An intelligent rockburst early warning method based on microseismic point cloud feature learning

By using a combination of fast approximation convolution and time coding in the intelligent rock burst warning, the space-time characteristics of microseismic event clouds are processed, and the problem of insufficient analysis of microseismic event clouds in the existing technology is solved, and high-precision rock burst warning is achieved.

CN116030598BActive Publication Date: 2024-11-12DALIAN UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310001570.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-01-03
Publication Date
2024-11-12
Estimated Expiration
2043-01-03

AI Technical Summary

Technical Problem

Existing machine learning methods are difficult to effectively deal with microseismic event clouds, resulting in insufficient space-time correlation analysis of rock burst warnings and the inability to accurately predict rock burst risks.

Method used

An intelligent early warning method for rock burst for micro-seismic point cloud feature learning is proposed. Using fast approximation convolution (FAC) technology and time coding, a rock burst pattern recognition network based on point cloud is constructed, and a rock is realized through high-dimensional point cloud feature learning is achieved. Explosive warning.

Benefits of technology

It achieves a rock burst warning accuracy of up to 90.1% in hard rock tunnel projects, prevents information loss caused by special feature extraction, and can reliably conduct early rock burst warning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116030598B_ABST
    Figure CN116030598B_ABST
Patent Text Reader

Abstract

A rockburst intelligent early warning method based on microseismic point cloud feature learning belongs to the field of rock engineering disaster monitoring and intelligent early warning technology. A microseismic monitoring system combining wired and wireless data transmission is established in the tunnel. The machine learning model regards microseismic data as a high-dimensional point cloud, and proposes a fast approximation convolution FAC method to effectively learn features from the original data. Taking into account the three elements of microseismic signal time, space, and intensity, a FAC-based point cloud model is proposed to directly learn all the original monitored microseismic data and perform rockburst early warning to prevent information loss caused by special feature extraction. Microseismic data are collected as training and test data sets for this model, and two competing models are constructed, including a time series model and a traditional threshold model, to prove the effectiveness of the model; the FAC-based model does not encode the timestamp of the microseismic data to study the importance of time. The model can reliably provide a rockburst early warning accuracy of up to 90.1% for hard rock engineering tunnels.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of rock engineering disaster monitoring and intelligent early warning, and in particular relates to a rockburst intelligent early warning method based on microseismic point cloud feature learning. Background Art

[0002] Rockbursts often occur during the excavation of deep tunnels with high stress, which not only delays the construction period but also threatens personal safety. Rockbursts are a geological disaster caused by surrounding rocks. Due to their high in-situ stress, they store huge elastic energy. After excavation, the energy is suddenly released in a short period of time in the hard and brittle rock.

[0003] Due to the extremely complex mechanism of rockburst, its prediction has become a common challenge faced by engineers around the world. Before a rockburst occurs, microseismic signals are generated, which can be detected by highly sensitive sensors; these microseismic signals contain a lot of information that can reflect the evolution of rock damage and geological defects. By monitoring and analyzing microseismic signals, the three-dimensional spatiotemporal distribution of rock mass damage, as well as some source parameters such as energy, seismic moment and stress drop, can be obtained, providing opportunities for rockburst prediction. In order to achieve real-time rockburst warning, the technology that ultimately needs to be overcome is to predict the risk of such disasters based on the spatiotemporal distribution of observed data and its related parameters (energy, seismic moment, stress drop, etc.).

[0004] Microseismic events are spatially discrete before rockbursts, but it is not clear what distribution pattern will lead to rockbursts. In general, if a large number of microseismic events or rockbursts occur, the surrounding rock releases accumulated energy, the stress level decreases, and the risk of rockquakes also decreases. On the contrary, in another case, the damage will lead to greater stress concentration, which will accelerate the occurrence of rockbursts and increase the risk of geological disasters. Although the microseismic events are distributed within a certain range near the tunnel, the spatiotemporal distribution shows great irregularity and strong discreteness (in the form of a point cloud). Each point is captured and generated independently by the microseismic monitoring system, and the distance between the points is not fixed, showing greater randomness. These points contain a lot of information related to rock mass damage, such as location coordinates and energy. It is difficult to determine the relationship between these points and rockbursts theoretically or experimentally. At present, most machine learning methods for rockburst early warning focus on the identification and extraction of specific rockburst features, while ignoring the spatiotemporal correlation between these microseismic events.

[0005] Machine learning methods based on point cloud data analysis may be more suitable for rockburst warning. Since existing models cannot process microseismic event clouds, point cloud analysis methods have not been used in rockburst warning research so far. The key to rockburst warning is to focus on the characteristics of microseismic events clustered in a limited space-time segment, rather than analyzing the shape of the point cloud, which is the main idea of ​​the traditional point cloud model. Therefore, in order to make full use of the spatiotemporal information of the microseismic event point cloud, the present invention proposes a fast approximate convolution (FAC) method that is more suitable for microseismic data processing, and combines it with time encoding to construct a point cloud-based rockburst pattern recognition network. In addition, the model was tested using microseismic monitoring data and field rockburst records obtained from a hard rock tunnel project, and the model has good early rockburst warning capabilities. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes: a rockburst intelligent early warning method based on microseismic point cloud feature learning, comprising the following steps:

[0007] S1. Establish a microseismic monitoring system;

[0008] S2, randomly select a data set;

[0009] S3, the rockburst prediction problem is formulated as a recognition task on high-dimensional point cloud patterns, taking high-dimensional point clouds as input and learning features from high-dimensional point clouds to identify different concentration patterns of event points;

[0010] S4. Perform convolution on the high-dimensional point set for feature learning and predict rockburst.

[0011] Furthermore, in step S1, the configuration of the microseismic monitoring system is as follows: the microseismic monitoring system is set as a workstation inside the tunnel and a workstation outside the tunnel. The workstation inside the tunnel consists of an acceleration sensor and an integrated box. The acceleration sensor is used to detect microseismic events in the rock. The acceleration sensor is installed in the borehole and fixed with an anchor. The integrated box includes a Paladin digital signal acquisition system and a signal processing system, which are responsible for collecting and recording microseismic data, and wirelessly transmitting the microseismic information to the workstation outside the tunnel. The workstation outside the tunnel consists of a wireless signal real-time receiving system and an artificial intelligence signal processing system. The wireless signal real-time receiving system receives the microseismic data wirelessly sent by the workstation inside the tunnel in real time. The artificial intelligence signal processing system consists of a microseismic event recognition and classification module, a time-picking module, a transfer learning module, a source positioning module, and an artificial intelligence-based early warning module. These modules process the observed microseismic information and provide rock burst warnings in a timely manner.

[0012] The internal workstation adopts a mobile layout. The tunnel face is constantly moving forward, and the position of the microseismic monitoring equipment is adjusted according to the position of the face to ensure that the acceleration sensor can monitor the damage in the surrounding rock in real time.

[0013] Furthermore, in step S2, microseismic events of each day are selected as samples, including positive samples and negative samples. If a rockburst occurs on a certain day, the sample is determined as a positive label, and if a rockburst does not occur on a certain day, the sample is determined as a negative label. Each sample contains dozens of microseismic events. The data of the first 5 days of the prediction are input into the neural network as a sample set. The distribution pattern of the point cloud is made to correspond to the risk of rockburst through the timestamp information of each microseismic event. Finally, all the features are connected and input into the model together.

[0014] Furthermore, in step S3, the microseismic event is a three-dimensional point cloud with high-dimensional properties, and the microseismic event is represented by spheres of different colors. The microseismic event is represented as a three-dimensional point set, in which each point has its own three-dimensional coordinates (x, y, z) and feature channels, and the color and size of each point represent the magnitude and energy respectively. Considering 11 different features of each event point, namely, seismic momentum, event energy, source radius, AspRadius, static stress, apparent stress, dynamic stress, maximum displacement, peak velocity, peak acceleration and local amplitude, each event is independently regarded as a high-dimensional point cloud set, and the spatiotemporal distribution of the high-dimensional point cloud is related to the rock burst risk. The rock burst prediction problem is restated as a pattern recognition task on a high-dimensional point cloud set. The method uses a high-dimensional point cloud as input, rather than a highly integrated descriptor or time series, identifies different patterns in the event point set, and learns features from the high-dimensional point cloud.

[0015] Furthermore, in step S3, the steps of point set deep learning are as follows: Establish the architecture of PointNet: The input of PointNet is the original point cloud P N × D , N is the number of points in the cloud, D is the dimension, representing the (x, y, z) of each point. Two T-Nets are used to model the unordered features with learnable weights. In the last step, a 1024-dimensional feature vector is generated using the maximum pooling symmetric function to aggregate global information. The feature vector represents the feature descriptor of the input for recognition tasks. Further improvements to PointNet are achieved by modeling local region structures, which are used to capture hierarchical local structures in a manner similar to grid convolutions.

[0016] Furthermore, in PointNet++, first in P N × 3The farthest point sampling algorithm is used to select the centroid, and the ball query is used to select the nearest neighbor of each centroid; then, the point network is applied to all local areas to produce a vector describing the characteristics of the area; this process is repeated in a hierarchical form, thereby reducing the resolution of the points as the process deepens. In the last layer of the hierarchy, the entire point feature generates a global feature vector through PointNet.

[0017] Furthermore, in step S4, the convolution step on the high-dimensional point set for feature learning is as follows: According to the mathematical definition of convolution, for the discrete version in CNN, the convolution function is discretized into regular grids, which are composed of several independent pixels with integral coordinates. The discrete version is extended to three-dimensional space, and a multi-layer perceptron MLP is used to approximate the weight function, as shown in equation (4):

[0018]

[0019] (x i ,y i ,z i ) is the coordinate of the i-th neighboring point, and the multilayer perceptron MLP is a fully connected neural network with multiple hidden layers. Represents the weight vector associated with this point, where d is the dimension of the feature. First, the KNN algorithm is applied to obtain each local area, assuming that all points are taken as the centroid in turn. Then, for each local area, each point is subtracted from the coordinates of the centroid. Finally, each area is scaled normalized.

[0020] Furthermore, in step S4, it is assumed that the timestamps of a local point set with n points are t0, t1, t2, ..., t n , equation (5) is applied to obtain the normalized temporal encoding,

[0021] T n =scale(t i -t0) (5)

[0022] Where “scale” represents a normalization function that normalizes the values ​​of the set to [0,1] by n With the coordinate vector (x i ,y i ,z i ) n The four-dimensional input data are concatenated and fed into the weight function. The weights obtained have the same dimension as the feature map. These weights are multiplied by the corresponding features and the results are averaged and aggregated to become part of the new feature map. The output dimension of the new feature map is obtained by superimposing several continuous filters.

[0023] Furthermore, an operation method is proposed: fast approximation convolution FAC, where fast approximation convolution FAC is regarded as a learnable approximation of high-dimensional continuous convolution operation, the input feature channel is c_in, the output channel is c_out, the weight vectors generated by different filters in a specific layer are related, and a mapping function is constructed to project one feature subspace to another feature subspace;

[0024] The entire process of the fast approximation convolution FAC is as follows: for the microseismic event cloud, each local area shares the same weight function learned using a multilayer perceptron MLP; first, a multilayer perceptron MLP with a hidden layer of 16 neurons is trained to fit the first continuous filter formula (6), which contributes to the first dimension of the output channel. No activation function or batch normalization layer is seen in equation (6). Then, the first weight vector is input to the w2ws network, and the output contains the o-1 vector, which is the other o-1 feature in the output feature space. The w2ws network is a multilayer perceptron MLP with two hidden layers, both of which have 8 neurons. In addition, the w2ws network has activation and batch normalization functions. The ownership vectors around this point are obtained and multiplied by the previous feature vector and aggregated to generate each dimension of the output vector. It is assumed that the w2ws transformation is performed by a function f and contains two points in this local point set: p1 and p2. In addition, the two weight vectors generated by equation (6) for the two points are W 11 and W 12 , the fast approximation convolution FAC process is described by equation (7), in which O1 and O2 include all output vectors except the vector produced by the first multilayer perceptron MLP. In addition, the training w2ws network projects the first feature vector onto other o-1 feature vectors instead of projecting the first weight vector.

[0025] V=W2(W1(x,y,z,t)+b1)+b2 (6)

[0026]

[0027] Where, O1=(W 21 ,W 31 ,…,W o1 ) 1n

[0028] Instead of being stored in memory, the continuous weight generation process is divided into two parts: the first step is to train a common filter, and the second part is to project this filter onto other o-1 feature vectors based on the first vector.

[0029] Further, in step S4, the fast approximation convolution FAC is directly applied to extract features from the microseismic event cloud, and several such layers are stacked to build a final architecture to perform rockburst prediction, which consists of three fast approximation convolution FAC layers and a classification layer. The input data is divided into two parts: one group contains 11 event indicators, called feature groups, and the other group is a coordinate set, called coordinate groups. The coordinate groups are input into the KNN module to generate local area sets, and the number in the KNN module is set to 15. Then, these sets are encoded with timestamps by equations. These local areas are reused on all convolutional layers to generate corresponding weights, which are multiplied by feature branches. The aggregation layer calculates the average of all point feature vectors to obtain a 256-dimensional global feature vector. The classification layer is a multi-layer perceptron MLP with softmax function activation.

[0030] The beneficial effects of the present invention are as follows: the present invention first establishes a microseismic monitoring system combining wired and wireless data transmission in the tunnel, the machine learning model regards the microseismic data as a high-dimensional point cloud, and proposes a new fast approximation convolution FAC method to effectively learn features from the original data. The proposed FAC-based point cloud model can directly learn all the original monitored microseismic data, and then perform rock burst warning, preventing the loss of information caused by special feature extraction, collecting microseismic data to prepare data sets for training and testing of this model, and constructing two competing models, including a time series model and a traditional threshold model, to prove the effectiveness of the model; the FAC-based model does not encode the timestamp of the microseismic data to study the importance of time. The results show that the model can reliably provide a rock burst warning accuracy of up to 90.1% for hard rock engineering tunnels. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is the architecture of the wave dot network of the present invention;

[0032] Figure 2 It is a schematic diagram of the principle of three-dimensional convolution operation of the present invention;

[0033] Figure 3 The present invention converts the global coordinates into local coordinates before inputting the global coordinates into the neural network;

[0034] Figure 4 It is a schematic diagram of the FAC method used in the feature learning process of aggregated point set data of the present invention;

[0035] Figure 5 is a schematic diagram of the entire architecture of the present invention for processing microseismic point cloud data to identify rockburst patterns;

[0036] Figure 6 It is the temporal regularity of the distribution of microseismic events along the tunnel of the present invention;

[0037] Figure 7 There is one example where the FAC-based model provides correct results; this is contrary to traditional time series models;

[0038] Figure 8 It is an example of the present invention to input data into different models for comparison;

[0039] Fig. 9 is the first counterintuitive example where the proposed model can make correct predictions while human experts cannot;

[0040] Fig.10 is the second counterintuitive example where the proposed model can make correct predictions while human experts cannot; DETAILED DESCRIPTION

[0041] The present invention comprehensively considers the three elements of microseismic signal time, space and intensity, and proposes a rockburst intelligent early warning method based on microseismic point cloud feature learning to perform rockburst early warning, thereby preventing information loss caused by special feature extraction, as follows:

[0042] During the tunnel excavation process, rock burst is the main risk factor that threatens the safety of workers and limits the progress of the project. In order to obtain early warning of rock burst, microseismic monitoring technology is used to capture the micro-fractures and aggregates in the surrounding rock to form rock burst damage, so as to ensure construction safety. However, due to the frequent occurrence of rock burst in the tunnel, it is difficult to provide rock burst warning in time by manually processing microseismic data. This is the main purpose of the present invention to use deep learning methods to automatically warn of rock burst.

[0043] Microseismic monitoring configuration:

[0044] Microseismic monitoring equipment requires real-time, continuous and long-term monitoring of surrounding rock damage during tunnel excavation. The Qinling water supply tunnel has high temperature and high humidity environment, long tunnel excavation distance, inconvenient transportation, and difficult system installation, which are not conducive to the long-term continuous operation of the microseismic monitoring system. Therefore, the microseismic monitoring system is set as an internal and external monitoring workstation. The internal workstation consists of two parts: the acceleration sensor detects microseismic events in the rock around the borehole and is fixed with an anchor. The Paladin signal acquisition system and signal processing system are responsible for collecting and recording microseismic data, and wirelessly transmit microseismic information to the workstation outside the tunnel and store it in a box to protect its safety and prevent dust. The workstation outside the tunnel is mainly composed of a wireless signal receiving system and a real-time artificial intelligence (AI) signal processing system. The wireless signal receiving system receives the microseismic data wirelessly sent by the workstation in the tunnel in real time. The real-time artificial intelligence signal processing system consists of a microseismic event recognition and classification module, a time-pickup module, and a transfer learning module, the source location module in the present invention, and an early warning module based on artificial intelligence; these modules process the observed microseismic information and provide early rock burst warnings in a timely manner.

[0045] Considering the complexity of the excavation process and the surrounding rock conditions in the tunnel, as well as the long-term effective work of microseismic monitoring, the internal workstation adopts a mobile layout. The advantage of this method is that it can adjust the position of the microseismic monitoring equipment through the movement of the tunnel face to ensure that the sensor is sensitive enough to the damage in the surrounding rock.

[0046] Dataset Description:

[0047] To prepare the dataset for training and testing the model, the engineering data of the Han-Wei River Diversion Project were collected. During this period, engineers recorded the level and occurrence time of all rockbursts. We selected each day as our sample, which mainly includes two categories: positive samples and negative samples. If a rockburst occurred on a certain day, the sample was determined as a positive label, and vice versa. In addition, each sample also contains dozens of microseismic events. Here, we only input the 5 days before the date of interest as a sample set into the neural network. Therefore, our purpose is to recognize whether the distribution pattern of the point cloud corresponds to the risk of rockburst occurrence. It is worth noting that the dataset also contains very detailed timestamp information for each microseismic event; therefore, in the proposed model, we embed the timestamps as different features, concatenate them with other features, and input them into the model together. Although the temporal resolution of the model is days, the dataset also shows the potential to form more fine-grained datasets, such as those with a temporal resolution of hours or even minutes, which will be the focus of future work.

[0048] Problem Statement:

[0049] Microseismic events can be considered as 3D point clouds with high-dimensional nature. Microseismic events (spheres of different colors) are represented as 3D point sets, where each point has its own 3D coordinates (x, y, z) and other additional feature channels. For example, the color and size of each point represent the magnitude and energy, respectively. In addition to the coordinates, 11 different features of each event point are considered, namely, ground motion, event energy, source radius, AspRadius, static stress, apparent stress, dynamic stress, maximum displacement, peak velocity, peak acceleration, and local amplitude. Each event is independently considered as a high-dimensional point cloud set. It is worth noting that the rockburst risk may be closely related to the spatiotemporal distribution of microseismic events during excavation. However, the traditional approach, that is, simply summing or averaging the indicators, completely deconstructs the spatiotemporal correlation. In addition, time series models, such as recurrent neural networks, do not exploit the 3D spatial location of events, although they consider temporal information. In addition, time series models can only process high-dimensional event data iteratively, which limits their computational efficiency. Another solution tends to aggregate events at the level of a day, which loses the time scale and leads to performance degradation. In summary, the spatiotemporal distribution of high-dimensional point clouds may be related to the risk of rockburst.

[0050] Rockburst risk is reflected in the patterns of related event point clouds. In other words, different patterns correspond to different degrees of risk. In this context, we reformulate the rockburst prediction problem as a pattern recognition task on a high-dimensional point cloud set. The method takes a high-dimensional point cloud as input instead of a highly integrated descriptor or time series. Therefore, to identify different patterns in an event point set, it is key to learn features from the high-dimensional point cloud.

[0051] Deep learning on point sets:

[0052] There are three main challenges in applying deep learning to point cloud data, as follows:

[0053] (1) Point cloud data is irregular, which means that points are not evenly sampled across different regions of the scene. In this case, the variance of the distribution density is large, and some areas have more points than other sparse areas.

[0054] (2) Unlike 2D images or 3D volumes with structured grids, point clouds are unstructured. Each point is sampled and generated independently, and the distance between a point and its neighboring points is not always fixed.

[0055] (3) Point clouds are unordered; the scene representation is independent of the order in which the points are stored. In addition, three other challenges arise in predicting rockbursts by identifying patterns in point clouds.

[0056] First, the point cloud composed of microseismic events is always high-dimensional. Each point carries more than ten features that represent the physical properties of the corresponding event.

[0057] Secondly, microseismic event clouds contain not only spatial locations, but also temporal information that traditional point cloud data do not have. Each event has a timestamp that triggered the event. Finally, and most importantly, the pattern of a general point cloud is usually determined by its shape, that is, by the coordinates of its points.

[0058] However, the indicators of these microseismic events (e.g., energy) are far more important than the coordinates of the predicted rockburst. It is easy to imagine that two event clouds with the same shape but different energy levels will lead to different results. Therefore, rockburst early warning models require spatiotemporal data of microseismic events.

[0059] The goal of this invention is to predict rockbursts by considering all the features of an event simultaneously while respecting their spatiotemporal correlations; this is similar to the functionality of CNNs, which aggregate information along a regular and structured grid, such as an image. However, this convolution operation, which can only be performed on structured data, is very challenging to use for processing point clouds, especially microseismic clouds. Early methods overcame these difficulties by converting point clouds into a structured grid format; these methods can be roughly divided into two categories, voxel-based methods and multi-view based methods. Voxel-based methods (reference) convert a given point cloud into a 3D voxel structure of shape (X, Y, Z) and convolve it with a 3D kernel of size (x, y, z). Voxel-based methods sometimes show good performance, but they suffer from high memory consumption because events are usually triggered sparsely rather than evenly distributed in the scene. In addition, high memory consumption also limits the voxel resolution to usually between 32 cubes and 128 cubes. Other methods convert point clouds into a collection of 2D images and apply existing 2D CNNs to leverage already mature image processing techniques, achieving better performance than voxel-based techniques. However, this approach has an important drawback; namely, they can only transform a normal 3D point cloud into multiple views, but not a high-dimensional point cloud. All of the above problems are easily addressed by transforming point clouds into more structured representations. Related techniques process point clouds directly and overcome the above challenges to some extent.

[0060] Figure 1 Provides information about the architecture of PointNet. The input of PointNet is the original point cloud P N×D, where N is the number of points in the cloud and D is the number of dimensions (usually equal to 3), representing the (x, y, z) of each point. Using two T-Nets, the unordered features are modeled with learnable weights. In the last step, a 1024-dimensional feature vector is generated using the max pooling symmetric function to aggregate global information. The feature vector represents the feature descriptor of the input and can be used for recognition tasks. However, PointNet does not consider local dependencies and cannot capture local structure because only the final max pooling operation can aggregate features. Further improvements to PointNet by modeling local region structures can capture local structures hierarchically in a manner similar to grid convolution; for example, PointNet++ can be used as a representative work. In PointNet++, first in P N×3 The farthest point sampling algorithm is used to select the centroids, and the ball query is used to select the nearest neighbors of each centroid. Then, the point net is applied to all local regions to produce a vector describing the characteristics of the region. Such a process is repeated in a hierarchical form, thereby reducing the resolution of the points as the process deepens. In the last layer of the hierarchy, the entire point feature is passed through PointNet to generate a global feature vector. The results are greatly improved due to the introduction of local attention. Although these point net expansion methods have achieved satisfactory results when processing traditional point cloud data, they cannot directly process microseismic event clouds. This is mainly because our goal is to focus on the aggregation of microseismic event indicators rather than analyzing the shape of the input point cloud. In other words, traditional methods mainly focus on determining the pattern of cloud shape or consider coordinates as the main features. In contrast, the method proposed in this invention only uses coordinates as an intermediate medium to aggregate information. In addition, the microseismic event cloud also contains timestamps that are not traditional point cloud components. According to the above reasons, training a filter for feature learning from given event coordinates and timestamps can solve the above problems. This principle is very similar to discrete convolution operation. The difference is that discrete convolution can only learn weights at specific locations, i.e., integer pixel locations in an image or volume, but the filter used to aggregate microseismic event sets should be able to learn parameters from continuous space and time space. In the rest of the invention, the traditional convolution operation is reviewed and a continuous convolution technique is proposed. In order to achieve robustness to poor in-situ computing environments, the FAC method is proposed. Then, the technique is combined with time encoding to build a complete architecture and provide microseismic event cloud data to predict rockburst risks.

[0061] Discrete 3D convolution operation:

[0062] Convolution is a mathematical operation, which allows to combine two sets of information. In the case of CNN, such an operation is applied to the input data to filter the contained information and generate a corresponding feature map. This filter is also called a kernel or feature detector. For example, when a CNN processes images, its size can be 3×3. The most common input data for CNN are images in computer vision. In this case, the CNN used is called a 2D CNN. To perform a convolution, a kernel is overlaid on the input image, performing an element-wise matrix multiplication. The results obtained for each receptive field (the area where the convolution occurs) are written down in a feature map. For 3D volumetric data, which is closer to the real world than 2D planar data, the convolution operation should be extended to three dimensions. A 3D CNN maintains a convolution that is very similar to a 2D CNN. It only applies a 3D filter to the data, and the filter moves along 3 directions (x, y, z) to calculate the low-level feature representation. In Figure 2 , the output shape is a three-dimensional volume space, such as a cube.

[0063] It is clear that the convolution operation has the nice properties mentioned above; that is, it treats coordinates as a medium rather than the primary features commonly seen in point networks. However, discrete convolution kernels can only define weights on constant and integral coordinates. Figure 2 As shown, there is a 3×3×3 kernel, which means that the weights only exist in 27 coordinates, namely (0, 0, 0), (0, 0, 1)…(1, 1, 1). In contrast, high-dimensional point cloud data has floating number coordinates, and the positions of neighbors for each point are also different; therefore, constant and fixed-shape kernels cannot be directly applied to extract features from point cloud sets. Even if a given point cloud can be converted to a volumetric format, the presence of many sparse regions will reduce computational efficiency, and the conversion process may cause errors. Therefore, a "deformable" convolution kernel is needed to learn features from high-dimensional point clouds. The term "deformable" means that the filter can be arbitrarily scaled ( Figure 3 ). In the next section, we describe the proposed FAC method, which convolves high-dimensional point clouds with timestamps, and present the final prediction architecture.

[0064] Convolution on high-dimensional point sets for feature learning:

[0065] Convolution is mathematically defined as equation (1), where g(t) is considered as the convolution function used to convolve the input signal f(t). For the discrete version in CNN, the convolution function is discretized into a regular grid consisting of several independent pixels with integral coordinates as shown in equation (2). If the discrete version is extended to three-dimensional space, this equation can be reformulated in equation (3),

[0066]

[0067]

[0068]

[0069] Where F represents the feature query function that returns the features of a point, W is the weight function, and δ is the coordinate increment in the local area. It is worth noting that the weights in traditional CNN are stored as table types; however, in continuous three-dimensional space, it is impossible to enumerate the weights in the point cloud because the points in the point cloud do not lie on a fixed grid but can take any continuous weights. Instead of learning a weight table that defines all parameters discretely, a weight function should be trained to map the three-dimensional coordinates to the corresponding weights. We use a multi-layer perceptron (MLP) to approximate the weight function, as shown in equation (4),

[0070] (x i ,y i ,z i ) where are the coordinates of the i-th neighboring point, and MLP is a fully connected neural network with multiple hidden layers. represents the weight vector associated with this point, where d is the dimension of the feature. The weights of each layer of the MLP are shared between all neighboring points to maintain permutation invariance. In addition, this continuous convolution operation should be performed on a local area. Before entering the layer, the coordinates of each point should be converted from global values ​​to local values ​​( Figure 3 ). In detail, we first apply the KNN algorithm to obtain each local region while assuming that all points are taken as the centroid in turn. Then, for each local region, we subtract each point from the coordinates of the centroid. Finally, we scale normalize each region.

[0071] In order to process the microseismic event cloud, the timestamp of each event must be considered, but the naive continuous three-dimensional convolution format cannot model the time dimension. In this invention, the simplest method is used to encode the timestamp of each local area. Assume that the timestamp of a local point set with n points is [t0, t1, t2…, t n ], equation (5) is applied to obtain the normalized temporal encoding.

[0072] T n =scale(t i -t0) (5)

[0073] “scale” represents a normalization function that normalizes the values ​​of a set to [0,1]. By n With the coordinate vector (x i ,y i ,z i ) nThe four-dimensional input data is concatenated and fed into the weight function. The weights obtained have the same dimension as the feature map. Next, these weights are multiplied by the corresponding features and the results are averaged and aggregated to become part of the new feature map. The output dimension of the new feature map can be obtained by stacking several consecutive filters.

[0074] However, the implementation of the continuous point cloud convolution operation is very inefficient and memory intensive, which prevents its use on poorly configured computers at construction sites, such as embedded systems and mobile devices. This paper proposes a new implementation of the operation, called FAC. This reconstruction can exploit existing parallel computing techniques and is easily implemented by mainstream deep learning frameworks; it can be viewed as a learnable approximation of the high-dimensional continuous convolution operation.

[0075] Let us imagine a situation where the input feature channel is c_in and the output channel is c_out. Let us take a point or volume pixel as an example and calculate the computational cost. For the case of discrete convolution, the number of parameters is simple. c_in×c_out For the case of continuous convolution, if we use MLP to fit the weight function, the cost depends on the number of hidden layers used in the MLP. For example, if the MLP has a hidden layer with 16 neurons, then the number of parameters should be c_in×c_out×16 because the input and output feature channels are fixed. In summary, continuous convolution should have c_in×c_out×∑ i num neurons (i), where i represents the i-th hidden layer, returns the number of neurons in the i-th hidden layer. Therefore, the number of parameters increases by dozens of times, which greatly limits the application of this operation.

[0076] Here, we introduce an assumption: the weight vectors generated by different filters in a specific layer should be correlated because they are in the same common space (output feature space), rather than independent of each other. Therefore, a mapping function can be constructed to project one feature subspace to another feature subspace. This assumption is used to simplify the above complex operations and make them sufficiently "lightweight". We call this mapping function a weight-to-weight network (w2ws network). The entire process of the proposed FAC method is as follows Figure 4As shown, where n represents the number of input channels and o represents the number of output channels. For the microseismic event cloud, each local area shares the same weight function that can be learned using MLP; therefore, we take a point as an example to describe the whole process. First, an MLP with a hidden layer containing 16 neurons is trained to fit the first continuous filter (Formula (6)), which contributes to the first dimension of the output channel. It can be seen in equation (6) that there is no activation function or batch normalization layer. Then, the first weight vector is input to the w2ws network, and the output contains the o-1 vector, which is the other o-1 features in the output feature space. The w2ws network in the present invention is an MLP with two hidden layers, both of which have 8 neurons. In addition, the w2ws network has activation and batch normalization functions. Here, the ownership vectors around this point are obtained, and they are multiplied by the previous feature vectors and aggregated to generate each dimension of the output vector. We assume that the w2ws transformation is performed by a function f and contains two points in this local point set: p1 and p2. In addition, the two weight vectors generated by equation (6) are W 11 and W 12 The FAC process can be described by equation (7), where O1 and O2 include all output vectors except the vector produced by the first MLP. This is why W in O1 21 The first subscript of starts with 2. In actual implementation, it is recommended to calculate the second half first to get the first eigenvector. In addition, the training w2ws network projects the first eigenvector onto the other o-1 eigenvectors instead of projecting the first weight vector. In this way, the efficiency can be further improved.

[0077] V=W2(W1(x,y,z,t)+b1)+b2 (6)

[0078]

[0079] Where O1=(W 21 ,W 31 ,…,W o1 ) 1n

[0080] In this method, the continuous weight generation process is not stored in memory, but divided into two parts: the first step is to train a common filter, and the second part is to project this filter onto other o-1 feature vectors based on the first vector. This is similar to dividing a continuous process into smaller continuous and discrete processes, which greatly reduces the consumption of computing resources.

[0081] In summary, this FAC module can be directly applied to extract features from microseismic event clouds. Inspired by traditional CNN, we also stack several such layers to build a final architecture to perform rockburst prediction. Figure 5 is the final framework, which consists of three FAC layers and a classification layer. The input data is divided into two parts: one group contains 11 event indicators, called feature groups, and the other is a set of coordinates, called coordinate groups. The coordinate groups are input into the KNN module to generate local region sets. In the present invention, we set the number of neighbors in the KNN module to 15 through experiments. These sets are then encoded with timestamps by equations, and these local regions are reused on all convolutional layers to generate corresponding weights, which are multiplied by the feature branches. The aggregation layer calculates the average of all point feature vectors to obtain a 256-dimensional global feature vector. The classification layer is an MLP with softmax activation.

[0082] Experiment and evaluation:

[0083] To evaluate the proposed method, we conducted multiple experiments on a dataset containing all rockburst events in the Han-Wei River Diversion Project. The microseismic events three days before each rockburst were selected as the input data belonging to the positive class. In contrast, the dates without rockbursts were randomly sampled, and the microseismic events three days before that were selected as the negative input data. It is worth noting that the rockburst event on the last day should not be included because it is our prediction target; however, the rockburst events before the last day should be included. Here, we constructed two competing models, including a time series model and a traditional threshold model, to demonstrate the effectiveness of the proposed model. In addition, we built another FAC-based model that does not encode timestamps (i.e., the data input to the FAC module is in the format of (x, y, z) without t to determine the importance of timestamps. All the comparison results are shown in Table 1. The traditional threshold model contains two vectors: the first is the sum of the above indicators, and the second is the average of these indicators. The threshold can be manually determined by experienced experts or trained based on previous data. Here, we use both methods: the TT-1 model has a manual threshold, while the TT-2 model adopts the learned classification boundary. It is worth noting that neither of these two studies considers any spatiotemporal correlation between events. The time series model is just a simple long short-term memory (LSTM) model that encodes events sequentially according to their timestamps. This model considers the temporal information of events, but ignores the spatial distribution of events. In addition, AF-Space in Table 1 represents the FAC model without time encoding, and AF-ST represents the entire model proposed above.

[0084] As can be seen from the results, the traditional thresholding method completely collapses, indicating that it is impossible to determine the rockburst risk with only a few indicators. In a real project, microseismic experts usually make decisions based on many factors. In addition, when LSTM is introduced to model the temporal correlation, the performance is greatly improved. The reason for this finding is that LSTM has better representation power and is able to encode information along the time dimension. However, the results generated by LSTM are still far behind those given by the FAC-based model. This shows that even if the time series model can capture some features related to the time dimension, it cannot extract reasonable features to identify event group patterns. In addition, AF-Space receives input with three-dimensional coordinates, but does not include timestamps, producing an accuracy of 87.3%. This performance is close to the 90.1% accuracy provided by the full model with time encoding, namely the AF-ST model. The point seems to be that spatial distribution is the most dominant factor, or at least more important than temporal correlation. However, in tunnel engineering, the spatial event distribution implicitly contains temporal correlation to a certain extent. It is easy to imagine that most of the new microseismic events are triggered around the tunnel face, because the rock mass around these areas is mostly disturbed during excavation processes such as blasting or tunnel construction. Therefore, the spatial distribution along the tunneling direction roughly reflects the order of these triggering events, such as Figure 6 As shown in Figure 2. Different colors correspond to different dates, and the redder the color, the closer it is to the future. It is obvious that large-scale temporal correlations can be identified along the excavation direction. Even if the timestamp information is lost, the performance is rarely degraded due to the large-scale temporal information contained in the spatial correlation. Therefore, in real-world applications, considering only the spatial coordinates of the event may be a trade-off between accuracy and computational complexity.

[0085] To deeply examine the results of the proposed method, we calculated three metrics, including precision, recall, and F1 score, and reported them in Table 2. As can be seen from the table, the performance of the model is stable. The three metrics of the positive class are higher than those of the negative class.

[0086] Table 1 Comparison of the results of the proposed method with existing methods

[0087]

[0088] Table 2 Precision, recall and F1 score of the results

[0089]

[0090] In the rest of this section, we will select some specific examples to illustrate the effectiveness of the proposed method when applied to the real world. Figure 7The example of TT-1 and TT-2 predicting that rock burst will not occur is given, but the model based on the FAC proposed by the present invention predicts that rock burst will occur, which is consistent with the actual record. Because even if the entire microseismic event is relatively small, the location of the damage is concentrated around the tunnel face, and the energy released in this process is still relatively small. The concentration of damage with small energy release indicates that these damages do not release stress, but are more likely to cause greater stress concentration, and the accumulation of high stress is conducive to the occurrence of rock burst.

[0091] Another example is Figure 8 This event set was predicted by the time series model as a rockburst. However, the report showed that no rockburst occurred on that day, and the results of the FAC-based model were consistent with the facts.

[0092] In addition, our algorithm can also lead to some counterintuitive conclusions. For example, Fig. 9 and Fig.10 Two groups of events are shown. The number of events in these two groups is almost equal. The first set of events is more concentrated and has a higher total energy level. Intuitively, the first group is more likely to cause a rockburst, while the second group seems safe. However, the fact is that the first group remained stable, while the second group had a rockburst on the same day. It is difficult for humans to make correct judgments, but our model can give accurate results. Therefore, people can learn new rules and knowledge by comparing the results of manual predictions, model predictions, and real cases.

[0093] in conclusion:

[0094] In order to predict rockbursts through microseismic monitoring, the present invention proposes a deep learning algorithm based on a point cloud model to replace the previous threshold-based and time series-based methods to provide more effective and accurate early rockburst warnings. Microseismic events can be considered as three-dimensional point clouds with high-dimensional features. These features are not summarized and compared with pre-set thresholds, nor are they averaged and analyzed through time series models to determine rockburst risks; the present invention treats each event independently as a high-dimensional point cloud set. In the present invention, the monitoring system generates 11 different raw features for each event and inputs them into a machine learning model to learn the characteristics of rockbursts and prevent information loss caused by special feature extraction.

[0095] Microseismic event clouds contain not only spatial locations but also temporal information that traditional point cloud data do not have. Traditional point cloud models cannot be directly applied to the prediction of rockbursts because this task requires incorporating the spatiotemporal data of microseismic events and their associated damage information into a machine learning model. Therefore, we proposed the AF-ST model to process complex microseismic monitoring data to comprehensively consider spatiotemporal characteristics. In addition, the present invention proposes a new FAC technique to effectively extract features from high-dimensional point clouds instead of using many independent MLPs to fit different filters. Finally, a new FAC-based point cloud model AF-ST model is proposed to predict the occurrence of rockfall eruptions. To the best of our knowledge, this invention is the first attempt to use a point cloud model to predict rockbursts.

[0096] To demonstrate the effectiveness of the proposed model for early rockburst warning, we also implemented two competing models: a time series model and a traditional threshold model. In addition, the present invention also established another FAC-based model that does not encode timestamps to study the importance of timestamp effect on rockburst warning. The results show that the model has an early rockburst warning accuracy of up to 90.1% on the Han-Wei River Diversion Project, which is significantly better than the traditional time series model and threshold model. In addition, this result also shows that spatial distribution is more important than temporal correlation when predicting rockbursts.

[0097] In addition, some specific rockburst predictions generated by the model were analyzed to validate the performance of the model. The results show that the model performs well in providing early warnings of rockbursts, including immediate rockbursts, delayed rockbursts, and structural rockbursts; even though it is difficult for human experts to judge whether a rockburst will occur based on microseismic data. Based on these results, our model is applicable to multiple types of rockbursts and is able to capture the dynamic changes and characteristics of the rock mass. The performance of the model can be further improved by modeling the temporal and spatial correlations separately and integrating them together.

[0098] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An intelligent rockburst early warning method based on microseismic point cloud feature learning, It is characterized in that The steps include: S1. Establish a microseismic monitoring system; S2, randomly select a data set; S3, the rockburst prediction problem is formulated as a recognition task on high-dimensional point cloud patterns, taking high-dimensional point clouds as input and learning features from high-dimensional point clouds to identify different concentration patterns of event points; In step S3, the microseismic event is a three-dimensional point cloud with high-dimensional properties. The microseismic event is represented by spheres of different colors. The microseismic event is represented as a three-dimensional point set, wherein each point has its own three-dimensional coordinates (x, y, z) and feature channels. The color and size of each point represent the magnitude and energy, respectively. Considering 11 different features of each event point, namely, seismic momentum, event energy, source radius, AspRadius, static stress, apparent stress, dynamic stress, maximum displacement, peak velocity, peak acceleration and local amplitude, each event is independently regarded as a high-dimensional point cloud set. The spatiotemporal distribution of the high-dimensional point cloud is related to the rock burst risk. The rock burst prediction problem is reformulated as a pattern recognition task on a high-dimensional point cloud set. The method uses a high-dimensional point cloud as input, rather than a highly integrated descriptor or time series, identifies different patterns in the event point set, and learns features from the high-dimensional point cloud. In step S3, the steps of point set deep learning are as follows: Establish the architecture of PointNet: The input of PointNet is the original point cloud P N×D , where N is the number of points in the cloud and D is the dimension, representing the (x, y, z) of each point. Two T-Nets are used to model the unordered features with learnable weights. In the last step, a 1024-dimensional feature vector is generated using the maximum pooling symmetric function to aggregate global information. The feature vector represents the feature descriptor of the input for recognition tasks. Further improvements to PointNet are achieved by modeling local region structures, which are used to capture hierarchical local structures in a manner similar to grid convolutions. In PointNet++, first in P N×3 The farthest point sampling algorithm is used to select the centroid, and the ball query is used to select the nearest neighbor of each centroid; then, the point network is applied to all local areas to produce a vector describing the characteristics of the area; this process is repeated in a hierarchical form, thereby reducing the resolution of the points as the process deepens. In the last layer of the hierarchy, the entire point feature is generated through PointNet to generate a global feature vector; S4. Perform convolution on the high-dimensional point set for feature learning and predict rockburst.

2. According to claim 1, the rockburst intelligent early warning method based on microseismic point cloud feature learning, It is characterized in that In step S1, the configuration of the microseismic monitoring system is as follows: the microseismic monitoring system is set as a workstation inside the tunnel and a workstation outside the tunnel. The workstation inside the tunnel consists of an acceleration sensor and an integrated box. The acceleration sensor is used to detect microseismic events in the rock. The acceleration sensor is installed in the borehole and fixed with an anchor. The integrated box includes a Paladin digital signal acquisition system and a signal processing system, which are responsible for collecting and recording microseismic data, and wirelessly transmitting the microseismic information to the workstation outside the tunnel. The workstation outside the tunnel consists of a wireless signal real-time receiving system and an artificial intelligence signal processing system. The wireless signal real-time receiving system receives the microseismic data wirelessly sent by the workstation inside the tunnel in real time. The artificial intelligence signal processing system consists of a microseismic event recognition and classification module, a time-pickup module, a transfer learning module, a source positioning module, and an early warning module based on artificial intelligence. These modules process the observed microseismic information and provide rock burst warnings in a timely manner. The internal workstation adopts a mobile layout. The tunnel face is constantly moving forward. The position of the microseismic monitoring equipment is adjusted according to the position of the face to ensure that the acceleration sensor can monitor the damage in the surrounding rock in real time. The microseismic monitoring system includes 6 acceleration sensors, which are installed on the side walls on both sides of the tunnel, with 3 sensors on each side. The distance between two adjacent sensors on each side wall is 50m, and they are fixed at the waist and shoulders of the two side walls of the tunnel.

3. According to claim 1, the rockburst intelligent early warning method based on microseismic point cloud feature learning, It is characterized in that In step S2, each day is selected as a sample, including positive samples and negative samples. If a rockburst occurs on a certain day, the sample is determined as a positive label, and if a rockburst does not occur on a certain day, the sample is determined as a negative label. Each sample also contains dozens of microseismic events. The data of the first 5 days of the prediction are input into the neural network as a sample set. The distribution pattern of the point cloud is made to correspond to the risk of rockburst through the timestamp information of each microseismic event. Finally, all the features are connected and input into the model together.

4. According to claim 1, the rockburst intelligent early warning method based on microseismic point cloud feature learning, It is characterized in that In step S4, the convolution step on the high-dimensional point set for feature learning is as follows: According to the mathematical definition of convolution, for the discrete version in CNN, the convolution function is discretized into regular grids, which are composed of several independent pixels with integral coordinates. The discrete version is extended to three-dimensional space, and a multi-layer perceptron MLP is used to approximate the weight function, as shown in equation (4) (x i ,y i ,z i ) is the coordinate of the i-th neighboring point, and the multilayer perceptron MLP is a fully connected neural network with multiple hidden layers. Represents the weight vector associated with this point, where d is the dimension of the feature. First, the KNN algorithm is applied to obtain each local area, assuming that all points are taken as the centroid in turn. Then, for each local area, each point is subtracted from the coordinates of the centroid. Finally, each area is scaled normalized.

5. According to claim 4, the rockburst intelligent early warning method based on microseismic point cloud feature learning, It is characterized in that In step S4, it is assumed that the timestamp of a local point set with n points is t 0 ,t 1 ,t 2 …,t n , equation (5) is applied to obtain the normalized temporal encoding, T n =scale(t i -t 0 ) (5) Among them, "scale" represents the normalization function that normalizes the values ​​of the set to [0,1], by n With the coordinate vector (x i ,y i ,z i ) n The four-dimensional input data is concatenated and fed into the weight function. The weights obtained have the same dimension as the feature map. These weights are multiplied by the corresponding features and the results are averaged and aggregated to become part of the new feature map. The output dimension of the new feature map is obtained by superimposing several continuous filters.

6. The rockburst intelligent early warning method based on microseismic point cloud feature learning according to claim 5, It is characterized in that A method of operation is proposed: fast approximation convolution FAC. Fast approximation convolution FAC is regarded as a learnable approximation of high-dimensional continuous convolution operation. The input feature channel is c_in and the output channel is c_out. The weight vectors generated by different filters in a specific layer are related. A mapping function is constructed to project one feature subspace to another feature subspace. The entire process of the fast approximation convolution FAC is as follows: where n represents the number of input channels and o represents the number of output channels. For the microseismic event cloud, each local area shares the same weight function learned using a multilayer perceptron MLP; first, a multilayer perceptron MLP with a hidden layer of 16 neurons is trained to fit the first continuous filter formula (6), which contributes to the first dimension of the output channel. No activation function or batch normalization layer is seen in equation (6). Then, the first weight vector is input to the w2ws network, and the output contains the o-1 vector, which is the other o-1 features in the output feature space. The w2ws network is a multilayer perceptron MLP with two hidden layers, both of which have 8 neurons. In addition, the w2ws network has activation and batch normalization functions. The ownership vectors around this point are obtained and multiplied by the previous feature vectors and aggregated to generate each dimension of the output vector. It is assumed that the w2ws transformation is performed by a function f and contains two points in this local point set: p 1 and p 2 , In addition, the two weight vectors generated by equation (6) are W 11 and W 12 , the fast approximation convolution FAC process is described by equation (7), where O 1 and O 2 Include all output vectors except the one produced by the first multilayer perceptron MLP, and train the w2ws network to project the first feature vector onto the other o-1 feature vectors instead of projecting the first weight vector. V=W 2 (W 1 (x,y,z,t)+b 1 )+b 2 (6) Among them, O 1 = (W 21 ,W 31 ,…,W o1 )p 1n Instead of being stored in memory, the continuous weight generation process is divided into two parts: the first step is to train a common filter, and the second part is to project this filter onto other o-1 feature vectors based on the first vector.

7. The rockburst intelligent early warning method based on microseismic point cloud feature learning according to claim 6, It is characterized in that In step S4, the fast approximation convolution FAC is directly applied to extract features from the microseismic event cloud. Several such layers are stacked to build a final architecture to perform rockburst prediction, which consists of three fast approximation convolution FAC layers and a classification layer. The input data is divided into two parts: one group contains 11 event indicators, called feature groups, and the other group is a coordinate set, called coordinate groups. The coordinate groups are input into the KNN module to generate local area sets. The number of neighbors in the KNN module is set to 15. Then, these sets are encoded with timestamps by equations. These local areas are reused on all convolutional layers to generate corresponding weights, which are multiplied by feature branches. The aggregation layer calculates the average of all point feature vectors to obtain a 256-dimensional global feature vector. The classification layer is a multi-layer perceptron MLP with softmax function activation.