Using Ambient Vibration Measurements for Risk Assessment at Urban - PowerPoint PPT Presentation

Using Ambient Vibration Measurements for Risk Assessment at Urban Scale : from Numerical Proof of Concept to a Case Study in Beirut (Lebanon) Christelle Salameh 1 , Pierre-Yves Bard 1 , Bertrand Guillier 1 , Jacques Harb 2 , Cécile Cornou 1 and Michelle Almakari 1 1 - ISTerre, University Grenoble-Alpes, France 2 - Notre-Dame University-Louaizé, Lebanon ESG5, Taipei, Taiwan, 15/08/2016

Outline Introduction ? Use of frequency information in large scale damage assessment Conceptual framework and comprehensive numerical simulation SDOF elastoplastic oscillators on multilayered 1D (linear) soil profiles ANN analysis Robustness and field applicability ? easily available site amplification proxy NL soil behavior (MDOF) Sense-check : example Application to Beirut City (Lebanon) Conclusions, caveats and further steps ESG5, Taipei, Taiwan, 15/08/2016

Introductory words • Many examples of larger damage due to coincidence between soil and building frequencies  Mexico 1985, Kathmandu 2015, …  Obvious for linear systems, not so much for NL systems • Building specific studies (detailed information)  best GM proxy = SA (f 0 ) or ASA ([0.6 – 1] f 0 ) – (Perrault & Gueguen, 2015; De Biasio, 2015) • ? Urban scale (or larger) : Damage / Risk maps  Microzonation, site effects : rather quantitative assessment • Site characterization : Geology, VS30, f0 (H/V, … ) • Site amplification Lack of consistency  Building surveys : most often only qualitative hazard / vulnerabilty • Gross typology ESG5, Taipei, Taiwan, 15/08/2016

Damage Estimation Bullding scale : Large scale (urban) ? Mechanical methods Macroseismic approach (Hazus, RISK-UE) Damage Intensity (or PGA) Vulnerability Curves for various Purple: Seismic demand building typologies Black: Building Resistance Large scale = qualitative Individual scale = quantitative ! Challenging ! (spatial variability) Estimate damages quantitatively on a large scale with more mechanical input including spectral coincidence ESG5, Taipei, Taiwan, 15/08/2016

Outline Introduction ? Use of frequency information in large scale damage assessment Conceptual framework and comprehensive numerical simulation Elastoplastic SDOF oscillator on a single layer Extension through comprehensive numerical simulation SDOF elastoplastic oscillators on multilayered 1D soil profiles Neural network analysis ESG5, Taipei, Taiwan, 15/08/2016

Soil response 3f0 5f0 f0 Soil Reflexion/Transmission properties Bedrock ESG5, Taipei, Taiwan, 15/08/2016

Oscillator response : weak input (linear response) Linear branch of the oscillator d max ESG5, Taipei, Taiwan, 15/08/2016

Oscillator response : strong input (non linear domain) Elastoplastic Behavior of the oscillator dmax ESG5, Taipei, Taiwan, 15/08/2016

Conceptual framework : a simple illustrative example V Plastic Linear Vy 54 SDOF Elastoplastic oscillators 9 x f structure = 1 → 9 Hz 6 x dy= 0.005 → 0.05 (m) du dy d H(m), 36 Linear single-layer sites (No SSI) : ρ , 4 x Velocity Contrast= 2 → 8 Vs (m/s) 9 x f soil = 1 → 9 (Hz) Qs 60 synthetic Input motion (Sabetta Vs=1000 m/s Bedrock and Pugliese 1996 : nonstationary): 5 x Magnitude= 3 → 7 4 x Distance= 5 → 100 (km) PGA= 0.02- 8.6 (m/s 2 ) 116 640 Models ESG5, Taipei, Taiwan, 15/08/2016

Comparison soil / rock dmax soil / dmax rock dmax soil On soil On outcropping bedrock dmax rock Bedrock Bedrock Same input motion ESG5, Taipei, Taiwan, 15/08/2016

Statistical analysis for the simple case Low PGA / linear response High PGA V V F struct = 9 Hz f struct =1 Hz dmax(rock) dmax(soil) dmax(soil) Vy Vy dmax(rock) fsoil=3 Hz H(m), ρ , Vs(m/s) d du dy d du dy Vs=1000 m/s Bedrock Non-linear C=2 behavior of f struct /f soil = 1 the dmax(soil) /dmax(rock C=4 54 oscillators, structure 9 site frequencies (1-9 Hz) f struct /f soil = 3 ) C=6 1 velocity contrast f struct /f soil =0.33 C=8 f struct /f soil = 3 0.1 1.0 10 1.0 10 f struct / f soil F struct / f soil ESG5, Taipei, Taiwan, 15/08/2016

Realistic (less unrealistic … ) case: real soil profiles Risk-UE typologies : 141 SDOF elastoplastic oscillators f struct, dy, du classified into 5 typology classes: 1 = Masonry; 2 = Non-designed RC; 3 = RC Low ductility; 4= RC Medium ductility; 5) RC High ductility 887 multilayered linear soils (still no SSI): 614 KiKnet + 251 USA + 22 Europe f soil = 0.2-39 Hz Vs30= 111 -2100 m/s depth= 7-1575 m 60 synthetic Input Signal: Magnitude= 3 → 7, Distance = 5 → 100 km PGA= 0.02- 8.6 m/s 2 ~7.5 MILLION MODELS!!! ESG5, Taipei, Taiwan, 15/08/2016

Oscillator characteristics Distribution dy – T Distribution du/dy – T struct struct Yield displacement dy Ductility du/dy Period (s) Period (s) ESG5, Taipei, Taiwan, 15/08/2016

Distribution of site characteristics Sediment Thickness Fundamental frequency 150 80 70 60 100 50 Frequency Frequency 40 30 50 20 10 0 0 1 3.16 10 31.62 100 316.22 1000 3162.27 0.1 0.31 1 3.16 10 31.62 100 1 10 100 1000 0.1 1. 10. 100 . f 0 (Hz) Depth (m) Thickness (m) f 0 (Hz) 100 80 Velocity contrast 90 70 V S30 80 60 70 50 Frequency 60 Frequency 50 40 40 30 30 20 20 10 10 0 0 100 158.48 251.18 398.10 630.95 1000 1584.89 2511.88 1 1.58 2.51 3.98 6.30 10 15.84 25.11 39.81 63.09 100 300. 1000. 3000. 1 3. 10. V s30 (m/s) Cv V S30 (m/s) 30. C V = V max / V min ESG5, Taipei, Taiwan, 15/08/2016

Classical statistical analysis? ~7.5 MILLION MODELS!!! Input parameters "SSS" Soil 1 output Structure Signal Damage increment Artificial Neural Network ANN ESG5, Taipei, Taiwan, 15/08/2016

Neural network approach Goal  to look for statistical relationships between pre-selected input and output variables, without any a priori on the functional forms Principle (ML perceptron)  Combination through weigthed sums ("synaptic weights") and "activation functions"  Introduction of a "hidden layer" Implementation  Selection of input and output parameters  Learning, validation and test sets : 70%, 15%, 15%  Optimizing • Number of neurons in the hidden layer • Activation functions • Training algorithm ESG5, Taipei, Taiwan, 15/08/2016

Neural Network : principle Hidden layer Activation Inputs ∑ w i1 functions ∑ w h j Computed Target Output Error Performance parameters RMSE w ij , w h j = synaptic weights R 2 Initially random, optimized ESG5, Taipei, Taiwan, 15/08/2016 through training

Neural Network: Our case study Input layer Output layer Hidden Layer PGA Target fstruct/ fsoil Damage increment Velocity One ANN for each building contrast typology class (1-5) ESG5, Taipei, Taiwan, 15/08/2016

Damage level index V Damage index DI D4 Complete 4 Vy 1.5dy 0.5(dy+du) D3 extensive du 3 0.7dy D2 moderate 2 D1 Slight 1 D0 No damage 0 du dy d Risk-UE project : correspondence between EMS98 damage states and maximum structural displacement (Lagomarsino and Giovinazzi, 2006) ESG5, Taipei, Taiwan, 15/08/2016

Performance of the ANN models Initial Coeffificient of ANN Model / Vulnerability Error RMSE Variance standard determination Class RMSE Reduction reduction deviation R 2 Class 1 (Masonry) 0.182 0.126 31% 52% 0.81 Class 2 (Non-designed RC) 0.170 0.102 40% 64% 0.80 Class 3 (Low ductility RC) 0.172 0.112 35% 58% 0.81 Class 4 (Medium ductility RC) 0.153 0.094 39% 62% 0.81 Class 5 (High ductility RC) 0.147 0.096 35% 57% 0.82 Variance Reduction 50-64% + Good R 2 Satisfactory performance (given the small number of input parameters) ESG5, Taipei, Taiwan, 15/08/2016

Relative importance of input parameters : synaptic weights 0.6 0.5 Synaptic weight Proportion 0.4 Class 1 Class 2 0.3 Class 3 Class 4 Class 5 0.2 0.1 0 fstruc/fsoil Impedance Contrast PGA ESG5, Taipei, Taiwan, 15/08/2016

Dependence of damage increment on SSS inputs (example: class 3 - Low Ductility RC) Pga (m/s 2 ) 0.05 0.5 2 4 Scale for D soil - D rock C=20 C=10 C = 5 C = 2 ESG5, Taipei, Taiwan, 15/08/2016

Outline Introduction Proof of concept : comprehensive numerical simulation Robustness and field applicability Field applicability : site amplification proxy NL soil behavior (MDOF) ESG5, Taipei, Taiwan, 15/08/2016

Fiel applicability : Input parameters ✔ Loading : PGA ✔ Spectral coincidence : fstruct / fsoil ✔ Building mechanical behavior : typology class ✗ Site amplification : velocity contrast Cv  ? Other site amplification proxies : V S30 , V S10 , A 0HV , …. ESG5, Taipei, Taiwan, 15/08/2016

Numerical simulation of ambient noise After Bonnefoy-Claudet et al., (2006) Step 1: Definition of sources – receiver configuration sources receiver Step 2: Computation of Greens functions : DWN [Hisada, 1995] Step 3: Summation of all the individual noise synthetics in the time domain. Total ambient noise synthetics for each of the 887 soil profiles (5-10 min) ESG5, Taipei, Taiwan, 15/08/2016

Derivation and check of the "expected" H/V spectral ratio Geopsy ESG5, Taipei, Taiwan, 15/08/2016