WO2020220416A1 - 一种基于深度学习的微震信号分类辨识方法 - Google Patents

一种基于深度学习的微震信号分类辨识方法 Download PDF

Info

Publication number
WO2020220416A1
WO2020220416A1 PCT/CN2019/088270 CN2019088270W WO2020220416A1 WO 2020220416 A1 WO2020220416 A1 WO 2020220416A1 CN 2019088270 W CN2019088270 W CN 2019088270W WO 2020220416 A1 WO2020220416 A1 WO 2020220416A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
signals
microseismic
coal
classification
Prior art date
Application number
PCT/CN2019/088270
Other languages
English (en)
French (fr)
Inventor
张杏莉
赵震华
卢新明
贾瑞生
Original Assignee
山东科技大学
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 山东科技大学 filed Critical 山东科技大学
Publication of WO2020220416A1 publication Critical patent/WO2020220416A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G01MEASURING; TESTING
    • G01VGEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28Processing seismic data, e.g. analysis, for interpretation, for correction
    • G01V1/288Event detection in seismic signals, e.g. microseismics
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01VGEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/28Processing seismic data, e.g. analysis, for interpretation, for correction
    • G01V1/30Analysis

Definitions

  • the invention belongs to the field of signal analysis and identification, and specifically relates to a method for classification and identification of microseismic signals based on deep learning.
  • the signals collected by the coal mine microseismic monitoring system also include a large number of blasting vibration signals generated by coal mine blasting operations.
  • the waveform of the coal rock fracture microseismic signal and the blasting vibration signal waveform are very similar.
  • the rock fracture microseismic signal is identified from the massive monitoring data, and the manual identification method is adopted, which is difficult to identify and low efficiency.
  • the identification methods of coal mine microseismic signals and blasting signals at home and abroad mainly include Fourier transform, wavelet transform, wavelet packet transform and empirical mode decomposition (EMD).
  • EMD empirical mode decomposition
  • the traditional Fourier transform is mainly used to analyze periodic stationary signals, and the analysis effect of random and non-stationary microseismic signals including spikes and sudden changes is not good; the EMD method has boundary effects and modal aliasing phenomena, leading to EMD
  • the decomposed signal has instability and non-uniqueness.
  • the purpose of the present invention is to provide a method for classification and identification of microseismic signals based on deep learning, which extracts identification sensitive features from the microseismic signals of coal and rock fracture and blasting vibration signals, and applies deep learning technology to construct two types of vibration signal classifications based on deep neural networks
  • the identification model uses a classifier to identify the test set data.
  • a method for classification and identification of microseismic signals based on deep learning includes the following steps in sequence:
  • Step 1 Select M coal and rock fracture microseismic signals and N blasting vibration signals to form sample data sets of two types of vibration signals;
  • Step 2 Extract the main frequency F m , post-peak attenuation coefficient b, and energy center of gravity coefficient C x of M coal and rock fracture microseismic signals and N blasting vibration signals to form the sample feature data training set and test set;
  • Step 3 Construct a classification and identification model of coal and rock fracture microseismic signals and blasting vibration signals based on a four-layer deep neural network, train the signal classification model with the training set data, and verify the classification and identification effect of the signal classification model with the test set data, through cross-training Continuously improve classification accuracy;
  • Step 4 Extract the feature vector of the signal to be identified and input it into the classification identification model to obtain the identification result.
  • step 2 the specific steps of extracting the main frequency F m of the M coal and rock fracture microseismic signals and the N blasting vibration signals are:
  • Step 2.1.1 Calculate the frequency spectrum of the signal according to formula (1);
  • Step 2.1.2 Calculate the dominant frequency of the signal according to formula (2):
  • STA/LTA Use long and short time window method
  • i is the i-th sampling point
  • sn is the length of the short-time window
  • ln is the length of the long-time window
  • is the trigger threshold of STA/LTA
  • CF(j) is j The value of the characteristic function of the information at the moment
  • Step 2.2.1 Use the long and short time window method to pick up the end time point of the signal
  • Step 2.2.2 Use cubic spline interpolation to find the envelope of the signal
  • Step 2.2.3 Use equation (7) to fit the envelope
  • x is the signal amplitude
  • t is the sampling point
  • a and b are fitting parameters
  • parameter a is related to the signal peak
  • parameter b is related to the signal attenuation rate.
  • Step 2.3.2 Calculate the energy corresponding to each component U k as E k according to formula (8), namely
  • K is the number of variational modes
  • Step 2.3.3 Calculate the percentage of the energy of each modal component to the total energy of the original signal according to formula (9);
  • Step 2.3.4 Calculate the energy center of gravity coefficient C x (0 ⁇ C x ⁇ 1) of the energy distribution on the X axis according to formula (10):
  • the four-layer deep neural network includes an input layer, an output layer and two hidden layers, and the two hidden layers each include 10 hidden neurons.
  • the present invention applies deep learning technology to use three sensitive identification features of two types of vibration signals as identification feature vectors to construct a classification and identification model based on deep neural networks.
  • the energy center of gravity coefficient C x constitutes the sample feature data training set and test set; based on the four-layer deep neural network to build a classification and identification model, use the training set data to train the signal classification model, and use the trained signal classification model to perform the test set data Recognition, the classification accuracy is continuously improved through the cross-training method; the feature vector of the signal to be recognized is extracted and input into the signal model to obtain the recognition result.
  • This method can realize the effective identification of the microseismic signal of coal and rock fracture and the blasting vibration signal.
  • the invention is based on the current coal mine monitoring big data environment, according to the characteristics and characteristics of the coal and rock fracture microseismic signals and blasting vibration signals, by extracting the identification sensitive features of the two types of vibration signals, deep learning technology is used to establish a deep neural network model to train deep nerves
  • the network classifier realizes accurate and effective identification of the two types of vibration signals.
  • This method has the characteristics of simple algorithm, strong adaptability and real-time performance, and high identification accuracy. It can effectively classify the microseismic signals and blasting vibration signals of coal and rock fractures in coal mines, and has good technical value and application prospects.
  • Figure 1 is a flowchart of a method for classification and identification of microseismic signals based on deep learning in the present invention
  • Figure 2 is a schematic diagram of the signal x(t) to be identified
  • Figure 3 is a schematic diagram of extracting the dominant frequency characteristics of the signal to be identified x(t);
  • Figure 4 is a schematic diagram of extracting the post-peak attenuation coefficient characteristics of the signal to be identified x(t);
  • Figure 5 is a schematic diagram of extracting features of the energy barycentric coefficient of the signal to be identified x(t);
  • FIG. 6 is a schematic diagram of the four-layer deep neural network structure of the signal classification and identification model of the present invention.
  • the present invention proposes a classification and identification method for microseismic signals based on deep learning.
  • Step 1 Select M coal and rock fracture microseismic signals and N blasting vibration signals to form sample data sets of two types of vibration signals;
  • Step 2 Extract the main frequency F m , post-peak attenuation coefficient b, and energy center of gravity coefficient C x of M coal and rock fracture microseismic signals and N blasting vibration signals to form the sample feature data training set and test set;
  • step 2 the specific steps of extracting the main frequency F m of the M coal and rock fracture microseismic signals and the N blasting vibration signals are:
  • Step 2.1.1 Calculate the continuous spectrum of the signal according to formula (1);
  • Step 2.2.2 Calculate the dominant frequency of the signal according to formula (2):
  • step 2 the specific steps of extracting the post-peak attenuation coefficient b of the M coal and rock fracture microseismic signals and the N blasting vibration signals are:
  • STA/LTA The long and short time window method (STA/LTA) is used to automatically pick up the end time point of the signal. among them:
  • i is the i-th sampling point
  • sn is the length of the short-term window
  • ln is the length of the long-term window
  • is the trigger threshold of STA/LTA
  • CF(j) is the characteristic function value of the information at time j.
  • Step 2.2.1 Use the long and short time window method to pick up the end time point of the signal
  • Step 2.2.2 Use the three-spline interpolation method to obtain the envelopes of the two signals
  • Step 2.2.3 Use equation (7) to fit the envelope
  • x is the signal amplitude
  • t is the sampling point
  • a and b are fitting parameters.
  • the parameter a is related to the signal peak value
  • the parameter b is related to the signal attenuation rate.
  • the larger the value of b the faster the signal attenuation rate, and vice versa. Therefore, the parameter b is defined as the attenuation coefficient of the signal.
  • step 2 the specific steps of extracting the energy-gravity coefficient C x of the M coal and rock fracture microseismic signals and the N blasting vibration signals are:
  • Step 2.3.2 Calculate the energy corresponding to each component U k as E k according to formula (8), namely
  • T is the number of signal sampling points
  • K is the number of variational modes.
  • Step 2.3.3 Calculate the percentage of the energy of each modal component to the total energy of the original signal according to formula (9);
  • Step 2.3.4 Calculate the X-axis center of gravity coefficient C x (0 ⁇ C x ⁇ 1) of the energy distribution according to formula (10):
  • Step 3 Construct a classification and identification model of coal and rock fracture microseismic signals and blasting vibration signals based on a four-layer deep neural network, train the signal classification model with the training set data, and verify the classification and identification effect of the signal classification model with the test set data, through cross-training Continuously improve classification accuracy;
  • Step 4 Extract the feature vector of the signal to be identified and input it into the signal classification model to obtain the identification result.
  • the signal sampling point data is shown in Table 1.
  • the cubic spline interpolation method is used to obtain the envelope of the signal and fit the envelope.
  • the fitting accuracy is represented by the correction correlation coefficient Adj.R-Square. The closer the coefficient is to 1, the higher the fitting accuracy and the smoother the signal attenuation process.
  • the obtained post-peak attenuation coefficient b 6.238, the result is shown in Figure 4, and the correlation value is shown in Table 2.
  • the energy center coefficient C x 0.7185 according to formula (10) in step 2, as shown in FIG. 5, which is a schematic diagram of the four-layer deep neural network structure of the signal identification model of the present invention.
  • step 4 input the eigenvector value (24.5, 6.238, 0.7185) of the signal to be identified into the signal classification and identification model trained in step 3, and the identification result is that the identification signal is a coal-rock fracture microseismic signal.
  • test set data is used to verify the classification and identification effect of the model, and the test results are shown in Table 5.
  • the coal and rock fracture microseismic signal and the blasting vibration signal are both vibration signals, the two types of vibration signals have significant differences in the three characteristics of main frequency, post-peak attenuation coefficient and energy center of gravity coefficient. Therefore, based on this feature, the Identify the feature vector of the signal of the category, use the deep learning technology to build a signal classifier, and input the feature vector of the signal to be identified into the classifier to realize the classification and identification of the signal to be detected.

Abstract

一种基于深度学习的微震信号分类辨识方法,包括以下步骤:步骤1、建立微震信号与爆破信号的样本数据库;步骤2、提取样本信号的主频、峰后衰减系数和能量重心系数特征,构成样本特征数据训练集和测试集;步骤3、使用样本特征数据训练集训练深度神经网络分类辨识模型,利用测试集数据验证信号分类辨识模型的分类辨识效果,并通过交叉训练不断提升分类精度;步骤4、提取待辨识信号的特征向量,输入信号分类模型中,得到辨识结果。该方法具有算法简单、自适应性和实时性强、辨识准确率高的特点,能对煤矿微震信号和爆破信号进行有效的分类。

Description

一种基于深度学习的微震信号分类辨识方法 技术领域
本发明属于信号分析及识别领域,具体涉及一种基于深度学习的微震信号分类辨识方法。
背景技术
煤矿微震监测系统采集到的信号中除了大量有效的煤岩破裂微震信号,还包含了煤矿爆破作业产生的大量爆破震动信号,煤岩破裂微震信号波形与爆破震动信号波形又极为相似,需将煤岩破裂微震信号从海量监测数据中识别出来,采用人工识别方式,识别难度大,工作效率低。
目前,国内外对煤矿微震信号和爆破信号的识别方法主要包括傅里叶变换、小波变换、小波包变换和经验模态分解(Empirical Mode Decomposition,EMD)等。如传统的傅里叶变换主要用于分析周期性平稳信号,对包含有尖峰和突变的随机性、非平稳性微震信号分析效果欠佳;EMD方法存在边界效应及模态混叠现象,导致EMD分解信号具有不稳定性和不唯一性,EMD的这些缺陷使得在信号辨识时难免存在弊端。这些方法用于信号分析时均存在一定程度的解决了两类震动信号的辨识问题,但忽略了当前煤矿监测大数据环境和深度学习等新一代信息化技术在信号分类辨识技术中的应用,影响了信号分类辨识精度的进一步提高。
发明内容
本发明的目的在于提供一种基于深度学习的微震信号分类辨识方法,从煤岩破裂微震信号和爆破震动信号中提取辨识敏感特征,应用深度学习技术,构建基于深度神经网络的两类震动信号分类辨识模型,利用分类器对测试集数据进行识别。
为了实现上述目的,本发明采用了如下技术方案:
一种基于深度学习的微震信号分类辨识方法,依次包括如下步骤:
步骤1:分别选取M个煤岩破裂微震信号和N个爆破震动信号构成两类震动信号的样本数据集;
步骤2:提取M个煤岩破裂微震信号和N个爆破震动信号的主频F m、峰后衰减系数b、能量重心系数C x构成样本特征数据训练集和测试集;
步骤3:基于四层深度神经网络构建煤岩破裂微震信号与爆破震动信号的分类辨识模型,用训练集数据训练该信号分类模型,利用测试集数据验证信号分类模型的分类辨识效果,通过交叉训练不断提升分类精度;
步骤4:提取待辨识信号的特征向量,输入所述的分类辨识模型中,得到辨识结果。
进一步的,步骤2中,提取所述的M个煤岩破裂微震信号和N个爆破震动信号的主频 F m的具体步骤为:
假设煤岩破裂微震信号或爆破震动信号为x(t),t=1,2,…,T;
步骤2.1.1:根据式(1)计算得到信号的频谱;
Figure PCTCN2019088270-appb-000001
式(1)中,X(ω)为信号x(t)的频谱,j 2=-1;
步骤2.1.2:根据式(2)计算信号的主频:
F m=max(X(ω))        (2)。
进一步的,步骤2中,提取所述的M个煤岩破裂微震信号和N个爆破震动信号的峰后衰减系数b的具体步骤为:假设煤岩破裂微震信号或爆破震动信号为x(t),t=1,2,…,T;
采用长短时窗法(STA/LTA)自动拾取信号的终止时刻点,其中:
Figure PCTCN2019088270-appb-000002
Figure PCTCN2019088270-appb-000003
Figure PCTCN2019088270-appb-000004
CF(j)=x(j) 2-x(j-1)·x(j+1)       (6)
上述式(3)-式(6)中:i为第i个采样点,sn为短时窗长度,ln为长时窗长度,λ为STA/LTA的触发阀值,CF(j)为j时刻的关于信息的特征函数值;
求解信号峰后衰减系数b的具体步骤如下:
步骤2.2.1:利用长短时窗法拾取信号的终止时刻点;
步骤2.2.2:利用三次样条插值法求取信号的包络线;
步骤2.2.3:利用式(7)对包络线进行拟合;
x=at b       (7)
式(7)中:x为信号振幅,t为采样点,a、b为拟合参数;参数a与信号峰相关,参数b与信号衰减速率相关,通常b值越大,信号的衰减速率越快,反之亦然;因此将参数b定义为信号的衰减系数。
进一步的,步骤2中,提取所述的M个煤岩破裂微震信号和N个爆破震动信号的能量重心系数C x的具体步骤为:假设煤岩破裂微震信号或爆破震动信号为x(t),t=1,2,…,T;
步骤2.3.1:对信号x(t),t=1,2,…,T进行VMD分解,得到的K个变分模态分量,记为{U 1,…,U k,…,U K};
步骤2.3.2:根据式(8)计算各分量U k对应的能量为E k,即
Figure PCTCN2019088270-appb-000005
式(8)中,x ki(k=1,2,…,K;i=1,2,…,T)为第k个变分模态分量U k的离散点幅值,T为信号的采样点个数,K为变分模态个数;
步骤2.3.3:根据式(9)计算各模态分量能量占原始信号总能量的百分比为;
Figure PCTCN2019088270-appb-000006
得到能量分布特征向量P=(P(1),…,P(k),…,P(K)),并构造能量分布平面;
步骤2.3.4:根据式(10)计算能量分布X轴能量重心系数C x(0<C x≤1):
Figure PCTCN2019088270-appb-000007
进一步的,在步骤3中,所述的四层的深度神经网络中包括输入层、输出层和两层隐含层,两层隐含层分别包含10个隐含神经元。
本发明原理如下:
为实现煤岩破裂微震信号和爆破震动信号的有效分类辨识,本发明应用深度学习技术,以两种震动信号的三个敏感辨识特征为辨识特征向量,构建基于深度神经网络的分类辨识模型。首先从历史监测数据中选择M个起跳明显的煤岩破裂微震信号和N个典型爆破震动信号构成样本数据集;分别对M个微震信号和N个爆破信号提取主频F m、峰后衰减系数b、能量重心系数C x构成样本特征数据训练集和测试集;基于四层深度神经网络构建分类辨识模型,利用训练集数据训练信号分类模型,利用训练好的信号分类模型,对测试集数据进行识别,通过交叉训练法不断提升分类精度;提取待辨识信号的特征向量,输入信号模型中,得到辨识结果。该方法可以实现对煤岩破裂微震信号和爆破震动信号的有效辨识。
与现有技术相比,本发明带来了以下有益技术效果:
本发明基于当前煤矿监测大数据环境,根据煤岩破裂微震信号和爆破震动信号的自身特性与特点,通过提取两类震动信号的辨识敏感特征,采用深度学习技术建立深度神经网络模型,训练深度神经网络分类器,实现对两类震动信号的准确、有效辨识。该方法具有算法简单、自适应性和实时性强、辨识准确率高的特点,能对煤矿煤岩破裂微震信号和爆破震动信号进行有效的分类,具有很好的技术价值和应用前景。
附图说明
下面结合附图对本发明做进一步说明:
图1为本发明一种基于深度学习的微震信号分类辨识方法的流程图;
图2为待辨识信号x(t)的示意图;
图3为提取待辨识信号x(t)的主频特征示意图;
图4为提取待辨识信号x(t)的峰后衰减系数特征示意图;
图5为提取待辨识信号x(t)的能量重心系数特征示意图;
图6为本发明信号分类辨识模型的四层深度神经网络结构示意图。
具体实施方式
本发明提出了一种基于深度学习的微震信号分类辨识方法,为了使本发明的优点、技术方案更加清楚、明确,下面结合具体实施例对本发明做详细说明。
一种基于深度学习的微震信号分类辨识方法,其流程如图1所示,具体包括如下步骤:
步骤1:分别选取M个煤岩破裂微震信号和N个爆破震动信号构成两类震动信号的样本数据集;
步骤2:提取M个煤岩破裂微震信号和N个爆破震动信号的主频F m、峰后衰减系数b、能量重心系数C x构成样本特征数据训练集和测试集;
进一步的,步骤2中,提取所述的M个煤岩破裂微震信号和N个爆破震动信号的主频F m的具体步骤为:
假设煤岩破裂微震信号或爆破震动信号为x(t),t=1,2,…,T;
步骤2.1.1:根据式(1)计算得到信号的连续频谱;
Figure PCTCN2019088270-appb-000008
式(1)中,X(ω)为信号x(t)的频谱,j 2=-1;
步骤2.2.2:根据式(2)计算信号的主频:
F m=max(X(ω))        (2)
进一步的,步骤2中,提取所述的M个煤岩破裂微震信号和N个爆破震动信号的峰后衰减系数b的具体步骤为:
假设煤岩破裂微震信号或爆破震动信号为x(t),t=1,2,…,T;
采用长短时窗法(STA/LTA)自动拾取信号的终止时刻点。其中:
Figure PCTCN2019088270-appb-000009
Figure PCTCN2019088270-appb-000010
Figure PCTCN2019088270-appb-000011
CF(j)=x(j) 2-x(j-1)·x(j+1)        (6)
其中,i为第i个采样点,sn为短时窗长度,ln为长时窗长度,λ为STA/LTA的触发阀值, CF(j)为j时刻的关于信息的特征函数值。
求解信号峰后衰减系数的具体步骤如下:
步骤2.2.1:利用长短时窗法拾取信号的终止时刻点;
步骤2.2.2:利用三条样条插值法求取两种信号的包络线;
步骤2.2.3:利用式(7)对包络线进行拟合;
x=at b         (7)
其中,x为信号振幅,t为采样点,a、b为拟合参数。参数a与信号峰值相关,参数b与信号衰减速率相关,通常b值越大,信号的衰减速率越快,反之亦然。因此将参数b定义为信号的衰减系数。
进一步的,步骤2中,提取所述的M个煤岩破裂微震信号和N个爆破震动信号的能量重心系数C x的具体步骤为:
假设煤岩破裂微震信号或爆破震动信号为x(t),t=1,2,…,T;
步骤2.3.1:对信号x(t),t=1,2,…,T进行VMD分解,得到的K个变分模态分量,记为{U 1,…,U k,…,U K};
步骤2.3.2:根据式(8)计算各分量U k对应的能量为E k,即
Figure PCTCN2019088270-appb-000012
式中,x ki(k=1,2,…,K;i=1,2,…,T)为第k个变分模态分量U k的离散点幅值,T为信号的采样点个数,K为变分模态个数。
步骤2.3.3:根据式(9)计算各模态分量能量占原始信号总能量的百分比为;
Figure PCTCN2019088270-appb-000013
得到能量分布特征向量P=(P(1),…,P(k),…,P(K)),并构造能量分布平面;
步骤2.3.4:根据公式(10)计算能量分布X轴重心系数C x(0<C x≤1):
Figure PCTCN2019088270-appb-000014
步骤3:基于四层深度神经网络构建煤岩破裂微震信号与爆破震动信号的分类辨识模型,用训练集数据训练该信号分类模型,利用测试集数据验证信号分类模型的分类辨识效果,通过交叉训练不断提升分类精度;
步骤4:提取待辨识信号的特征向量,输入信号分类模型中,得到辨识结果。
如图2所示,从监测数据中获取以时间(s)为横轴,振幅为纵轴,采样频率fs=1000Hz的待辨识信号x(t),t=1,2,…,4000,微震信号采样点数据见表1。
表1待辨识信号采样点数据(可以存储于Excel表中)
序号 采样点(N) 振幅
1*1/fs 1 -2.097e-06
2*1/fs 2 -8.842e-06
3*1/fs 3 -9.590e-06
4*1/fs 4 3.960e-06
5*1/fs 5 6.416e-06
3999*1/fs 3999 2.920e-05
4000*1/fs 4000 2.804e-05
按照步骤2中的提取主频的步骤提取待辨识信号x(t),t=1,2,…,4000主频F m的结果为Fm=24.5,如图3所示。
按照步骤2中的提取峰后衰减系数的步骤提取待辨识信号x(t),t=1,2,…,4000峰后衰减系数b,取微震信号的短时窗sn=10,长时窗ln=180,触发阀值λ=7。用三次样条插值法对信号求取包络线,并对包络线进行拟合。拟合精度用校正相关系数Adj.R-Square表示,该系数越接近1,说明拟合精度越高,信号衰减过程越平稳。得到的峰后衰减系数b=6.238,结果如图4所示,相关值如表2所示。
表2峰后衰减系数b的相关值
Equation Adj.R-Square a b 峰后衰减系数b
x=a×t b 0.9696 1.231e+17 -6.238 6.24
按照步骤2中的提取能量重心系数C x的步骤提取待检测信号x(t),t=1,2,…,4000的能量重心系数C x,对信号x(t)进行VMD分解,取K=6,计算分解各模态分量能量占原始信号总能量的百分比如表3所示。
表3待辨识信号各模态分量能量占原始信号总能量的百分比(%)
P(1) P(2) P(3) P(4) P(5) P(6)
0.4495 1.8778 1.1972 69.4277 16.6905 10.3571
得到的微震信号的能量分布特征向量P,即P=(0.4495,1.8778,1.1972,69.4277,16.6905,10.3571)。并按照步骤2中公式(10)计算能量中心系数C x=0.7185,如图5所示,图6为本发明信号辨识模型的四层深度神经网络结构示意图。
按照步骤4,将待辨识信号的特征向量值(24.5,6.238,0.7185)输入由步骤3训练好的信号分类辨识模型中,得到辨识结果为该辨识信号是煤岩破裂微震信号。
为进一步验证本发明中的信号分类辨识模型的分类精度,从测试集中分别选择15个煤岩破裂微震信号和15个爆破震动信号,其特征数据如表4所示。
表4 15个微震信号和15个爆破信号的样本数据测试集
Figure PCTCN2019088270-appb-000015
利用上述测试集数据验证模型的分类辨识效果,其测试结果如表5所示。
表5 15个微震信号和15个爆破信号的测试结果
信号类型 序号 标签 辨识结果 信号类型 序号 标签 辨识结果
煤岩破裂微震信号 1 0 0 爆破震动信号 1 1 1
煤岩破裂微震信号 2 0 0 爆破震动信号 2 1 1
煤岩破裂微震信号 3 0 0 爆破震动信号 3 1 1
煤岩破裂微震信号 4 0 0 爆破震动信号 4 1 1
煤岩破裂微震信号 5 0 0 爆破震动信号 5 1 1
煤岩破裂微震信号 6 0 0 爆破震动信号 6 1 1
煤岩破裂微震信号 7 0 0 爆破震动信号 7 1 1
煤岩破裂微震信号 8 0 0 爆破震动信号 8 1 1
煤岩破裂微震信号 9 0 0 爆破震动信号 9 1 1
煤岩破裂微震信号 10 0 0 爆破震动信号 10 1 1
煤岩破裂微震信号 11 0 0 爆破震动信号 11 1 0
煤岩破裂微震信号 12 0 0 爆破震动信号 12 1 1
煤岩破裂微震信号 13 0 0 爆破震动信号 13 1 1
煤岩破裂微震信号 14 0 0 爆破震动信号 14 1 1
煤岩破裂微震信号 15 0 0 爆破震动信号 15 1 1
由表5可知,在该30个测试信号中,29个信号的辨识结果正确,仅1组辨识错误,总的分类辨识准确率为96.67%。
煤岩破裂微震信号和爆破震动信号虽然同为震动信号,但两类震动信号在主频、峰后衰 减系数及能量重心系数三个特征上差异显著,因此可根据这一特点,通过提取已明确类别的信号的特征向量,利用深度学习的技术建立信号分类器,将待辨识信号的特征向量输入分类器中,即可实现对待检测信号的分类辨识。
本发明中未述及的部分借鉴现有技术即可实现。
需要说明的是,在本说明书的教导下本领域技术人员所做出的任何等同方式,或明显变型方式均应在本发明的保护范围内。

Claims (5)

  1. 一种基于深度学习的微震信号分类辨识方法,其特征在于,依次包括如下步骤:
    步骤1:分别选取M个煤岩破裂微震信号和N个爆破震动信号构成两类震动信号的样本数据集;
    步骤2:提取M个煤岩破裂微震信号和N个爆破震动信号的主频F m、峰后衰减系数b、能量重心系数C x构成样本特征数据训练集和测试集;
    步骤3:基于四层深度神经网络构建煤岩破裂微震信号与爆破震动信号的分类辨识模型,用训练集数据训练该信号分类模型,利用测试集数据验证信号分类模型的分类辨识效果,通过交叉训练不断提升分类精度;
    步骤4:提取待辨识信号的特征向量,输入所述的分类辨识模型中,得到辨识结果。
  2. 根据权利要求1所述的一种基于深度学习的微震信号分类辨识方法,其特征在于:步骤2中,提取所述的M个煤岩破裂微震信号和N个爆破震动信号的主频F m的具体步骤为:假设煤岩破裂微震信号或爆破震动信号为x(t),t=1,2,…,T;
    步骤2.1.1:根据式(1)计算得到信号的频谱;
    Figure PCTCN2019088270-appb-100001
    式(1)中,X(ω)为信号x(t)的频谱,j 2=-1;
    步骤2.1.2:根据式(2)计算信号的主频:
    F m=max(X(ω))       (2)。
  3. 根据权利要求2所述的一种基于深度学习的微震信号分类辨识方法,其特征在于:步骤2中,提取所述的M个煤岩破裂微震信号和N个爆破震动信号的峰后衰减系数b的具体步骤为:假设煤岩破裂微震信号或爆破震动信号为x(t),t=1,2,…,T;
    采用长短时窗法(STA/LTA)自动拾取信号的终止时刻点,其中:
    Figure PCTCN2019088270-appb-100002
    Figure PCTCN2019088270-appb-100003
    Figure PCTCN2019088270-appb-100004
    CF(j)=x(j) 2-x(j-1)·x(j+1)       (6)
    上述式(3)-式(6)中:i为第i个采样点,sn为短时窗长度,ln为长时窗长度,λ为STA/LTA的触发阀值,CF(j)为j时刻的关于信息的特征函数值;
    求解信号峰后衰减系数b的具体步骤如下:
    步骤2.2.1:利用长短时窗法拾取信号的终止时刻点;
    步骤2.2.2:利用三次样条插值法求取信号的包络线;
    步骤2.2.3:利用式(7)对包络线进行拟合;
    x=at b               (7)
    式(7)中:x为信号振幅,t为采样点,a、b为拟合参数;参数a与信号峰相关,参数b与信号衰减速率相关,通常b值越大,信号的衰减速率越快,反之亦然;因此将参数b定义为信号的衰减系数。
  4. 根据权利要求3所述的一种基于深度学习的微震信号分类辨识方法,其特征在于:步骤2中,提取所述的M个煤岩破裂微震信号和N个爆破震动信号的能量重心系数C x的具体步骤为:假设煤岩破裂微震信号或爆破震动信号为x(t),t=1,2,…,T;
    步骤2.3.1:对信号x(t),t=1,2,…,T进行VMD分解,得到的K个变分模态分量,记为{U 1,…,U k,…,U K};
    步骤2.3.2:根据式(8)计算各分量U k对应的能量为E k,即
    Figure PCTCN2019088270-appb-100005
    式(8)中,x ki(k=1,2,…,K;i=1,2,…,T)为第k个变分模态分量U k的离散点幅值,T为信号的采样点个数,K为变分模态个数;
    步骤2.3.3:根据式(9)计算各模态分量能量占原始信号总能量的百分比为;
    Figure PCTCN2019088270-appb-100006
    得到能量分布特征向量P=(P(1),…,P(k),…,P(K)),并构造能量分布平面;
    步骤2.3.4:根据式(10)计算能量分布X轴能量重心系数C x(0<C x≤1):
    Figure PCTCN2019088270-appb-100007
  5. 根据权利要求1所述的一种基于深度学习的微震信号分类辨识方法,其特征在于,在步骤3中,所述的四层的深度神经网络中包括输入层、输出层和两层隐含层,两层隐含层分别包含10个隐含神经元。
PCT/CN2019/088270 2019-04-28 2019-05-24 一种基于深度学习的微震信号分类辨识方法 WO2020220416A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910348744.5 2019-04-28
CN201910348744.5A CN110133714B (zh) 2019-06-05 2019-06-05 一种基于深度学习的微震信号分类辨识方法

Publications (1)

Publication Number Publication Date
WO2020220416A1 true WO2020220416A1 (zh) 2020-11-05

Family

ID=67575459

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/088270 WO2020220416A1 (zh) 2019-04-28 2019-05-24 一种基于深度学习的微震信号分类辨识方法

Country Status (2)

Country Link
CN (1) CN110133714B (zh)
WO (1) WO2020220416A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114563826A (zh) * 2022-01-25 2022-05-31 中国矿业大学 基于深度学习融合驱动的微震稀疏台网定位方法

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2020220416A1 (zh) * 2019-04-28 2020-11-05 山东科技大学 一种基于深度学习的微震信号分类辨识方法
CN110632643B (zh) * 2019-09-23 2021-03-16 北京无线电计量测试研究所 一种防第三方施工开挖检测报警方法
CN110956764B (zh) * 2019-12-03 2021-07-27 西安科技大学 基于神经网络的地埋线破坏预警装置及预警方法
CN112380949B (zh) * 2020-11-10 2024-03-26 大连理工大学 一种微震波到时点检测方法及系统
CN113514878B (zh) * 2021-07-09 2022-06-24 长沙矿山研究院有限责任公司 一种矿山微震波形类型自动识别方法
CN116990865B (zh) * 2023-09-28 2024-01-16 中国石油大学(华东) 基于深度迁移学习的微地震事件检测方法及系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013089764A1 (en) * 2011-12-16 2013-06-20 Halliburton Energy Services, Inc. Methods of calibration transfer for a testing instrument
CN106405640A (zh) * 2016-08-26 2017-02-15 中国矿业大学(北京) 基于深度信念神经网络的微震信号到时自动拾取方法
CN106407649A (zh) * 2016-08-26 2017-02-15 中国矿业大学(北京) 基于时间递归神经网络的微震信号到时自动拾取方法
CN106646587A (zh) * 2016-12-29 2017-05-10 北京知觉科技有限公司 基于声振动信号的目标检测与识别方法及系统
CN109613610A (zh) * 2019-01-17 2019-04-12 中南大学 微震信号到时差的自动拾取方法
CN110133714A (zh) * 2019-06-05 2019-08-16 山东科技大学 一种基于深度学习的微震信号分类辨识方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105740840B (zh) * 2016-02-29 2017-02-22 中南大学 一种岩体破裂信号与爆破振动信号的非线性识别方法
CN105956526B (zh) * 2016-04-22 2020-01-24 山东科技大学 基于多尺度排列熵的低信噪比微震事件辨识方法
CN107505652B (zh) * 2017-07-26 2018-11-13 山东科技大学 一种基于能量分布特征的矿山微震信号辨识方法
CN107944469B (zh) * 2017-11-03 2021-12-21 桂林电子科技大学 一种基于重排st的低信噪比微震事件辨识方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013089764A1 (en) * 2011-12-16 2013-06-20 Halliburton Energy Services, Inc. Methods of calibration transfer for a testing instrument
CN106405640A (zh) * 2016-08-26 2017-02-15 中国矿业大学(北京) 基于深度信念神经网络的微震信号到时自动拾取方法
CN106407649A (zh) * 2016-08-26 2017-02-15 中国矿业大学(北京) 基于时间递归神经网络的微震信号到时自动拾取方法
CN106646587A (zh) * 2016-12-29 2017-05-10 北京知觉科技有限公司 基于声振动信号的目标检测与识别方法及系统
CN109613610A (zh) * 2019-01-17 2019-04-12 中南大学 微震信号到时差的自动拾取方法
CN110133714A (zh) * 2019-06-05 2019-08-16 山东科技大学 一种基于深度学习的微震信号分类辨识方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZHANG, XINGLI ET AL.: "Identifi Cation of Blasting Vibration and Coal-Rock Fracturing Microseismic Signals", APPLIED GEOPHYSICS, vol. 15, no. 2, 30 June 2018 (2018-06-30), pages 280 - 289, XP036586755, DOI: 20200119135618Y *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114563826A (zh) * 2022-01-25 2022-05-31 中国矿业大学 基于深度学习融合驱动的微震稀疏台网定位方法
CN114563826B (zh) * 2022-01-25 2023-03-03 中国矿业大学 基于深度学习融合驱动的微震稀疏台网定位方法

Also Published As

Publication number Publication date
CN110133714A (zh) 2019-08-16
CN110133714B (zh) 2020-11-10

Similar Documents

Publication Publication Date Title
WO2020220416A1 (zh) 一种基于深度学习的微震信号分类辨识方法
WO2022134868A1 (zh) 一种基于特征融合的道岔尖轨伤损状态监测方法
CN107290741B (zh) 基于加权联合距离时频变换的室内人体姿态识别方法
CN102289522B (zh) 一种对于文本智能分类的方法
CN109389180A (zh) 一款基于深度学习的电力设备图像识别方法及巡查机器人
CN108564005A (zh) 一种基于卷积神经网络的人体跌倒辨识方法
WO2019019565A1 (zh) 一种基于能量分布特征的矿山微震信号辨识方法
CN110767216A (zh) 一种基于pso算法的语音识别攻击防御方法
CN104808107A (zh) 一种xlpe电缆局部放电缺陷类型识别方法
CN106228979B (zh) 一种公共场所异常声音特征提取及识别方法
CN108830328B (zh) 融合空间知识的微震信号smote识别方法及监测系统
CN102201236A (zh) 一种高斯混合模型和量子神经网络联合的说话人识别方法
CN103743980A (zh) 基于pso优化的svm的电能质量扰动识别与分类方法
CN103093133B (zh) 面向ieee802.15.6的生物身份认证方法
CN106447086A (zh) 一种基于风电场数据预处理的风电功率组合预测方法
CN103558519A (zh) 一种gis局部放电超声波信号识别方法
CN112529341A (zh) 一种基于朴素贝叶斯算法的钻井漏失机率预测方法
CN104887263A (zh) 一种基于心音多维特征提取的身份识别算法及其系统
CN110632662A (zh) 一种利用DCNN-Inception网络自动识别微地震信号的算法
CN102411709A (zh) 一种虹膜分割识别方法
CN102982347A (zh) 一种基于kl距离的电能质量扰动分类方法
CN105160359A (zh) 一种基于超声导波的复杂结构损伤协同识别方法
CN103337248A (zh) 一种基于时间序列核聚类的机场噪声事件识别方法
CN113296148A (zh) 基于时域和小波域双通道卷积神经网络的微震识别方法
Wang et al. An automatic recognition method of microseismic signals based on s transformation and improved Gaussian mixture model

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19927247

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19927247

Country of ref document: EP

Kind code of ref document: A1