Neural Emulator

I have implemented a deep neural network in Matlab and TensorFlow that acts as neural emulator which estimates the future acceleration responses of the structure in the next five time-steps given the accelerations in the current time-step. This way, we can substantially reduce the computation time comparing to the finite element analysis which allows us to enter such estimation of the future structural responses to the calculations of the control forces.

Developing such neural emulator was a challenging task as it requires altering several network architectures, properties, and types. It also requires a large dataset of structural responses in different types of external excitations including earthquake records, harmonic loads, white noise, and stochastic loads. The neural emulator has been tested with multiple scenarios and the results show that it can estimate future responses of the building under any external excitation with an acceptable accuracy. This neural emulator can be utilized as part of the intelligent controller in further steps.

After designing the emulator parameters, training NN is the main step in developing the neural emulator. Considering main points about the nature of the structural control problems, application limits, and the future expected function of the neural emulator may help to make appropriate assumptions in the training phase. Some of this points can be mention as follow:

1. The earthquake excitation is stochastic and unpredictable

2. During the earthquake, the excitations propagate thorough soil layers into the structural base, therefore, measuring the displacement cannot be directly performed and the feasible direct measurements can include the accelerations.

3. After the earthquake induced vibration phase, the free vibration phase of the structure would begin which may include harmonic responses.

4. As in the reality, some time delays exist in the system, therefore, the emulator shall be able to estimate the response of the structure in some time steps for ahead.

Taking these issues into account, an step by step training scenario for the neural emulator have been developed as follow:

a- The neural network’s variables randomly are initialized.

b- The NN is trained with data from analytical model under white noise load values as actuator forces.

c- The NN being retrained with a harmonic loads as actuator forces with 1/3 epochs of full the training course.

d- The NN being retrained under data obtained from analytical model in which the building is analyzed under existence of earthquake excitations and also randomly generated white noise forces as actuator forces with 1/3 of full training epoch number.

e- The NN being trained with training data obtained as randomly mix data from previous three training data sets. At the end of the each step, the performance of the NN has been investigated under randomly generated white noise as actuator forces.

The results show that Neural Emulator is able to predict the response of the structure to the actuator forces with a small error which is about 0.001. As next step, the trained Neural Network is tested under the expected situation in the intelligent controller under earthquake excitation. In this regard, without additional retraining procedure, the Neural Emulator is used to predict the future response of the structure using history of the accelerations and actuator forces in 5 time steps.