• Neither neural network was able to pick up the residual drift of GM7. • For both NN, 30% percent of 12000 points for the smallest magnitude earthquake was sufficient to predict a good portion of the rest of the GMs with good accuracy. More training epochs and data should be implemented to see if better performance is reached

• Results also suggest that features need to be selected carefully to avoid noise in the target output. The ground displacement was an important feature.

• Double numerical integration is still a more practical way to compute the displacement from accelerations as it takes less time than training neural networks.

The author is grateful to the TAs from the class whom helped with guidance. In addition, author is

thankful to the National Science Foundation through the Network for Earthquake Engineering Simulation (award NEESR-1135029), which made the data available.

IMPLEMENTING MACHINE LEARNING IN EARTHQUAKE ENGINERING CS 229 Machine Learning, Autumn 2016 Poster Presentation

Cristian Acevedo (cacevedo@stanford.edu)

INTRODUCTION

Motivation: A measure of structural performance is the amount of displacement, x(t), a structure experiences during an earthquake. The higher the displacement, the more damage the structure sees. The displacement can be typically derived by double integrating the acceleration, .. Nevertheless, this process is not trivial as the data requires pre-processing (e.g., baseline correction and blind filtering) to obtain reasonable results. As seen below, residual displacements are not well captured from the acceleration data.

OBJECTIVE

RESULTS CONCLUSIONS

ACKNOWLEDGMENTS

FUTURE WORK

The use of accelerometers in earthquake engineering is very important for engineers as it helps them understand and quantify the magnitude of seismic forces. Structures are typically instrumented at the ground and floor levels; thus, when an earthquake occurs, the ground and floor accelerations are recorded. This data can then be used to understand the performance of the structure and it allows engineers to create prediction models to design future ones.

DATA

ML FEATURES RESULTS

1000 2000 3000 4000 5000 6000 7000−4

−2

0

2

4

Time Step

Floo

r Acc

eler

atio

n [g

]

1000 2000 3000 4000 5000 6000 7000−20

−10

0

10

20

30

Time Step

Dis

plac

emen

t [in

]

IntegratedTarget

•  Apply machine learning to acceleration data to obtain displacement data

•  Compare its results to double integration method

Data was obtained from the experimental results of a 2-story wood-frame Unibody house. The structure was tested on an earthquake simulator or shake table. The 1989 Loma Prieta Earhtquake (GM) was simulated at seven amplitudes. Each floor was instrumented with five accelerometers. The displacement was measured with string potentiometers. A total of 12,000 data points were recorded per instrument.

Shake Table

Unibody House

ML Implementation Neural networks were investigated in this project using Matlab libraries. Since these are time sequences, Feedforward Neural Networks (FFNN) and Recurrent Neural Networks, respectively, were particularly implemented.

Eight features were considered for this project. These features were selected based on data that can be collected from an earthquake. These features were:

Building EarthquakeAcceleration at floor of interest Acceleration of the ground

Vertical acceleration of floor of interest

Ground displacement (derived from acceleration)

Power spectral density of horizontal acceleration

Vertical acceleration

Ratio of Fz floor / Fz ground Power spectral density of ground acceleration

FFNN RNN

The total amount of features used were six. The ones in red were not used as they added noise to the output.

Data trained with 30% of GM1 and used to predict the behavior at higher intensities. Small windows of these predictions are shown.

3000 4000 5000 6000 7000−20

−10

0

10

20

30

Time Step

Dis

plac

emen

t [in

]

FFNNIntegratedTarget

4000 4200 4400 4600 4800 5000−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

Time Step

Dis

plac

emen

t [in

]

FFNNIntegratedTarget

4000 4200 4400 4600 4800 5000−4

−3

−2

−1

0

1

2

Time Step

Dis

plac

emen

t [in

]

FFNNIntegratedTarget

Predicting 70% of GM1 Predicting GM6 Does a fairly good job, sometimes not c a p t u r i n g t h e peaks.

Predicting GM7 Does not do a good job at picking up peaks o r r e s i d u a l displacement

This was just a preliminary study incorporating machine learning to this problem, which arose from the interest of the author. More research is needed before discarding neural networks as a possible method to pick up residual displacements. More data (either experimental or simulated) should be used to generalize the model and check if performance is improved.

Predicting 70% of GM1

FFNN RNN

Predicting GM7 Does not do a good job at picking up peaks o r r e s i d u a l displacement

4000 4200 4400 4600 4800 5000−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

Time Step

Dis

plac

emen

t [in

]

RNNIntegratedTarget

Predicting GM6 Does a fairly good job with a bit more noise

4000 4200 4400 4600 4800 5000−5

−4

−3

−2

−1

0

1

2

Time Step

Dis

plac

emen

t [in

]

RNNIntegratedTarget

3000 4000 5000 6000 7000−20

−10

0

10

20

30

Time Step

Dis

plac

emen

t [in

]

RNNIntegratedTarget

0 10 20 30 40 50 60 70 8010−4

10−3

10−2

10−1

100Best Training Performance is 0.00013496 at epoch 85

Mea

n Sq

uare

d Er

ror

(mse

)

85 Epochs

TrainBest

0 5 10 15 2010−4

10−3

10−2

10−1

100

101Best Validation Performance is 0.00014245 at epoch 15

Mea

n Sq

uare

d Er

ror

(mse

)

21 Epochs

TrainValidationTestBest

Looking at the performance, the RNN has better performance but takes longer to train and the predictions are not always as smooth.

It was noticed that if the data was trained multiple times for the same neural network, different performance training weights were obtained.

Double Numerical Integration