WO2021196367A1 - 基于三层数据集神经网络的性能地震动危险性分析方法 - Google Patents

基于三层数据集神经网络的性能地震动危险性分析方法 Download PDF

Info

Publication number
WO2021196367A1
WO2021196367A1 PCT/CN2020/091544 CN2020091544W WO2021196367A1 WO 2021196367 A1 WO2021196367 A1 WO 2021196367A1 CN 2020091544 W CN2020091544 W CN 2020091544W WO 2021196367 A1 WO2021196367 A1 WO 2021196367A1
Authority
WO
WIPO (PCT)
Prior art keywords
performance
magnitude
ground motion
data
neural network
Prior art date
Application number
PCT/CN2020/091544
Other languages
English (en)
French (fr)
Inventor
刘文锋
周正
李建峰
Original Assignee
青岛理工大学
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 青岛理工大学 filed Critical 青岛理工大学
Publication of WO2021196367A1 publication Critical patent/WO2021196367A1/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. for interpretation or for event detection
    • G01V1/30Analysis
    • G01V1/307Analysis for determining seismic attributes, e.g. amplitude, instantaneous phase or frequency, reflection strength or polarity
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01VGEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V1/00Seismology; Seismic or acoustic prospecting or detecting
    • G01V1/01Measuring or predicting earthquakes
    • 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. for interpretation or for event detection
    • G01V1/30Analysis
    • G01V1/306Analysis for determining physical properties of the subsurface, e.g. impedance, porosity or attenuation profiles
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10Complex mathematical operations
    • G06F17/18Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00Computing arrangements based on specific mathematical models
    • G06N7/01Probabilistic graphical models, e.g. probabilistic networks
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01VGEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V2210/00Details of seismic processing or analysis
    • G01V2210/30Noise handling
    • G01V2210/32Noise reduction
    • G01V2210/324Filtering
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01VGEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V2210/00Details of seismic processing or analysis
    • G01V2210/60Analysis
    • G01V2210/62Physical property of subsurface
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01VGEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
    • G01V2210/00Details of seismic processing or analysis
    • G01V2210/60Analysis
    • G01V2210/63Seismic attributes, e.g. amplitude, polarity, instant phase

Definitions

  • the invention relates to a seismic technical analysis method, in particular to a seismic hazard analysis method based on neural network.
  • the Earthquake "time, space, intensity” are the three core issues of seismic risk analysis technology.
  • Cornell proposed the probabilistic seismic hazard method.
  • the magnitude-frequency function relationship reflects the distribution demand of earthquake magnitude intensity; Poisson distribution reflects the probability of earthquake occurrence over time; the uniform distribution of earthquakes at various points in the potential source area reflects the earthquake Spatial distribution requirements.
  • the 1990 "China Earthquake Intensity Zoning Map” proposed a comprehensive probabilistic seismic hazard method, which divided the potential seismic source areas into secondary structures.
  • the 2015 China Earthquake Parameter Zoning Map The potential source area is structured as a three-level structure.
  • Seismic hazard analysis mainly includes deterministic hazard analysis and probabilistic seismic hazard analysis.
  • Deterministic risk analysis is based on considerations of seismic geology, seismic activity, and the largest earthquake in history, and gives a deterministic estimate of seismic events. Although the engineering meaning is clear and the expression is simple and intuitive, the uncertainty of the time, space, and intensity of future earthquakes cannot be objectively quantified and expressed.
  • Probabilistic seismic hazard analysis is a probabilistic analysis method in which the seismic parameters (intensity, acceleration, velocity, response spectrum, etc.) of a specific area exceed a given value during a certain design reference period in the future. The uncertainty of future earthquake time, space, and intensity is given A quantified probability index is given.
  • the seismic hazard analysis method has been applied to practical engineering, there are still many technical difficulties that need to be improved: the rationality of the b-value power law form of the GR relationship in a small area; the activity of the potential source area and the potential source are determined based on the activity of the seismic zone The objectivity of the regional spatial distribution function; the super-reality of probabilistic seismic hazard analysis on large-scale time series prediction results; the empirical statistical method has limitations in regional application; the random vibration method based on the source spectrum and the earthquake based on the empirical Green function Study methods, the uncertainty of complex seismic theoretical calculation models, etc. Therefore, big data directly based on actual seismic parameter records drive earthquake risk prediction and improve the effect of earthquake risk analysis, which is a new technology for earthquake risk analysis.
  • the training set + test set there are currently two types of structural models trained, one is training set + test set; the other is training set + verification set + test set.
  • the function of the second method validation set is to modify and improve the neural network model of the training set in the validation set. In this sense, the two methods are essentially the same. But when the amount of data is small, there will be problems that fail to pass in the test set. In reality, the most widely used method is to evaluate the generalization ability of the learning method through test errors. But this kind of evaluation depends on the test data set, because the test data set is limited, it is very likely that the result of the evaluation is unreliable. Therefore, how to change the training structure model and improve the robustness of neural network deep learning is also an important technology of machine learning.
  • the existing methods are based on the risk of earthquakes to determine seismic parameters, that is, considering the problem from the starting point of the earthquake.
  • the seismic design based on performance (behavior) is considered from the performance (behavior) of the structure in earthquake, that is, the problem is considered from the result of structural behavior caused by earthquake occurrence.
  • Existing documents and patents have proposed seismic design methods based on performance levels, but there is no risk analysis method for performance ground motions.
  • ground motion affects the results of seismic hazard analysis most significantly and sensitively.
  • the attenuation relationship of ground motion is restricted by the form of expression.
  • the technical effect of the present invention can overcome the above-mentioned shortcomings, and provides a method for analyzing the performance ground motion hazard based on a three-layer data set neural network.
  • This method uses the big data of seismic records, multi-parameter input to the neural network, and adopts a new neural network.
  • the training method predicts the attenuation relationship of ground motions, which improves the broadness and resilience of the attenuation relationship.
  • the present invention adopts the following technical solution: it includes the following steps:
  • the seismic record has not yet reached one million or tens of millions of big data.
  • existing neural network deep learning to predict the attenuation relationship of ground motions, there may be problems that fail to pass in the test set.
  • this patent The training structure model of training set + interval set + test set is used, which is different from the existing training structure model and improves the robustness of neural network deep learning.
  • this patent uses a deep neural network based on a three-layer data set to predict the attenuation relationship of ground motion, and adopts a probabilistic seismic hazard analysis method to predict the year of performance ground motion. The beyond probability and return period, based on performance ground motion and consistent probability, determine the magnitude and epicenter distance of the set earthquake.
  • Figure 1 is a schematic flow diagram of the method
  • Fig. 4 The set earthquake magnitude of the consistent beyond probability to determine the performance ground motion.
  • the performance ground motion hazard analysis method based on the three-layer data set neural network of the present invention includes the following steps (see Figure 1):
  • Ground motion data collection and data noise removal Collect a large number of ground motion records (including ground motion acceleration, velocity and displacement), perform baseline correction and band-pass filtering and other cleaning tasks, and unify the ground motion record data format.
  • Data feature parameter extraction and initialization processing extract the input feature parameter x of the neural network from the ground motion record, perform correlation test and data initialization.
  • the characteristic parameters are: performance ground motion logarithm LnY, magnitude M, focal distance R, focal depth H, fault type marker F, fault dip angle ⁇ 1 , fault strike ⁇ 2 , fault tendency ⁇ 3 , slip angle ⁇ 4 , record longitude ⁇ 0X , dimension ⁇ 0Y , venue V 30 .
  • MinMaxScaler algorithm To initialize the final feature x, take the MinMaxScaler algorithm as an example to scale all features to between 0 and 1.
  • the conversion function is:
  • x imin is the minimum value of the feature parameter x i in the data set, and the maximum value of the feature parameter x i in the x imax data set.
  • the data used in the training set, interval set, and test set are allocated in proportions of 60%, 20%, and 20% of the total data volume respectively.
  • train the multi-layer neural network MLP train the multi-layer neural network MLP
  • the training method is the conventional neural network training method.
  • MLP multi-layer neural network
  • Python by calling the MLPClassfier() function in the Scikit-Learn machine learning library, passing in the external parameters hidden_layer_sizes (hidden layer size), activation (activation function), slover (optimizer), alpha( Regularization parameters), max_iter (iteration depth), etc., to realize the architecture of neural network objects.
  • step (6) to verify the M * Yhmin and M * Yhmax of the ground motion Y that have reached the performance level according to the step length of 0.1. from M * Yhmin M * Yhmax to achieve the performance obtained seismic Y of the q M * Yhp, wherein M * Yhmin minimize vibration magnitude value Y corresponding to the performance, M * Yhmax seismic performance of the maximum value of magnitude corresponding to Y .
  • M imin M imax and M i are the minimum and maximum speed magnitude, Determine the k-th potential source area, exceeding Earthquake occurrence rate
  • S ⁇ is a ground motion parameter, which can be acceleration, velocity, response spectrum, etc.
  • v 0 is the annual average occurrence rate of earthquakes in the potential source area
  • the site has m potential seismic source areas, and the probability of exceeding the performance level within t years of the site:
  • This patent uses data-driven, deep neural network to construct the attenuation relationship of ground motion; adopts probabilistic seismic hazard analysis method to predict the annual exceedance probability and return period of performance ground motion; based on performance ground motion and consistent probability, determine the set earthquake Magnitude and epicenter distance.
  • the essence of the deep neural network of this patent is the attenuation relationship of ground motion.
  • Traditional statistical analysis will also obtain the attenuation relationship of ground motion, and give the error range.
  • the steps (7) and (8) of this patent the performance ground motion risk analysis method is carried out. Similarly, the exceedance probability, the return period and the corresponding set earthquake magnitude and epicenter distance of the performance ground motion are obtained, which will not be repeated.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Remote Sensing (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Environmental & Geological Engineering (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Geophysics (AREA)
  • General Life Sciences & Earth Sciences (AREA)
  • Geology (AREA)
  • Acoustics & Sound (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Algebra (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Operations Research (AREA)
  • Databases & Information Systems (AREA)
  • Geophysics And Detection Of Objects (AREA)
  • Business, Economics & Management (AREA)
  • Emergency Management (AREA)

Abstract

本发明涉及一种抗震技术分析方法,本发明的基于三层数据集神经网络的性能地震动危险性分析方法,包括以下步骤:(1)地震动数据采集与数据噪声去除;(2)数据特征参数提取与初始化处理;(3)训练集、区间集、测试集生成;(4)基于训练集,训练多层神经网络;(5)基于区间集,训练神经网络输出值,并计算输出值相对误差的均值和标准差;(6)基于测试集,训练神经网络确定输出值,基于区间置信度,计算震级区间;(7)概率地震危险性分析,确定性能地震的年超越概率和重现期;(8)基于性能地震动和一致概率,确定达到性能地震动的震级和震中距。采用新的神经网络训练方法,预测地震动衰减关系,提升了衰减关系适用的广泛性和韧性。

Description

基于三层数据集神经网络的性能地震动危险性分析方法 技术领域
本发明涉及一种抗震技术分析方法,尤其涉及一种基于神经网络的地震危险性分析方法。
背景技术
地震“时间、空间、强度”是地震危险性分析技术的三个核心问题。1968年Cornell提出了概率地震危险性方法,震级-频数函数关系反映了地震震级强度的分布需求;泊松分布反映了随时间变化地震发生的概率;潜在震源区内各点地震均匀分布反映了地震空间分布需求。为了修正潜在震源区均匀分布假定的近似性,1990年《中国地震烈度区划图》提出综合概率地震危险性方法,将潜在震源区按二级构造,2015年《中国地震动参数区划图》,将潜在震源区按三级构造。
地震危险性分析主要包括确定危险性分析和概率地震危险性分析。确定危险性分析基于地震地质、地震活动性、历史最大地震等的考量,给出地震事件的确定性估计。虽然工程意义明确、表达形式简单直观,但未来地震时间、空间、强度的不确定性,无法客观地量化表达。概率地震危险性分析是特定区域在未来一定设计基准期内地震参数(烈度、加速度、速度、反应谱等)超过某一给定值的概率分析方法,未来地震时间、空间、强度不确定性给出了量化的概率指标。主要包括三种方法:一是经验统计法;二是基于震源谱的随机振动法;三是基于经验格林函数的地震学方法。地震危险性分析方法虽已经应用于实际工程,但仍然存在很多技术难点需要改进:小区域的G-R关系b值幂律形式的合理性;基于地震带的活动性确定潜在震源区活动性和潜在震源区空间分布函数的客观性;概率地震危险性分析在大尺度时间序列上预测结果的超真实性;经验统计法存在区域适用的局限性;基于震源谱的随机振动法和基于经验格林函数的地震学方法,复杂地震理论计算模型的不确定性等。因此,直接基于实际地震参数记录的大数据驱动地震危险性预测,提升地震危险性分析的效果, 是地震危险性分析的一项新技术。
在基于神经网络的深度学习方面,目前训练的结构模型有二种,一是训练集+测试集;二是训练集+验证集+测试集。第二种方法验证集的作用是在验证集里修正、改进训练集的神经网络模型,从这个意义上讲,这二种方法本质是一致的。但当数据量较少,在测试集里会出现无法通过的难题。现实中采用最多的方法是通过测试误差来评价学习方法的泛化能力。但这种评价是依赖于测试数据集的,因为测试数据集是有限的,很有可能由此得到的评价结果是不可靠的。因此,如何改变训练结构模型,提升神经网络深度学习的鲁棒性,也是机器学习的一项重要技术。
在地震危险性分析方法方面,目前已有的方法是从地震发生的危险性出发确定地震参数的,即从地震发生的始点考虑问题的。而基于性能(性态)的抗震设计是从结构地震中的性能(性态)考虑的,即从地震发生导致的结构行为结果考虑问题。现有文献和专利都已有提出了基于性能水准的抗震设计方法,但是,没有性能地震动的危险性分析方法。
地震动衰减关系影响地震危险性分析结果最为显著和敏感。目前地震动衰减关系受表达形式的制约,仅有地震动(含烈度)、震级和震中距三参数,难以全面体现地震复杂性,仅适用区域地震动衰减模拟。
发明内容
本发明的技术效果能够克服上述缺陷,提供一种基于三层数据集神经网络的性能地震动危险性分析方法,本方法利用地震记录的大数据,多参数输入神经网络,并采用新的神经网络训练方法,预测地震动衰减关系,提升了衰减关系适用的广泛性和韧性。
为实现上述目的,本发明采用如下技术方案:其包括以下步骤:
(1)地震动数据采集与数据噪声去除;
(2)数据特征参数提取与初始化处理;
(3)训练集、区间集、测试集生成;
(4)基于训练集,训练多层神经网络;
(5)基于区间集,训练神经网络输出值,并计算输出值相对误差的均值和标准差;
(6)基于测试集,训练神经网络确定输出值,基于区间置信度,计算震级区间;
(7)概率地震危险性分析,确定性能地震的年超越概率和重现期;
(8)基于性能地震动和一致概率,确定达到性能地震动的震级和震中距。
目前地震记录尚未达到百万级、千万级的大数据,采用已有神经网络深度学习,预测地震动衰减关系,在测试集里可能会出现无法通过的难题,为了解决这一不足,本专利采用训练集+区间集+测试集的训练结构模型,区别于现有的训练结构模型,提升了神经网络深度学习的鲁棒性。目前尚未有基于性能地震动的危险性分析方法,为了解决这一不足,本专利采用基于三层数据集深度神经网络预测地震动衰减关系,采用概率地震危险性分析方法,预测性能地震动的年超越概率和重现期,基于性能地震动和一致概率,确定设定地震的震级和震中距。
附图说明
图1为本方法的流程示意图;
图2性能地震动-超越概率的地震危险性曲线;
图3性能地震动超越概率的确定方法;
图4一致超越概率确定性能地震动的设定地震震级。
具体实施方式
本发明的基于三层数据集神经网络的性能地震动危险性分析方法,包括以下步骤(见图1所示):
1、地震动数据采集与数据噪声去除:收集大量的地面运动记录(包括地面运动的加速度、速度以及位移),进行基线校正和带通滤波等清洗工作,统一地面运动记录数据格式。
2、数据特征参数提取与初始化处理:从地面运动记录中提取神经网络的输入特征参数x,进行相关性检验和数据初始化。
特征参数分别是:性能地震动对数LnY、震级M、震源距R、震源深度H、断层类型标记F、断层倾角θ 1、断层走向θ 2、断层倾向θ 3、滑动角θ 4,记录经度θ 0X、维度θ 0Y、场地V 30
进行特征参数x的相关性检验,计算两两特征的信息增益Gain(A,B)=H(A)-H(A|B),A为特征参数,Gain(A,B)为信息增益在得知特征参数A一定的情况下,逾期概率B不确定性的减少程度,H(A/B)为特征参数x被固定时的条件熵。保证10%以下的相关程度,若大于10%,考虑去除相应特征,形成新的数据表,最终特征参数为x。
对最终特征x初始化处理,以MinMaxScaler算法为例,将所有特征缩放到0和1之间,转换函数为:
Figure PCTCN2020091544-appb-000001
式中,x imin为数据集中特征参数x i的最小值,x imax数据集中特征参数x i的最大值。
将震级M作为数据列表的最后一列,按震级M由小到大的顺序排列,建立新的输入列表。
3、训练集、区间集、测试集采用的数据分别按总数据量的60%、20%、20%比例分配。
4、在训练集内,训练多层神经网络MLP,训练方法为常规的神经网络训练方法。
①按照输入层、隐藏层和输出层,按照前馈BP算法,架构多层神经网络MLP。以Python开发的MLP神经网络为例,通过调用Scikit-Learn机器学习库中的MLPClassfier()函数,传入外部参数hidden_layer_sizes(隐藏层尺寸)、activation(激活函数)、slover(优化器)、alpha(正则化参数)、max_iter(迭代深度)等,实现对神经网络对象的架构。确定自定义的上述参数的全部或部分,分别进行取值列表的给定,如activation:{‘identity’,‘logistic’,‘tanh’,‘relu’},构建交叉验证的网格搜索对象,将训练集转换为Python Numpy 运算库中的ndarray对象,输入至网格搜索对象,确定模型最佳参数组合。
②依据上一步骤给出最佳参数组合,定义多层神经网络MLP,将训练集标签的ndarray对象输入网络,计算当前神经网络输出值与该条数据标签的差值,基于BP算法,将误差函数后反馈至每层每一个节点上计算梯度,逐步更新每个节点的权重与偏置值,训练多层神经网络MLP。
5、在区间集内,基于训练神经网络输出震级,计算区间集输出值相对误差的均值和标准差。
将区间集的数据输入至步骤(4)训练好的神经网络中,获取每一条数据的输出
Figure PCTCN2020091544-appb-000002
与真实数据M h进行对比,计算每条输出数据的相对误差
Figure PCTCN2020091544-appb-000003
计算区间集输出值相对误差的均值μ和标准差ε。
6、在测试集内,基于训练神经网络输出震级
Figure PCTCN2020091544-appb-000004
根据区间置信度计算震级区间。
将测试集的数据输入至步骤(4)训练好的神经网络中,获取每一条数据的输出
Figure PCTCN2020091544-appb-000005
并基于步骤(5)得出的相对误差的均值μ和标准差ε,计算其与真实标签M h的相对误差ΔM,将M h表示为
Figure PCTCN2020091544-appb-000006
设定置信度95%,以验证集中95%M h落在区间
Figure PCTCN2020091544-appb-000007
为原则,计算验证集中满足置信度要求的最小l值,则
Figure PCTCN2020091544-appb-000008
满足置信度要求的震级区间,
Figure PCTCN2020091544-appb-000009
Figure PCTCN2020091544-appb-000010
分别是震级区间的最小震级和最大震级。
7、概率地震危险性分析,确定性能地震的年超越概率和重现期。
①确定场地建筑物达到性能地震动Y,具体方法可参见本发明人前期已经授权或公开的发明专利。
②按照《中国地震动参数区划图GB18306-2015》三级划分原则,确定潜在震源区。
③确定潜在震源区内起算震级M 0、震级上限M uz,从起算震级M 0到震级上限M uz,确定n个震级档M j(j=1,2,……,n)。
④依据G-R震级频度关系公式lgN=a-bM,根据实际地震震级M、频度N进行统计,确定a和b值,令中间参数β=b ln 10,计算潜在震源区起算震级及以上的年发生率v 0
⑤根据确定的达到性能水准的地震动Y和潜在震源区的震源距R,采用步骤(6)验证通过达到性能地震动Y的M * Yhmin和M * Yhmax,按照震级档为0.1的步长,从M * Yhmin到M * Yhmax可获得达到性能地震动Y的q个M * Yhp,其中M * Yhmin为性能地震动Y对应的最小震级值,M * Yhmax为性能地震动Y对应的最大震级值。
⑥任意一个潜在震源区k内,M imin和M imax分别是M i震级档内最小值和最大值,
Figure PCTCN2020091544-appb-000011
确定第k个潜在震源区内,超过
Figure PCTCN2020091544-appb-000012
的地震发生率
Figure PCTCN2020091544-appb-000013
式中,S α为地震动参数,可以是加速度、速度、反应谱等;v 0为潜在震源区的地震年平均发生率;
Figure PCTCN2020091544-appb-000014
为潜在震源区k、震级分档M j的空间分布系数,根据地震活动特征因子、区划图发生率因子、地震构造条件因子、地震活动度因子、网格活动性因子、大震发生率因子、中长期危险性因子、离逝时间因子等确定。
⑦潜在震源区k在t年内场地发生达到性能水准的地震超越概率为:
P tk(S a≥Y|M,R)=1-e -vt    (2)
⑧场地有m个潜在震源区,场地t年内发生达到性能水准的超越概率:
Figure PCTCN2020091544-appb-000015
⑨场地t年内发生达到性能水准的重现期:
Figure PCTCN2020091544-appb-000016
绘制q条性能地震动-超越概率的地震危险性曲线,见图2。
8、基于性能地震动和一致概率,确定设定震级和震中距
①挑选任一条地震动—超越概率的地震危险性曲线Yhp,根据性能地震动Y i,确定相应的超越概率,见图3;依据一致超越概率,确定达到性能地震动Y i的设定地震震级M i,见图4。
②将m个潜在震源区划分为等间距微小震源区,场点到m个潜在震源区的微小震源区会有相应数量的震中距R i
③取历遍R i的方法,将Y i、R i等参数带回已训练好的深度神经网络,计算
Figure PCTCN2020091544-appb-000017
筛选具有一致超越概率,且达到性能地震动Y i的设定地震震中距R i
其中,R i的筛选规则:
1)任意潜在震源区内M imin≤M i≤M imax
2)
Figure PCTCN2020091544-appb-000018
3)R i在相应的震源区内;
确定了达到性能地震动Y i,且一致概率的设定地震的震级M i和震中距R i
本专利采用数据驱动,深度神经网络构建地震动的衰减关系;采用概率地震危险性分析方法,预测性能地震动的年超越概率和重现期;基于性能地震动和一致概率,确定设定地震的震级和震中距。
需要指出,本专利的深度神经网络的本质是地震动的衰减关系,传统采用统计分析也会得到地震动的衰减关系,且给出误差范围。按照本专利步骤(7)和步骤(8)进行性能地震动危险性分析方法,同样,得到达到性能地震动的超越概率、重现期和相应设定地震的震级、震中距,不赘述。

Claims (10)

  1. 一种基于三层数据集神经网络的性能地震动危险性分析方法,其特征在于包括以下步骤:
    (1)地震动数据采集与数据噪声去除;
    (2)数据特征参数提取与初始化处理;
    (3)训练集、区间集、测试集生成;
    (4)基于训练集,训练多层神经网络;
    (5)基于区间集,训练神经网络输出值,并计算输出值相对误差的均值和标准差;
    (6)基于测试集,训练神经网络确定输出值,基于区间置信度,计算震级区间;
    (7)概率地震危险性分析,确定性能地震的年超越概率和重现期;
    (8)基于性能地震动和一致概率,确定达到性能地震动的震级和震中距。
  2. 根据权利要求1所述的分析方法,其特征在于,所述的步骤(1)中地震动数据包括地面运动的加速度、速度以及位移。
  3. 根据权利要求1或2所述的分析方法,其特征在于,所述的步骤(2)中特征参数分别为:性能地震动对数LnY、震级M、震源距R、震源深度H、断层类型标记F、断层倾角θ 1、断层走向θ 2、断层倾向θ 3、滑动角θ 4、记录经度θ 0X、维度θ 0Y、场地V 30
  4. 根据权利要求3所述的分析方法,其特征在于,对上述特征参数进行相关性检验,计算两两特征的信息增益Gain(A,B)=H(A)-H(A|B),其中A为特征参数,Gain(A,B)为信息增益在得知特征参数A的条件下,逾期概率B不确定性的减少程度,H(A/B)为特征参数被固定时的条件熵;保证10%以下的相关程度,若大于10%,去除相应特征参数,形成新数据表,最终特征参数为x。
  5. 根据权利要求4所述的分析方法,其特征在于,对最终特征参数x初始化处理,将所有最终特征参数x缩放到0和1之间,转换函数为:
    Figure PCTCN2020091544-appb-100001
    式中,x i min为特征参数x i的最小值,x i max为特征参数x i的最大值;震级M作为数据列表的最后一列,将初始化处理后的数据按震级M由小到大的顺序排列,建立输入列表。
  6. 根据权利要求5所述的分析方法,其特征在于,步骤(3)中,训练集、区间集、测试集采用的数据分别按总数据量的60%、20%、20%比例分配。
  7. 根据权利要求6所述的分析方法,其特征在于,步骤(5)中,将区间集的数据输入至步骤(4)训练好的神经网络中,获取每一条数据的输出
    Figure PCTCN2020091544-appb-100002
    与真实数据M h进行对比,计算每条输出数据的相对误差
    Figure PCTCN2020091544-appb-100003
    计算区间集输出值相对误差的均值μ和标准差ε。
  8. 根据权利要求7所述的分析方法,其特征在于,步骤(6)中,将测试集的数据输入至步骤(4)训练好的神经网络中,获取每一条数据的输出
    Figure PCTCN2020091544-appb-100004
    并基于步骤(5)得出的相对误差的均值μ和标准差ε,计算其与真实数据M h的相对误差ΔM,将M h表示为
    Figure PCTCN2020091544-appb-100005
    设定置信度95%,以验证集中95%M h落在区间
    Figure PCTCN2020091544-appb-100006
    为原则,计算测试集中满足置信度要求的最小l值,则
    Figure PCTCN2020091544-appb-100007
    满足置信度要求的震级区间,令
    Figure PCTCN2020091544-appb-100008
    Figure PCTCN2020091544-appb-100009
    其分别是震级区间的最小震级和最大震级。
  9. 根据权利要求8所述的分析方法,其特征在于,步骤(7)包括以下步骤:
    (7.1)确定场地建筑物达到性能地震动Y;
    (7.2)按照三级划分原则,确定潜在震源区;
    (7.3)确定潜在震源区内起算震级M 0、震级上限M uz,从起算震级M 0到震级上限M uz,确定n个震级档M j(j=1,2,……,n);
    (7.4)依据G-R震级频度关系公式lgN=a-bM,根据实际地震震级M、频度N进行统计,确定a和b值,令中间参数β=bln 10,统计潜在震源区起算震级及以上的年发生率v 0
    (7.5)根据确定的达到性能水准的地震动Y和潜在震源区的震源距R,采用步骤(6)验证通过达到性能地震动Y的M * Yhmin和M * Yhmax,按照震级档为0.1的步长,从M * Yhmin到M * Yhmax可获得达到性能地震动Y的q个M * Yhp,其中M * Yhmin为性能地震动Y对应的最小震级值,M * Yhmax为性能地震动Y对应的最大震级值;
    (7.6)任意一个潜在震源区k内,M i min和M i max分别是震级M i档内最小值和最大值,
    Figure PCTCN2020091544-appb-100010
    确定第k个潜在震源区内,超过性能地震动Y的地震发生率
    Figure PCTCN2020091544-appb-100011
    式中,S α为地震动参数,是加速度、速度、反应谱;v 0为潜在震源区的地震年平均发生率;
    Figure PCTCN2020091544-appb-100012
    为潜在震源区k、震级分档M j的空间分布系数,根据地震活动特征因子、区划图发生率因子、地震构造条件因子、地震活动度因子、网格活动性因子、大震发生率因子、中长期危险性因子、离逝时间因子确定;
    (7.7)潜在震源区k在t年内场地发生达到性能水准的地震超越概率为:
    P tk(S a≥Y|M,R)=1-e -vt
    (7.8)场地有m个潜在震源区,场地t年内发生达到性能水准的超越概率:
    Figure PCTCN2020091544-appb-100013
    (7.9)场地t年内发生达到性能水准的重现期:
    Figure PCTCN2020091544-appb-100014
    绘制q条性能地震动—超越概率的地震危险性曲线。
  10. 根据权利要求9所述的分析方法,其特征在于,步骤(8)包括以下步骤:
    (8.1)挑选任一条地震动—超越概率的地震危险性曲线,根据任意一个性能 地震动Y i,确定相应的超越概率;依据一致超越概率,确定达到性能地震动Y i的设定地震震级M i
    (8.2)将m个潜在震源区划分为等间距微小震源区,场点到m个潜在震源区的微小震源区会有相应数量的震中距R i
    (8.3)采取历遍R i的方法,将Y i、R i等参数带回已训练好的深度神经网络,计算
    Figure PCTCN2020091544-appb-100015
    筛选具有一致超越概率,且达到性能地震动Y i的设定地震震中距R i
    其中,R i的筛选规则:
    1)任意潜在震源区内M i min≤M i≤M imax
    2)
    Figure PCTCN2020091544-appb-100016
    3)R i在相应的震源区内;
    确定了达到性能地震动Y i,且一致概率的设定地震的震级M i和震中距R i
PCT/CN2020/091544 2020-03-30 2020-05-21 基于三层数据集神经网络的性能地震动危险性分析方法 WO2021196367A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010236208.9 2020-03-30
CN202010236208.9A CN111458748B (zh) 2020-03-30 2020-03-30 基于三层数据集神经网络的性能地震动危险性分析方法

Publications (1)

Publication Number Publication Date
WO2021196367A1 true WO2021196367A1 (zh) 2021-10-07

Family

ID=71685019

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/091544 WO2021196367A1 (zh) 2020-03-30 2020-05-21 基于三层数据集神经网络的性能地震动危险性分析方法

Country Status (5)

Country Link
US (1) US11215719B2 (zh)
EP (1) EP3889652A1 (zh)
JP (1) JP6948740B2 (zh)
CN (1) CN111458748B (zh)
WO (1) WO2021196367A1 (zh)

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110751399B (zh) * 2019-10-22 2023-02-24 中国地震局地球物理研究所 地震危险性并行分析方法、装置及设备
CN113009553A (zh) * 2021-03-02 2021-06-22 电子科技大学 基于NGBoost和SHAP值的可解释地震动参数概率密度分布预测方法
CN112946318A (zh) * 2021-03-24 2021-06-11 苏州康旺聚贤智能科技有限公司 一种加速度传感器的标定算法
CN113268852B (zh) * 2021-04-14 2022-02-22 西南交通大学 一种基于蒙特卡洛模拟的地震滑坡概率危险性分析方法
CN113379105B (zh) * 2021-05-21 2022-05-31 浙江大学 一种基于人工智能的场地地震液化灾害快速评估方法
US11200788B1 (en) * 2021-06-28 2021-12-14 1st Street Foundation, Inc. Systems and methods for forecasting and assessing hazard-resultant effects
CN113484913B (zh) * 2021-08-16 2023-06-16 成都理工大学 一种多粒度特征融合卷积神经网络的地震资料去噪方法
CN114329974B (zh) * 2021-12-29 2023-07-18 武汉大学 基于蒙特卡洛模拟的城市供水管网地震损伤评估方法
CN114693066B (zh) * 2022-02-28 2024-03-15 哈尔滨工业大学(深圳) 地震危险性分析方法、装置、设备及存储介质
CN114675325A (zh) * 2022-03-13 2022-06-28 中国地震台网中心 基于Mindlin解的地震地表破裂永久位移的估计方法
CN114779329B (zh) * 2022-04-24 2023-05-23 福州大学 一种近断层海域地震动反应谱标定方法
CN115343751A (zh) * 2022-08-04 2022-11-15 中国地震局地震预测研究所 一种活动地块边界带断层震级确定方法及系统
CN115291281B (zh) * 2022-09-30 2022-12-20 中国科学院地质与地球物理研究所 基于深度学习的实时微地震震级计算方法及装置
CN115826039B (zh) * 2022-11-09 2023-07-14 中国地质科学院 一种时间切片分类模型训练方法、系统及应用方法、系统
CN116256803B (zh) * 2023-03-09 2024-01-19 中国矿业大学 一种融合开采信息与地质信息的煤矿微震区域定位方法
CN116956046B (zh) * 2023-09-07 2023-12-29 西南交通大学 一种基于循环神经网络的地震滑坡危险性分析方法及装置
CN118090107A (zh) * 2024-04-24 2024-05-28 华侨大学 一种考虑地震动影响的多结构地震反应智能预测方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101699317A (zh) * 2009-11-09 2010-04-28 东南大学 利用地震参量进行短期地震预测的神经网络方法
CN103926621A (zh) * 2014-05-06 2014-07-16 中国地震局工程力学研究所 两阶段残差分析建立地震动衰减关系的方法
KR20150035633A (ko) * 2013-09-27 2015-04-07 한국전력공사 지진측정장치 및 지진측정방법
KR20160114804A (ko) * 2015-03-24 2016-10-06 현대자동차주식회사 해석성능 평가 방법
CN109799533A (zh) * 2018-12-28 2019-05-24 中国石油化工股份有限公司 一种基于双向循环神经网络的储层预测方法

Family Cites Families (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5490062A (en) * 1994-05-11 1996-02-06 The Regents Of The University Of California Real-time neural network earthquake profile predictor
US6928367B2 (en) * 2002-09-27 2005-08-09 Veritas Dgc Inc. Reservoir fracture characterization
US7280920B1 (en) * 2005-06-29 2007-10-09 Whiteside Lowell S Method and apparatus for forecasting earthquakes and making informed risk management decisions
CN102200588B (zh) * 2011-03-22 2012-08-22 成都理工大学 地震资料的波形相似性体曲率分析方法
CN102787676B (zh) * 2011-05-20 2014-07-16 青岛理工大学 工程结构多目标性能化抗震设计方法
CN105005079B (zh) * 2015-07-14 2017-12-01 北京博达瑞恒科技有限公司 一种井曲线反演方法
CN106019359A (zh) * 2016-05-17 2016-10-12 浪潮集团有限公司 一种基于神经网络的地震预测系统
CN106154309A (zh) * 2016-06-29 2016-11-23 合肥民众亿兴软件开发有限公司 一种基于网络的地震12月内短期预测方法
CN109635461B (zh) * 2018-12-18 2022-04-29 中国铁建重工集团股份有限公司 一种应用随钻参数来自动识别围岩级别的方法和系统
JP7180946B2 (ja) * 2019-01-11 2022-11-30 株式会社竹中工務店 地震情報処理装置
JP2020193935A (ja) * 2019-05-30 2020-12-03 新医療リアルワールドデータ研究機構株式会社 評価システム及び評価プログラム
JP7258682B2 (ja) * 2019-07-16 2023-04-17 株式会社東芝 プラント評価システム及びプラント評価方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101699317A (zh) * 2009-11-09 2010-04-28 东南大学 利用地震参量进行短期地震预测的神经网络方法
KR20150035633A (ko) * 2013-09-27 2015-04-07 한국전력공사 지진측정장치 및 지진측정방법
CN103926621A (zh) * 2014-05-06 2014-07-16 中国地震局工程力学研究所 两阶段残差分析建立地震动衰减关系的方法
KR20160114804A (ko) * 2015-03-24 2016-10-06 현대자동차주식회사 해석성능 평가 방법
CN109799533A (zh) * 2018-12-28 2019-05-24 中国石油化工股份有限公司 一种基于双向循环神经网络的储层预测方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
LUO, QIFENG: "PROBABILITY-CONSISTENT SCENARIO EARTHQUAKE AND ITS DETERMINATION", EARTHQUAKE ENGINEERING AND ENGINEERING VIBRATION, vol. 16, no. 3, 30 September 1996 (1996-09-30), ISSN: 1000-1301 *
PANAKKAT, ASHIF ET AL.: "Neural Network Models for Earthquake Magnitude Prediction Using Multiple Seismicity Indicators", INTERNATIONAL JOURNAL OF NEURAL SYSTEMS, vol. 17, no. 1, 28 February 2007 (2007-02-28), ISSN: 0129-0657 *

Also Published As

Publication number Publication date
EP3889652A1 (en) 2021-10-06
CN111458748A (zh) 2020-07-28
US20210302603A1 (en) 2021-09-30
US11215719B2 (en) 2022-01-04
CN111458748B (zh) 2021-03-05
JP6948740B2 (ja) 2021-10-13
JP2021162591A (ja) 2021-10-11

Similar Documents

Publication Publication Date Title
WO2021196367A1 (zh) 基于三层数据集神经网络的性能地震动危险性分析方法
CN107529651B (zh) 一种基于深度学习的城市交通客流预测方法和设备
Lin et al. Dynamics‐based cross‐domain structural damage detection through deep transfer learning
Wu et al. A hybrid support vector regression approach for rainfall forecasting using particle swarm optimization and projection pursuit technology
CN105653518A (zh) 一种基于微博数据的特定群体发现及扩充方法
CN115564114B (zh) 一种基于图神经网络的空域碳排放短期预测方法及系统
CN109635010B (zh) 一种用户特征及特征因子抽取、查询方法和系统
CN111079999A (zh) 一种基于cnn和svm的洪水灾害易发性预测方法
CN116304546A (zh) 基于声音信号的供热系统热力站故障诊断方法及系统
Sajedi et al. Intensity-based feature selection for near real-time damage diagnosis of building structures
CN114912798A (zh) 一种基于随机森林和震害大数据的地震损失评估系统
CN113312596B (zh) 一种基于深度学习和异步轨迹数据的用户身份识别方法
CN112201036B (zh) 一种基于Inception-CNN的城市快速路行程速度短时预测方法
Wang et al. Improvement of the seismic resilience of regional buildings: A multi-objective prediction model for earthquake early warning
Zhang et al. A real‐time seismic damage prediction framework based on machine learning for earthquake early warning
CN112183847B (zh) 基于QRNN改进Stacking算法的列车行驶风速概率预测方法
CN115858763A (zh) 多模态数据融合的城管事件分析方法及其应用
CN113240161B (zh) 净现值预测模型建立方法、装置、存储介质及电子设备
Woolford et al. Convergent data sharpening for the identification and tracking of spatial temporal centers of lightning activity
CN115391523A (zh) 风电场多源异构数据处理方法及装置
KR102211842B1 (ko) 상관관계 점수 행렬 생성 알고리즘을 이용한 인공 신경망 기반의 장소 인식 장치 및 방법
CN111104298A (zh) 一种基于lstm的电网服务器运行状态预测装置
Ramírez-Rojas et al. Time series analysis in seismology: practical applications
CN113205685B (zh) 基于全局-局部残差组合模型的短时交通流预测方法
CN117911800B (zh) 一种多尺度流域人居环境分类管理系统

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: 20928599

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: 20928599

Country of ref document: EP

Kind code of ref document: A1