• No se han encontrado resultados

BATTERY RECIPROCATOR II (530.715)

In document BATTERY POWER LINE II (página 44-47)

Previous section defines the most appropriate surrogate models for bridge deck

unseating classification. The results from the FSI Model methodology were

implemented to develop these surrogate models. However, the efficiency of fragility

analysis can be further improved by using more simplified structural analysis

models, as long as it is capable of capturing the failure mode of interest. Fluid-

structure interaction models such as the one presented in Section 4.2 have the

advantage of accurate estimation of the bridge behavior and providing insight on

the response of the bridge but they are computationally intense as noted in Section

6.3.

Therefore, this section constructs reliability surrogate models on the output

of the MCS Static and Dynamic Models, and compares them with the reliability

surrogate model constructed from the FSI Model. The same case study bridge

model with the same experimental design and approach presented in Section 6.3 is

utilized here; however, the output data used to derive the surrogate model is based

and the Static Model. The full set of random variables for this simulation is

presented in Appendix V.

Two random forests are trained over the results of the MCS Static Model

output and Dynamic Model output, respectively. The resultant random forests are

designated as S-RF and D-RF, respectively. Since all random forests algorithms are

already constructed based on the training data, predicted failure probabilities can

be directly estimated for any Hmax and Zc. A sample size of 10,000 points over the

hazard intensity measures is generated and used to identify the misclassification

error for the other models versus FSI. The trained random forests are compared

with the FSI-RF in Figure 7-3. The S-RF and D-RF are in agreement with the

FSI-RF. The misclassification errors for these two random forests with respect to

FSI-RF are 0.19 and 0.13, respectively. Since the random forests surrogate model

has high accuracy (the surrogate model error is around 0.01), the misclassification

error is due to the reduction in the structural analysis and load modeling accuracy

from the FSI Model to the Dynamic and Static Models. Confusion matrices are

shown in Table 7-3. As it can be concluded, the results of the MCS Static and

Dynamic Models are conservative; (i.e., more false negatives are reported than false

susceptible bridges structures under hurricane events. The F1-measures for the S-

RF and D-RF models compared to FSI-RF are 0.65 and 0.75 respectively. This

result shows that the higher accuracy is achieved by using more advanced models.

However, fast screening of vulnerable coastal bridges, especially for real time

application, is more practical with the MCS Static and Dynamic Models.

Figure 7-3. Comparison of two random forests models trained over different analysis output data: (a) FSI Model versus Static Model; (b) FSI Model versus Dynamic

Model.

Table 7-3. Confusion matrices for three random forests models trained over FSI, Static and Dynamic Models output.

A ct u a l cl a ss (F S I) Predicted class (Static) Predicted class (Dynamic)

Survived Failed Survived Failed

Survived 6484 1599 7119 964

7.4. Summary

This chapter and Chapter 6 present the core contribution of this research for the

development of fragility models for coastal bridges subjected to hurricane induced

wave and surge loads. The results of the structural analysis models presented in

Chapter 5 reveal that deck unseating is a brittle failure mode. Therefore, the

output becomes a categorical data; i.e., failed (unseated) or survived (seated). A

continuous model over the entire range of hazard intensity measures is required to

be constructed on this categorical data that can provide failure probabilities for

any realization of hazard parameter (that may not be simulated). Surrogate models

that relate continuous input to categorical output are required. Therefore

traditional response surface models are not appropriate. Three statistical learning

methods —logistic regression, support vector machines, and random forests— are

applied to the result of the analysis of bridge deck unseating. These methods are

nonlinear classifiers that can provide high accuracy classification for binary data.

Logistic regression, support vector machines with the Gaussian radial basis

function kernel, and random forests provide a high quality approximation of the

bridge deck unseating failure mode. Nonetheless, the performance of the random

Random forests surrogate models are trained over the results of Static, Dynamic,

and FSI Models and compared to each other, where the output of FSI Model is

considered as the most accurate and the basis for comparison. The result of the

Dynamic Model is in good agreement with the FSI Model, and therefore, can be

used for a more computationally efficient reliability assessment. The error increases

as the modeling accuracy decreases; i.e., Static Model has higher error than

Dynamic Model. However, the Static Model can provide an effective means to

142

Chapter 8

Retrofit Measures for Coastal Bridges

and Definition of New Capacity Limit

State Functions

This chapter introduces potential retrofit measures for coastal bridges to prevent

the deck unseating failure mode. Providing strong connections between the bridge

super- and substructure is one of the recommended methods for retrofitting coastal

bridges (Padgett et al. 2008; Padgett et al. 2009; Sawyer 2008). Other retrofit

measures, such as shear keys and restrainer cables that traditionally have been

used for seismic hazard mitigation of highway bridges, can also be used for coastal

bridges to prevent deck unseating. All of these measures can potentially transfer

state is not simply a deck displacement value, as considered in the fragility analysis

in prior chapters of this dissertation. The transferred forces may introduce damages

to the substructure which have not been explored in past research. This chapter

introduces the potential retrofit measures to prevent deck unseating failure mode.

Also, new capacity limit state functions for retrofitted bridges are defined. Chapter

9 will apply the approach to derive capacity limit state functions to evaluate the

viability of the prospective retrofits to improve the reliability of coastal bridges in

the Houston/Galveston area inventory.

In document BATTERY POWER LINE II (página 44-47)

Documento similar