically analyzed, and then noise is superimposed on the clean rupture waveforms to obtain enhanced samples.

[0078] For the waveform recognition model, the given input data dimension is (1, 2000). After the first layer of convolution, the dimension becomes (32, 2000). Then, after pooling and the second layer of convolution, a feature output with a dimension of (32, 1000) is obtained. At the same time, the outputs of the above two layers of features are added together to ensure that the model can remember the data features in the shallow layer. And so on, the present invention designs a waveform recognition model with a 5-layer structure to extract features from the input microseismic data, as Figure 5 shown. Its structure from top to bottom is respectively: a convolutional layer with a convolutional kernel of 7 and a stride of 3, a batch normalization layer, a GELU activation layer, a max pooling layer, a convolutional layer with a convolutional kernel of 5 and a stride of 2, a batch normalization layer, a GELU activation layer, a max pooling layer, and so on, until the obtained output dimension is (64, 256, 5). After feature extraction, the data features are flattened into one-dimensional features through global average pooling, and then mapped to an output with a dimension of (64, 2) through a fully connected layer, so as to realize the prediction of waveform categories.

[0079] When it comes to the arrival time picking model D, convolution is also used as the basic component. The present invention adopts a three-layer U-Net structure, and at the same time changes the feature addition process in the U-Net to feature subtraction to extract information on feature mutations. As Figure 6 shown, its structure from top to bottom is respectively: a convolutional layer with a convolutional kernel of 5 and a stride of 2, a batch normalization layer, a GELU activation layer, a max pooling layer, a convolutional layer with a convolutional kernel of 5 and a stride of 2, a batch normalization layer, a GELU activation layer, a max pooling layer, and so on, until the obtained output dimension is (64, 128, 32). Then, an upsampling operation is performed on this feature, that is, upsampling, a convolutional layer with a convolutional kernel of 5 and a stride of 2, a batch normalization layer, and a GELU activation layer are a basic combination. Until the input data is restored to the original size through upsampling. Among them, in the middle process of downsampling and upsampling, data with the same dimension needs to be subtracted. Finally, the arrival times of P waves and S waves and the prediction results of point E are respectively output after being integrated by three local attention mechanisms.

[0080] The rupture signal vibration end picking model also needs to utilize the main body position of the waveform. Considering that the amplitude at the end position of the waveform is small and is easily affected by weak noise, it is more difficult to label point E. Therefore, the present invention considers using unlabeled data for point E prediction. The specific idea is to introduce the vibration attenuation equation in physics as the basis for prediction, and perform regression prediction on the attenuation coefficient in the vibration attenuation equation through a simple fully connected layer neural network, so as to obtain the attenuation process that conforms to the rupture waveform. When the waveform amplitude decays to 5% of the maximum amplitude, it is regarded as the end position of the waveform, and the end position can be predicted through simple mathematical calculations. The specific implementation process is as Figure 7As shown in the figure, the main body of the waveform obtained by clipping the waveform using the mask obtained in 2 is then subjected to peak extraction. After that, outliers are removed through opening and closing operations. Finally, the remaining data is sorted and the first 20 numbers are selected to establish an array a. Using this array as the input, a single-layer fully connected layer is used to predict the coefficient k of the vibration attenuation equation. The obtained attenuation coefficient k is substituted into the equation to calculate the amplitude at each point and establish an array b. When the amplitude is less than 5% of the maximum amplitude, it is regarded as the waveform end position.

[0081] Through the reasonable design of the analysis model structure, the waveform recognition, arrival time picking, and E-point prediction tasks are organically combined. Waveform recognition provides an alternative band for arrival time picking, reducing the influence of redundant information on arrival time picking and improving the picking accuracy. At the same time, the prediction quality of the arrival time picking module can, in turn, fine-tune the prediction results of the waveform recognition module, ensuring the robustness of waveform recognition.

[0082] Embodiment 2

[0083] In this embodiment, collaborative enhancement optimization is adopted in two parts of the microseismic waveform integrated analysis model optimization: waveform recognition provides an alternative segment for arrival time picking and arrival time picking corrects the prediction results of waveform recognition.

[0084] The present invention believes that the main parts concerned by the waveform recognition and arrival time picking models are the same, that is, the main body position of the rupture waveform. In the past methods, whether it is the waveform recognition method or the arrival time picking method, the whole input data is concerned and then predicted. This often contains a large amount of redundant information, causing a huge waste of computing power and also causing strong interference to the model prediction process. The present invention takes the features of the waveform recognition module as the input, and only uses a single-layer fully connected layer to map the features to a number v between 0 and 1800. Then, this value is used as the starting position of the main body position of the rupture waveform, and two reparameterization techniques designed by the present invention are used to block and clip the useless bands of the original waveform. To ensure the accuracy of the starting position v, a self-supervised training method is introduced here, and the calculation method is as follows:

[0085] Obtaining v: v = L v (fea)

[0086] Obtaining m: m = δ(T - v)·δ(v + 200 - T)

[0087] Obtaining mask: mask = δ(I - Z + v)·δ(Z - I - v)

[0088] Optimization process of v: min MMD(F(x), F(m·x))

[0089] In the above formula, δ represents the sigmoid function, T is a fixed tensor with a dimension of (1, 1999), and its values are sequential integers from 0 to 1999. Z is a (256, 1999) tensor obtained by repeating T 256 times on the horizontal axis, and I is a tensor similar to T but with a dimension of (1, 256). By adding the invariant tensors I, Z, and T to the optimization parameter v, the design ensures the continuity of the gradient during the occlusion and cropping processes, enabling the model to be conveniently optimized in an end-to-end manner. The calculation process of the mask is as Figure 2 shown.

[0090] On the other hand, the output of the arrival-time picking module can, in turn, correct the prediction results of waveform recognition. Specifically, the main feature of a rupture waveform is that it consists of a P-wave and an S-wave. The arrival times of the P-wave and S-wave, as the starting points of the regular vibrations that occur in the sensor when the P-wave and S-wave reach the sensor, essentially manifest as mutation points in the continuous waveform data. The more obvious the mutation points, the more obvious the arrival times of the P-wave and S-wave, and the greater the probability that the waveform is a rupture wave. After analyzing a large amount of data, it is found that the quality of the arrival-time prediction results is inversely proportional to the predicted probability distribution map output by the model. Therefore, the present invention designs a set of arrival-time picking quality evaluation indicators to evaluate the obviousness of arrival-time features. Specifically, the entropy of the predicted probabilities of the P-wave and S-wave and is used as its quality evaluation indicator. The larger the entropy of the prediction result, the lower its quality, that is, the smaller the certainty. This evaluation indicator is used to fine-tune the probability output of the waveform recognition module during the training process . If the waveform recognition module identifies a waveform as a rupture, but the quality evaluation indicator of the prediction result of the arrival-time picking module is extremely low, then the waveform recognition result needs to be appropriately adjusted, and vice versa.

[0091] 1) Arrival-time evaluation indicator:

[0092] 2) Method for adjusting recognition results:

[0093] 3) Optimization process of the waveform recognition module:

[0094] Through the above two-part design, the present invention cleverly unifies waveform recognition and the arrival-time picking module, achieving collaborative enhancement optimization of the two. In terms of the selection of the optimization algorithm, the present invention uses the Adam algorithm to optimize the model, that is, to minimize the combination of the above losses. After training for 100 epochs, an integrated microseismic waveform analysis model can be obtained.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An integrated analysis method for tunnel microseismic waveform recognition, arrival time picking, and waveform end prediction, characterized in that Including: Obtain the microseismic single-component waveform to construct a waveform dataset, and annotate the waveform; Preprocess the data in the waveform data; Construct an integrated microseismic waveform analysis model; Optimize the integrated microseismic waveform analysis model; Use the optimized integrated microseismic waveform analysis model to analyze the waveform data, and output the waveform category, P-wave arrival time, S-wave arrival time, and E-point position.

2. The integrated analysis method for tunnel microseismic waveform recognition, arrival time picking and waveform end prediction according to claim 1, characterized in that The annotation of the waveform includes waveform type, P-wave arrival time, S-wave arrival time, and E-point position.

3. The integrated analysis method for tunnel microseismic waveform recognition, arrival time picking and waveform end prediction according to claim 1, characterized in that The preprocessing includes: data clipping, data augmentation, and data denoising; The data clipping means clipping the original data into equal-length samples; the data augmentation means horizontally flipping the waveform to increase the number of waveforms; the data denoising means removing the noise outside the rupture waveform frequency band through wavelet denoising.

4. The integrated analysis method for tunnel microseismic waveform recognition, arrival time picking and waveform end prediction according to claim 1, characterized in that The integrated microseismic waveform analysis model includes a waveform recognition model, an arrival time picking model, and a rupture signal vibration end picking model.

5. The integrated analysis method for tunnel microseismic waveform recognition, arrival time picking, and waveform end prediction according to claim 4, characterized in that, The waveform recognition model extracts features from the input microseismic waveform, flattens the data features into one-dimensional features through global average pooling, and uses a fully connected layer to map the one-dimensional features to the output dimension to realize the prediction of the waveform category; the feature extraction is expressed as: fea = F(x) where fea represents the data features learned through the deep learning model, F(x) represents the non-linear mapping function of the deep learning model, the input is x, and the output is the feature fea; The waveform category prediction is expressed as: Among them, The category results predicted by the linear layer L class include two waveform categories: noise and rupture.

6. The integrated analysis method for tunnel microseismic waveform recognition, arrival time picking and waveform end prediction according to claim 4, characterized in that The arrival time picking model adopts a U-Net structure, extracts feature mutation information through feature subtraction, and outputs the P-wave arrival time, S-wave arrival time, and E-point position after integration through a local attention mechanism: p, s, e = D(mask·x) where D(mask·x) represents the non-linear mapping function of the arrival time picking model, the input is mask·x, mask represents the masking information for intercepting x, and x represents the original data input.

7. The integrated analysis method for tunnel microseismic waveform recognition, arrival time picking, and waveform end prediction according to claim 4, wherein The rupture signal vibration end picking model introduces a vibration attenuation equation as the basis for prediction, performs regression prediction on the attenuation coefficient in the vibration attenuation equation through a fully connected layer neural network, obtains the attenuation process conforming to the rupture waveform, and predicts the end position through calculation: y = y0e -kt k = f(a) where y represents the amplitude after vibration attenuation, y0 represents the vibration amplitude in the initial state, k represents the attenuation coefficient, t represents the sampling point; f(a) represents the mapping function of the deep neural network, the input is the set of all peaks of the absolute value of the original input waveform, and the output is the attenuation coefficient k.

8. The integrated analysis method for tunnel microseismic waveform recognition, arrival time picking, and waveform end prediction according to claim 1, wherein Optimizing the integrated microseismic waveform analysis model specifically includes: For the waveform recognition model and the arrival time picking model, use cross-entropy as the loss function, calculate the cross-entropy between the predicted value and the true value, and optimize the model; for the rupture signal vibration end picking model, use the L2 distance between the predicted value and the true value as the loss to optimize the model.

Images