CN117609874A - Rock fault friction microseismic detection method and system based on integrated deep learning - Google Patents

Rock fault friction microseismic detection method and system based on integrated deep learning Download PDF

Info

Publication number
CN117609874A
CN117609874A CN202311484390.XA CN202311484390A CN117609874A CN 117609874 A CN117609874 A CN 117609874A CN 202311484390 A CN202311484390 A CN 202311484390A CN 117609874 A CN117609874 A CN 117609874A
Authority
CN
China
Prior art keywords
deep learning
integrated
phase
integrated deep
acoustic emission
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202311484390.XA
Other languages
Chinese (zh)
Other versions
CN117609874B (en
Inventor
李德康
戴仕贵
谢凡
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Sichuan Earthquake Administration
INSTITUTE OF GEOPHYSICS CHINA EARTHQUAKE ADMINISTRATION
Original Assignee
Sichuan Earthquake Administration
INSTITUTE OF GEOPHYSICS CHINA EARTHQUAKE ADMINISTRATION
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Sichuan Earthquake Administration, INSTITUTE OF GEOPHYSICS CHINA EARTHQUAKE ADMINISTRATION filed Critical Sichuan Earthquake Administration
Priority to CN202311484390.XA priority Critical patent/CN117609874B/en
Publication of CN117609874A publication Critical patent/CN117609874A/en
Application granted granted Critical
Publication of CN117609874B publication Critical patent/CN117609874B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28—Processing seismic data, e.g. for interpretation or for event detection
    • G01V1/282—Application of seismic models, synthetic seismograms
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28—Processing seismic data, e.g. for interpretation or for event detection
    • G01V1/288—Event detection in seismic signals, e.g. microseismics
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00—Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28—Processing seismic data, e.g. for interpretation or for event detection
    • G01V1/32—Transforming one recording into another or one representation into another
    • G01V1/325—Transforming one representation into another
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V2210/00—Details of seismic processing or analysis
    • G01V2210/40—Transforming data representation
    • G01V2210/48—Other transforms
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Remote Sensing (AREA)
  • Software Systems (AREA)
  • Environmental & Geological Engineering (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Acoustics & Sound (AREA)
  • Evolutionary Biology (AREA)
  • Geology (AREA)
  • General Life Sciences & Earth Sciences (AREA)
  • Geophysics (AREA)
  • Business, Economics & Management (AREA)
  • Emergency Management (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)

Abstract

The invention discloses a rock fault friction microseismic detection method and system based on integrated deep learning, wherein the method comprises the steps of constructing an integrated deep learning model, and the integrated deep learning model comprises the following steps: the first backbone characteristic extraction network is used for extracting local P-phase characteristics in acoustic emission waveform data; a second backbone feature extraction network for extracting complete waveform features having global context information in the acoustic emission waveform data; the integrated learning layer is used for carrying out weighted integration on output results of the first backbone characteristic extraction network and the second backbone characteristic extraction network to construct a P-phase integrated predictor; and acquiring continuous waveform data monitored in real time in the fault friction process, obtaining a P-phase probability sequence by using an integrated deep learning model, and detecting and picking up acoustic emission phases according to the P-phase probability sequence. The invention can automatically and accurately detect and pick up the microseism event by utilizing an integrated deep learning model based on a limited training data set.

Description

Rock fault friction microseismic detection method and system based on integrated deep learning
Technical Field
The invention relates to the technical field of seismic data analysis, in particular to a rock fault friction microseism detection method and system based on integrated deep learning.
Background
Disturbance of underground engineering such as energy exploration (deep mineral and geothermal), environmental protection (nuclear waste sealing), geological disaster early warning (rock slope) and the like can cause change of local stress fields around a fault plane, and fault activation is induced under certain conditions, so that fault sliding rock burst, mineral earthquake, earthquake and other related dynamic disasters are further caused. Thus, revealing the characteristics and mechanisms of rock fault friction is of great importance for predicting and preventing fault activation and related disasters.
Acoustic emission or microseismic events generated by rock fault friction are the representation of fault fracture or dislocation, and are the connection between a crack evolution rule on a microscopic scale and deformation characteristics on a macroscopic scale; therefore, it is critical to detect and pick up its dominant arrival phase (P-phase) from the large volume of waveform data of a multi-channel continuous acoustic emission or microseismic recording system.
Compared with the traditional phase picking algorithm, such as a long-short time window contrast method (STA/LTA), a template matching technology (cross correlation algorithm) and the like, the data processing method based on deep learning is widely applied to analysis and interpretation of seismic data due to high precision and high efficiency.
While deep learning based network models have demonstrated dramatic accuracy and efficiency in natural seismic detection and seismic facies pickup, they rely on the large amount of seismic annotation data accumulated in community seismic catalogs and the powerful feature extraction capabilities of deep network architecture. However, acoustic emission or microseismic waveform data is difficult to crowd-sourced or professionally labeled, lacks sufficient training data and corresponding tags, and therefore, in the case of artificially labeled acoustic emission or microseismic data being limited, detecting and picking up P-phases from multichannel continuous waveform data in a complete manner remains a challenging task.
At present, the research of detecting acoustic emission events by using a deep learning model is less, the existing model has a single type and relatively shallow network structure, and the problems of low precision or poor generalization performance of the model exist, so that a related method is not mature; the deep learning model specially designed for acoustic emission or microseismic event detection in the fault friction process is absent, so that the space-time evolution precision of acoustic emission or microseismic activity in the fault friction instability process is low.
Therefore, it is necessary to provide an integrated deep learning method for modeling based on limited manual annotation data, so as to avoid the problems of low precision and poor generalization caused by the over-fitting phenomenon of a single base model.
Disclosure of Invention
In view of the above, the invention provides a rock fault friction microseismic detection method and system based on integrated deep learning, which aim to improve the detection precision and generalization of a deep learning model by using an integrated deep learning method based on limited manual annotation data.
In order to achieve the above purpose, the present invention adopts the following technical scheme:
a rock fault friction microseismic detection method based on integrated deep learning comprises the following steps:
constructing an integrated deep learning model, the integrated deep learning model comprising:
the first backbone feature extraction network is used for extracting local P-phase features in the waveform data; the first backbone feature extraction network comprises a first downsampling section, a local phase feature extraction layer and a first upsampling section in sequence,
a second backbone feature extraction network for extracting complete waveform features with global context information in the waveform data; the second backbone feature extraction network sequentially comprises a second downsampling section, a global waveform feature extraction layer, a second upsampling section and
the integrated learning layer is used for constructing a P-phase integrated predictor by carrying out weighted integration on output results of the first backbone feature extraction network and the second backbone feature extraction network, and the P-phase integrated predictor is used for obtaining a P-phase probability sequence;
and acquiring continuous waveform data monitored in real time in the fault friction process, obtaining a P-phase probability sequence by using the integrated deep learning model, and detecting and picking up acoustic emission phases according to the obtained P-phase probability sequence.
Preferably, the first downsampling section and the second downsampling section have the same structure and respectively comprise a plurality of layers of convolution networks, and each layer of convolution network comprises a one-dimensional convolution layer and a maximum pooling layer.
The one-dimensional convolution layer learns a high-order characteristic mode of waveform data by sliding convolution kernels at different positions of input data, the maximum pooling layer reduces the dimension of the data, reduces the computational complexity and enhances the extraction of useful information by a network.
In addition, in the convolution layer, the number of convolution kernels is gradually increased, and the size of the convolution kernels is gradually reduced, so that the perceptibility of the neural network to the multi-scale information is also improved.
The encoder integrated with the deep learning model learns the information features of the input waveform data. The one-dimensional convolution layer and the maximum pooling layer are used for encoder construction, mainly to extract multi-scale useful information from the raw data and compress into a small number of neural units.
Preferably, the convolutional network has 7 layers.
The convolution network is set to be 7 layers, so that multi-scale feature representation can be learned, and the calculation cost is low.
Typically, convolution structures with less than 3-4 layers are often referred to as shallow networks, which are typically only capable of learning low-order features of the input waveform data for a limited data set; deep networks in turn increase the risk of feature learning over-fitting and high computational costs.
The invention considers learning multi-scale characteristic representation, has lower calculation cost, and sets the convolution network as 7 layers.
Preferably, the first upsampling section and the second upsampling section have the same structure and each include multiple one-dimensional convolutional layers.
Preferably, the one-dimensional convolution layer in the first/second upsampling section has 8 layers.
The up-sampling structure is arranged approximately symmetrically to the down-sampling structure, also in 7 layers,
the downsampling structure can effectively extract multi-scale useful information through the convolution kernel and the pooling layer, and the upsampling structure can better restore local and global information extracted from the waveform sequence by utilizing the convolution kernel size.
The application sets an 8 th one-dimensional convolution layer, which is used for changing a plurality of convolution kernels in the last layer into 1 convolution kernel, namely, the application aims to integrate the characteristics of a plurality of channels into one-dimensional sequence information and restore one-dimensional sequence waveforms.
Preferably, the local phase feature extraction layer sequentially comprises
A two-way long and short term memory layer for converting the downsampled features into a higher order representation having a time dependence;
a transducer-global attention mechanism for directing neural network attention to a portion associated with the acoustic emission waveform; and
the transducer-local attention mechanism directs neural network attention to local features associated with waveform phases by setting a large weight to a narrow window around the P-phase first arrival.
Preferably, the transducer-global attention mechanism is provided with a one-way long-short-term memory layer and a one-way short-term memory layer at the front and back for integrating position information to enable the following encoder states to have position perception.
Preferably, the global waveform feature extraction layer includes a plurality of bidirectional long-short-term memory layers for converting downsampled features into higher-order timing features having global context information.
Preferably, the two-way long-short-term memory layer is set to be 4 layers so as to ensure that the model integration has a better pick-up effect.
Preferably, a Softmax function is connected after the first upsampling segment and is used for mapping the local P-phase feature extracted by the first neural network into a first probability vector point-to-point;
and the second upsampling section is connected with a Softmax function for mapping the global high-order time sequence features extracted by the second neural network into a second probability vector point to point.
Preferably, a training data set is constructed, the integrated deep learning model is trained, and the training data set comprises:
a limited training data set is constructed, wherein the training data set comprises complete front earthquake, main earthquake and afterearthquake, the magnitude distribution follows the Gordon's Barbell-Rickett law, and the statistical characteristic distribution of the training set is consistent with the characteristic distribution of the rest earthquake period.
In another aspect, the present application discloses a rock fault friction microseismic detection system based on integrated deep learning, the system comprising,
the acoustic emission monitoring device is used for monitoring and recording acoustic emission waveform data in the fault friction process;
the microseism event detection device is internally provided with an integrated deep learning model constructed in the rock fault friction microseism detection method based on integrated deep learning, the integrated deep learning model is used for processing real continuous waveform data recorded in real time in a fault friction process, a P-phase probability sequence is obtained by utilizing the integrated deep learning model, and acoustic emission phases are detected and picked up according to the obtained P-phase probability sequence.
Compared with the prior art, the invention discloses a rock fault friction microseismic detection method and system based on integrated deep learning, which are characterized in that an integrated deep learning model is built, a parallel deep neural network is arranged, global and local diversified features are extracted from a limited training data set, and an integrated learning layer is used for building a P-phase integrated predictor for acoustic emission phase pickup;
the integrated deep learning is a synergistic effect of the weighted integration strategy and the deep learning model, and the integrated deep learning model is obtained by fusing the heterogeneous base models, so that the integrated deep learning model has the capability of coping with key challenges (such as small sample size, unbalanced class distribution, heterogeneous data and the like), and can effectively solve the problem of over-fitting of the base models caused by insufficient training data and coping with heterogeneous continuous waveform data recorded in a plurality of fault friction periods.
Meanwhile, the detection method provided by the invention can effectively reduce the requirement for manual annotation data, and based on limited data, the advantages of high precision, high accuracy and generalization of the integrated deep learning model can be fully exerted, so that the acoustic emission or microseism event can be automatically detected and picked up with high precision.
In addition, the method can be suitable for detecting acoustic emission or microseismic events and picking up phases in continuous waveform data in the frequency range of 1000 Hz-3 MHz, and has wide application potential for microseismic monitoring of internal fault friction sliding of rock slopes, underground reservoirs and the like.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims thereof as well as the appended drawings.
The technical scheme of the invention is further described in detail through the drawings and the embodiments.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings that are required to be used in the embodiments or the description of the prior art will be briefly described below, and it is obvious that the drawings in the following description are only embodiments of the present invention, and that other drawings can be obtained according to the provided drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic diagram of an integrated deep learning model in a second embodiment of the present invention;
FIG. 2 is a graph showing the performance comparison of two-way long-short-term memories of different layers according to the present invention, (a) shows 2-layer two-way long-short-term memories, (b) shows 3-layer two-way long-short-term memories, (c) shows 4-layer two-way long-short-term memories, and (d) shows 5-layer two-way long-short-term memories;
FIG. 3 is a graph showing the performance of two-way long-short term memories of different layers under different signal-to-noise ratios according to the present invention, wherein (a) represents 2 layers of two-way long-short term memories, (b) represents 3 layers of two-way long-short term memories, (c) represents 4 layers of two-way long-short term memories, and (d) represents 5 layers of two-way long-short term memories;
FIG. 4 is a performance display of the overall model of the present invention; (a) Representing a performance map at different time residuals, (b) representing a performance map at different signal-to-noise ratios;
FIG. 5 is a flow chart of training and detection of the integrated deep learning model of the present invention;
FIG. 6 is a graph of test results of four continuous waveform data according to the present invention, wherein (a) - (d) are serial numbers of continuous waveforms;
FIG. 7 is an acoustic emission event constructed from 3 acoustic emission waveform correlations in accordance with the present invention
FIG. 8 is an acoustic emission event constructed from 6 acoustic emission waveform correlations in accordance with the present invention;
FIG. 9 is an acoustic emission event constructed from 8 acoustic emission waveform associations of the present invention;
FIG. 10 is an acoustic emission event constructed from 12 acoustic emission waveform correlations in accordance with the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The continuous waveform data during the process of the stick-slip instability of the large-scale rock fault is monitored by using 12 detectors with a sampling rate of 3MHz, 13 stick-slip events (namely earthquake events) are recorded in total, each stick-slip event and the front 70s and the back 70s are defined as a stick-slip period (141 s) to be analyzed, wherein the continuous waveform data of each second is stored as a separate file containing 300 ten thousand data points. Therefore, there are significant challenges in detecting and picking up acoustic emission phases in such long-period continuous complex multi-channel waveform data.
Although the deep learning model is the current main research direction, the data in the field is difficult to crowd-source or manually label, so that the labeled acoustic emission data is limited, and the development and application of the deep learning model in the field are bottleneck.
In this regard, the embodiment of the invention discloses a rock fault friction microseismic detection method based on integrated deep learning, which is realized by the following scheme:
example 1
In this embodiment, the rock fault friction microseismic detection method based on integrated deep learning includes the following steps:
constructing an integrated deep learning model, wherein the integrated deep learning model is provided with a new network architecture and comprises two independent backbone feature extraction networks, namely:
the first backbone feature extraction network is used for extracting local P-phase features in the waveform data; sequentially comprises a first downsampling section, a local phase characteristic extraction layer and a first upsampling section,
the second backbone characteristic extraction network is used for extracting complete waveform characteristics in the waveform data; the method comprises a first downsampling section, a global waveform characteristic extraction layer and a second upsampling section in sequence;
the integrated learning layer is used for constructing a P-phase integrated predictor by carrying out weighted integration on output results of the first backbone feature extraction network and the second backbone feature extraction network, and the P-phase integrated predictor is used for obtaining a P-phase probability sequence;
and
Continuous waveform data monitored in real time in the fault friction process are obtained, the real-time monitoring data are preprocessed, the preprocessed data are input into an integrated deep learning model, a P-phase probability sequence is obtained, and acoustic emission phases are detected and picked up according to the obtained P-phase probability sequence.
In this embodiment, preprocessing of the recorded actual monitoring data includes: cutting real continuous waveform data of each second into a window of 6144 data points with certain overlap, removing the mean value of the data, and dividing the mean value by the standard deviation for data normalization; the data volume size of real-time monitoring data per second is therefore: 533*1*6144.
According to the invention, two independent backbone feature extraction networks run in parallel, waveform data information is comprehensively learned from two aspects of local phase features and full waveform high-order time sequence features based on limited training data, and the advantage complementation of the two backbone feature extraction networks is promoted through an integrated learning strategy, so that the model is prevented from being over-fitted, the accuracy and generalization of the model are improved, and the model can effectively detect and pick up multi-channel acoustic emission events.
The method and the device face limited acoustic emission or microseismic training data sets, fully exert the advantages of high precision and generalization of the integrated deep learning model, and can automatically and accurately detect and pick up acoustic emission or microseismic events.
The specific structure is shown in fig. 1, wherein the first downsampling section and the local P-phase extraction layer, the second downsampling section and the global waveform characteristic extraction layer which are connected in parallel form an encoder together, and the encoder is used for encoding received waveform data and extracting local P-phase characteristics and high-order time sequence characteristics of full waveforms in the waveform data;
the first upsampling section and the second upsampling section together with a common ensemble learning layer form a decoder for decoding the extracted features and for weighted integration of the two probability vectors, thereby constructing a P-phase ensemble predictor.
In the encoder, the first downsampling section and the second downsampling section have the same structure and respectively comprise a plurality of layers of convolution networks, and each layer of convolution network comprises a one-dimensional convolution layer and a maximum pooling layer.
The one-dimensional convolution layer and the max-pooling layer are used for encoder construction, mainly to extract multi-scale useful information from the raw data and compress into a small number of neurons.
Wherein the first/second downsampling stage takes into account the learning of the multi-scale feature representation and the low computational cost, the convolutional network layer is preferentially set to 7 layers to facilitate learning of higher-order feature representations of different levels, while having lower computational cost.
For the first neural network, the local phase feature extraction layer sequentially comprises
A two-way long and short term memory layer for converting the downsampled features into a higher order representation having a time dependence;
a transducer-global attention mechanism for directing neural network attention to a portion associated with an acoustic emission/microseismic waveform; and
a transducer-local attention mechanism for setting a large weight to a narrow window around the P-phase first arrival to direct neural network attention to a local feature associated with the waveform phase.
In one embodiment, the transducer-global attention mechanism is provided with one-way long-short-term memory layers before and after the transducer-global attention mechanism, and the one-way long-short-term memory layers are used for integrating position information so that the following encoder states have position awareness.
The one-way long-short-term memory layer is used for integrating the position information of each state in the features extracted by the two-way long-short-term memory;
the one-way long-short-term memory layer arranged at the back is used for integrating state position information related to acoustic emission/microseismic waveforms guided by a global attention mechanism;
for the second neural network, after the downsampling section of the same structure, a global waveform feature extraction layer is set, in this embodiment, the global waveform feature extraction layer is formed by multiple layers of two-way long-short-term memory layers, preferably, a deep long-short-term memory stack structure formed by 4 layers of two-way long-short-term memory modules is set, and is used for converting downsampling features into high-order time sequence features with global context information.
In this application, models containing 2 layers of two-way long-short-term memory, 3 layers of two-way long-short-term memory, 4 layers of two-way long-short-term memory and 5 layers of two-way long-short-term memory are compared, respectively, as shown in fig. 2 and 3.
As can be seen from FIG. 2, compared with other tests, the Bi-LSTM model with 4 layers of two-way long-short-term memory built by the invention has higher precision, recall rate and F1 fraction,
fig. 3 shows the model performance test under different snr conditions, and although the single base model has poor generalization in the high snr range, the model with 4 layers of bidirectional long-short term memory can be found to be relatively stable compared with the other layers of models.
Therefore, the Bi-directional long-short-term memory layer of the Bi-LSTM model is set to be 4 layers, so that the model integration is guaranteed to have a good pickup effect.
The deep long-short term memory stack structure is provided with more hidden nerve units (see fig. 1), and can more effectively capture long-term time dependence among full waveforms.
Further, the first upsampling section and the second upsampling section in the decoder have the same structure and respectively comprise a plurality of one-dimensional convolution layers.
Wherein the one-dimensional convolution layer in the first/second upsampling segment is preferably set to 8 layers.
In fact, the upsampling segments and downsampling segments are symmetrical structures in the present application to better understand the global and local information in the waveform sequence, i.e. the upsampling segments are also arranged in 7 layers,
in order to integrate the features acquired by the up-sampling section into one-dimensional sequence information so as to restore one-dimensional sequence waveforms, an 8 th one-dimensional convolution layer is arranged.
In order to further optimize the technical scheme, a Softmax function is connected after the first upsampling section and is used for mapping the local P-phase characteristics extracted by the first neural network into a first probability vector point to point;
meanwhile, a Softmax function is connected after the second upsampling segment and is used for mapping the global high-order time sequence features extracted by the second neural network into a second probability vector point to point.
The integrated deep learning model established by the method has a new network architecture, can learn global and local high-order characteristic representation in waveform data under the condition of limited data, and also has a high-efficiency and stable P-phase probability predictor, and can detect micro-fracture events recorded in the rock fault friction process in a complete mode.
The network is not a simple technical stack but creates a new, over-the-air effect of the individual components. For a deep network architecture, it is quite difficult to avoid the occurrence of overfitting with limited training data.
As shown in fig. 4, from the performance of the integrated deep learning model at different signal-to-noise ratios, it can be seen that the model in the present invention can address the risk of such overfitting (i.e., as the signal-to-noise ratio increases, the precision and recall increase) and has a smaller pick-up error (i.e., the mean absolute error MAE is 8.4 μs, see fig. 8).
In addition, the application of the network design and the weighted integration strategy in the invention is based on domain knowledge, and can process continuous waveform data in media and mechanical mechanisms which continuously change in a plurality of fault friction periods. Therefore, the design of the network architecture and the application of the integration strategy are not simple stacks, aiming at effectively solving the complex problems faced by us at present.
The integrated deep learning model is trained before use, specifically as shown in fig. 5, a backbone feature extraction network on the left side of a decoder and an encoder is combined to be used as a deep network model, named as a transducer model, and a backbone feature extraction network on the right side of the decoder and the encoder is combined to be used as another deep network model, named as a Bi-LSTM model; in the training process, the two base models are independently trained, and model training parameters in the two base models are iteratively updated and stored respectively.
In one embodiment, a limited training data set, a validation data set and a test data set are constructed; wherein,
a training data set for training the two backbone feature extraction networks,
the verification data set is used for weighing the relative contributions of different networks, carrying out normalization weight accurate calculation by utilizing the harmonic performance indexes (F1 scores) of the two deep neural networks, and carrying out weight integration on the output result;
and the test data set is used for testing the integrated deep learning model with the training and verification completed.
In this embodiment, the training data set, the verification data set and the test data set are respectively in a ratio of 4:1:1, and the different data sets all include acoustic emission event data with two types of labels, namely, the label of each noise data point is 0, and the label of the P-phase data point is 1;
wherein the training data set is derived from a complete seismic cycle comprising a pre-earthquake, a main earthquake and a aftershock, the magnitude distribution follows the Gordon's Barbell-Rickett law, and the statistical feature distribution of the training set is substantially similar to the feature distribution of the rest of the seismic cycle, which is crucial for successful migration of the integrated deep learning model to unknown data in the case of limited acoustic emission training data.
In one embodiment, the dataset consists of 1D acoustic emission waveform data, wherein the data point label of a triangular window of 100 data points centered on the P-phase is set to 1 and the labels of the remaining data points are 0; the entire dataset is structured as (data volume and type tag 3D). For the training data set, nt (1) channels are included, and each channel includes Ns (6144) sampling points, and Nk (10000) data are included, so that the data body of the training set is ns×nk, and the type label is also ns×nk. The ratio of the verification data set to the test data set is 1:1:4, so that the data volumes of the verification set and the test set are Nt Nk/4. The verification set is from other earthquake periods and has data isomerism, which is helpful for more accurately and fairly evaluating the relative contribution of two backbone feature extraction networks, so as to perform weight distribution; the test set is also from different earthquake periods, which is helpful for carrying out accurate performance evaluation on the integrated deep learning model.
In addition, the method also comprises the steps of amplifying the training data set, wherein the process is as follows: the original training data set is a 1D time sequence waveform with P phase as center 40000 data points, we start from random points between 1 and 20000 and extend to 6144 data points (which can cover most of acoustic emission waveforms), which not only allows the P phase to be at any position in the training process, but also adds a large amount of trainable random noise data, thereby constructing a data volume of the training data set as 6144×1×10000.
After the training data set is constructed, training the two backbone neural networks by using the training data set, and updating network parameters by the prediction probability value output by the backbone neural network and the error of the accurate label; and multi-classification cross entropy is adopted as a damage function during updating;
after training is completed, model training parameters of the two deep neural networks are stored;
and then evaluating the relative contribution of each backbone characteristic extraction network by using the verification data set, and carrying out normalized weight distribution by using the harmonic performance index (F1 score) of each network so as to carry out weighted integration on the two output prediction sequences.
The method and formula of the weighted integration are as follows:
applying the verification set to two independent backbone feature extraction networks, calculating True Positives (TP), false Positives (FP) and False Negatives (FN) of the verification set sample in each backbone feature extraction network, further calculating F1 score, and defining two F1 score indexes (respectively as F1 T And F1 B ) And carrying out weight normalization, and then carrying out weighted combination on the prediction results of the two backbone feature extraction networks by using the calculated weights to construct the P-phase integrated classifier.
1)
2)
3)P AEbagging ＝w 1 *P Transformer +w 2 *P Bi-LSTM (w 1 +w 2 ＝1)
Wherein P is m Is the probability output of model m (m.epsilon. { transducer, bi-LSTM, AEbagging }).
In the invention, when acoustic emission data or microseismic data are processed, the specific flow of the deep learning model is integrated:
the input acoustic emission (microseismic) training dataset contained a 1-dimensional waveform 6144 data points long. Firstly, two base models (namely a transducer model and a Bi-LSTM model) are independently and parallelly trained by using limited acoustic emission or microseismic artificial annotation data (training data of about 10000 one-dimensional waveform sequences), and trained model parameters and model structures are saved.
Then, two trained models are integrated and built into a new network architecture model using an integrated deep learning algorithm. For true continuous waveform data, it was cut into a window of 6144 data points (2.048 ms with overlap) and the waveform data normalized. These slice waveforms are then applied to an integrated deep learning model where the two backbone feature extraction networks extract the different Gao Jiete collection independently and where the integrated learning layer is used to construct an efficient robust P-phase probability predictor output P-phase probability sequence.
In order to verify the detection effect of the integrated deep learning model, performance test is carried out on the integrated deep learning model based on a test data set and multi-channel real continuous waveform data, and the test result is as follows:
and testing the performance of the model by utilizing the acoustic emission data.
First, an acoustic emission test dataset was created, containing 2500 6144 data points long one-dimensional waveform sequences, 25% of the training set. And dividing the acoustic emission test data into 6 categories according to the signal-to-noise ratio range, and carrying out prediction result statistics. The statistical results are shown in table 1,
TABLE 1
As can be seen from the data in table 1: in the test of different signal-to-noise ratios, three evaluation indexes (namely, the precision rate, the recall rate and the F1-fraction) are gradually increased along with the improvement of the signal-to-noise ratio, and even for the acoustic emission waveform with low signal-to-noise ratio (4-10 dB), the model of the invention keeps higher precision rate and recall rate (0.924 and 0.922). Therefore, the model of the invention has the characteristics of high detection precision and strong generalization performance.
Next, the model of the present invention is applied to the actual continuous waveform data.
From the 12 channels, 6 channels of continuous waveform data were selected to demonstrate the effect of the model application. Fig. 6 (a) to (d) show the model pickup results of 4 representative continuous waveform data, respectively.
The data indicate that: the deep learning model of the present invention is capable of accurately detecting and picking up acoustic emission events in multi-channel data when processing multiple events in unseen and long-time continuous waveform data. Meanwhile, from fig. 6 (b), it is found that the integrated deep learning method uses complementary correction of the prediction results mapped by two independent high-order feature sets to reduce the sensitivity of the model of the present application to noise level, so that the model can well pick up small events with high background noise.
In addition, the present invention applies the conventional STA/LTA method to the continuous waveform sequence of the channel 11 (CH 11), the short sliding window is set to 50 μs, and the long sliding window is set to 1000 μs. The data of fig. 6 (a) and (c) show that the STA/LTA algorithm cannot detect event waveforms with signal-to-noise ratios below 5.65dB (or about 6 dB) compared to the model of the present invention; and as can be seen from fig. 6 (d), when processing an event waveform with a signal-to-noise ratio of 8.76dB, the peak value outputted using the STA/LTA algorithm exhibits a phenomenon of oscillation, and thus when facing waveform data with different background noise, it may be difficult to set a threshold value to ensure that this type of signal is always accurately captured.
Furthermore, in order to analyze the space-time evolution mechanism of the fault friction sliding process, the acoustic emission waveforms picked up by the multiple channels are correlated, and an acoustic emission event catalog is constructed. The present application refers to acoustic emission signals recorded in a single channel as acoustic emission waveforms, and to acoustic emission waveforms having P-phase arrival intervals of less than 750 data points (corresponding to a sampling frequency of 3 MHz) recorded in three or more channels as acoustic emission events. Fig. 7-10 are correlation results of 3, 6, 8, and 12 acoustic emission waveforms, respectively, after model picking, each set randomly selecting two acoustic emission events for presentation, while also demonstrating the error of the model picking (dark line/red line) and manual labeling (light line/green line) of the present invention. These results indicate that the model of the present invention has lower pick-up errors and the acoustic emission catalogue has better correlation effect.
Example III
The embodiment discloses a rock fault friction microseismic detection system based on integrated deep learning, which comprises an acoustic emission monitoring device, a data acquisition device and a data acquisition device, wherein the acoustic emission monitoring device is used for monitoring and recording acoustic emission/microseismic waveform data in a fault friction process;
the microseism event detection device is internally provided with an integrated deep learning model constructed in the rock fault friction microseism detection method based on integrated deep learning, and is used for processing real continuous waveform data in the rock fault friction process, obtaining a P-phase probability sequence by using the integrated deep learning model, and detecting and picking up acoustic emission phases according to the obtained P-phase probability sequence.
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. The rock fault friction microseismic detection method based on integrated deep learning is characterized by comprising the following steps of:
constructing an integrated deep learning model, the integrated deep learning model comprising:
the first backbone characteristic extraction network is used for extracting local P-phase characteristics in waveform data and sequentially comprises a first downsampling section, a local phase characteristic extraction layer and a first upsampling section;
the second backbone feature extraction network is used for extracting complete waveform features with global context information in waveform data and sequentially comprises a second downsampling section, a global waveform feature extraction layer and a second upsampling section; and
the integrated learning layer is used for constructing a P-phase integrated predictor by carrying out weighted integration on output results of the first backbone feature extraction network and the second backbone feature extraction network, and the P-phase integrated predictor is used for obtaining a P-phase probability sequence;
and acquiring continuous waveform data monitored in the fault friction process, inputting the continuous waveform data into an integrated deep learning model, obtaining a P-phase probability sequence, and detecting and picking up acoustic emission phases according to the obtained P-phase probability sequence.
2. The rock fault friction microseismic detection method based on integrated deep learning of claim 1, wherein the first downsampling section and the second downsampling section have the same structure and respectively comprise a plurality of layers of convolution networks, and each layer of convolution network comprises a one-dimensional convolution layer and a maximum pooling layer.
3. The method for detecting rock fault friction microseismic according to claim 1, wherein the first upsampling section and the second upsampling section have the same structure and respectively comprise a plurality of one-dimensional convolution layers.
4. A rock fault friction microseismic detection method based on integrated deep learning according to claim 3, wherein the convolutional network layer in the first/second downsampling section has 7 layers, and the one-dimensional convolutional layer in the first/second upsampling section has 8 layers.
5. The rock fault friction microseismic detection method based on integrated deep learning according to claim 1, wherein the local phase feature extraction layer sequentially comprises,
a two-way long-short-term memory layer for converting the downsampled features into higher-order features with time dependence;
a transducer-global attention mechanism for directing neural network attention to a portion associated with the acoustic emission waveform; and
a transducer-local attention mechanism for setting a large weight to a narrow window around the P-phase first arrival to direct neural network attention to a local feature associated with the waveform phase.
6. The method for detecting the friction microseismic of the rock fault based on integrated deep learning according to claim 5, wherein a one-way long-short-term memory layer is arranged before and after the transducer-global attention mechanism and used for integrating position information so that the states of the following encoders are all provided with position sensing.
7. The method for rock fault friction microseismic detection based on integrated deep learning of claim 1, wherein the global waveform feature extraction layer comprises a plurality of two-way long and short term memory layers for converting downsampling features into higher order timing features with global context information.
8. The rock fault friction microseismic detection method based on integrated deep learning according to claim 1, wherein a Softmax function is connected after the first upsampling segment and is used for mapping the P-phase features extracted by the first neural network into a first probability vector point to point;
and the second upsampling section is connected with a Softmax function for mapping the global high-order time sequence features extracted by the second neural network into a second probability vector point to point.
9. The rock fault friction microseismic detection method based on integrated deep learning according to claim 1, wherein a training data set is constructed, the integrated deep learning model is trained, the training data set comprises a front earthquake, a main earthquake and a aftershock, the magnitude distribution follows the Gordon's Barbell-Rickett law, and the statistical characteristic distribution is consistent with the characteristic distribution of the rest earthquake period.
10. The rock fault friction microseismic detection system based on integrated deep learning is characterized by comprising an acoustic emission monitoring device, wherein the acoustic emission monitoring device is used for recording acoustic emission waveform data in a fault friction process;
the microseism event detection device is internally provided with an integrated deep learning model constructed in the rock fault friction microseism detection method based on integrated deep learning according to any one of claims 1-9, wherein the integrated deep learning model is used for processing real continuous waveform data recorded in a fault friction process, a P-phase probability sequence is obtained by using the integrated deep learning model, and acoustic emission phases are detected and picked up according to the obtained P-phase probability sequence.
CN202311484390.XA 2023-11-09 2023-11-09 Rock fault friction microseismic detection method and system based on integrated deep learning Active CN117609874B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311484390.XA CN117609874B (en) 2023-11-09 2023-11-09 Rock fault friction microseismic detection method and system based on integrated deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311484390.XA CN117609874B (en) 2023-11-09 2023-11-09 Rock fault friction microseismic detection method and system based on integrated deep learning

Publications (2)

Publication Number Publication Date
CN117609874A true CN117609874A (en) 2024-02-27
CN117609874B CN117609874B (en) 2024-05-10

Family

ID=89948752

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311484390.XA Active CN117609874B (en) 2023-11-09 2023-11-09 Rock fault friction microseismic detection method and system based on integrated deep learning

Country Status (1)

Country Link
CN (1) CN117609874B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220351037A1 (en) * 2021-04-30 2022-11-03 Saudi Arabian Oil Company Method and system for spectroscopic prediction of subsurface properties using machine learning
CN115436993A (en) * 2022-07-29 2022-12-06 太原理工大学 Seismic signal P wave detection method based on PPNet network
US20230119211A1 (en) * 2021-10-15 2023-04-20 Hohai University Method For Extracting Dam Emergency Event Based On Dual Attention Mechanism
CN116594057A (en) * 2023-05-17 2023-08-15 中国地震局地球物理研究所 Earthquake early warning method and device based on deep learning and edge calculation
CN116719936A (en) * 2023-06-15 2023-09-08 湖北大学 Network unreliable information early detection method based on ensemble learning
CN116822580A (en) * 2023-08-08 2023-09-29 西安石油大学 Method, system, equipment and medium for predicting mechanical drilling speed time sequence characteristics based on transducer model

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220351037A1 (en) * 2021-04-30 2022-11-03 Saudi Arabian Oil Company Method and system for spectroscopic prediction of subsurface properties using machine learning
US20230119211A1 (en) * 2021-10-15 2023-04-20 Hohai University Method For Extracting Dam Emergency Event Based On Dual Attention Mechanism
CN115436993A (en) * 2022-07-29 2022-12-06 太原理工大学 Seismic signal P wave detection method based on PPNet network
CN116594057A (en) * 2023-05-17 2023-08-15 中国地震局地球物理研究所 Earthquake early warning method and device based on deep learning and edge calculation
CN116719936A (en) * 2023-06-15 2023-09-08 湖北大学 Network unreliable information early detection method based on ensemble learning
CN116822580A (en) * 2023-08-08 2023-09-29 西安石油大学 Method, system, equipment and medium for predicting mechanical drilling speed time sequence characteristics based on transducer model

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
李健 等: "基于深度卷积神经网络的地震震相拾取方法研究", 地球物理学报, vol. 63, no. 04, 30 April 2020 (2020-04-30), pages 1591 - 1606 *

Also Published As

Publication number Publication date
CN117609874B (en) 2024-05-10

Similar Documents

Publication Publication Date Title
Zhu et al. Deep learning for seismic phase detection and picking in the aftershock zone of 2008 Mw7. 9 Wenchuan Earthquake
Münchmeyer et al. Which picker fits my data? A quantitative evaluation of deep learning based seismic pickers
Zhu et al. PhaseNet: a deep-neural-network-based seismic arrival-time picking method
Ross et al. P wave arrival picking and first‐motion polarity determination with deep learning
Michelini et al. INSTANCE–the Italian seismic dataset for machine learning
Guo et al. AEnet: Automatic picking of P-wave first arrivals using deep learning
Zhu et al. Seismic signal augmentation to improve generalization of deep neural networks
Del Pezzo et al. Discrimination of earthquakes and underwater explosions using neural networks
CN106707099A (en) Monitoring and locating method based on abnormal electricity consumption detection module
CN112749637B (en) SNN-based distributed optical fiber sensing signal identification method
CN113568043B (en) Three-stage seismic phase picking method based on deep convolutional neural network
Cai et al. Anomaly detection of earthquake precursor data using long short-term memory networks
Shi et al. MALMI: An automated earthquake detection and location workflow based on machine learning and waveform migration
CN109613610B (en) Automatic picking method for microseism signal arrival time difference
CN111027452A (en) Microseismic signal arrival time and seismic phase identification method and system based on deep neural network
Rojas et al. Artificial neural networks as emerging tools for earthquake detection
Gillings et al. Nocturnal flight calling behaviour of thrushes in relation to artificial light at night
Zheng et al. Deep neural networks design and analysis for automatic phase pickers from three-component microseismic recordings
CN111898644A (en) Intelligent identification method for health state of aerospace liquid engine under fault-free sample
Giudicepietro et al. Fast discrimination of local earthquakes using a neural approach
Kohlsdorf et al. An auto encoder for audio dolphin communication
Song et al. MEANet: Magnitude estimation via physics-based features time series, an attention mechanism, and neural networks
Martínez et al. Advanced signal recognition methods applied to seismo-volcanic events from Planchon Peteroa Volcanic Complex: Deep Neural Network classifier
Chen et al. CubeNet: Array‐based seismic phase picking with deep learning
Zhang et al. Integrated processing method for microseismic signal based on deep neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant