Building a neural network for the analysis of fracture in concrete
(2021) FKMM01 20211Materials Engineering
- Abstract
- SAAB Dynamics produces and develops warheads and uses computer simulations to evaluate its products. Recently, SAAB has begun using neural networks to quickly obtain accurate outputs, based on simulated data, from a set of input variables with success. This thesis investigated whether interim data in the form of video frames can improve the output of the network. This thesis investigated the use of a neural network in order to perform frame forecasting on simulated videos of reinforced concrete walls being subjected to impacts from high velocity projectiles.
The inputs were based on simulated data where each video consists of 88 frames where all frames are in greyscale. Two cost functions were be evaluated, Mean Squared Error (MSE) and... (More) - SAAB Dynamics produces and develops warheads and uses computer simulations to evaluate its products. Recently, SAAB has begun using neural networks to quickly obtain accurate outputs, based on simulated data, from a set of input variables with success. This thesis investigated whether interim data in the form of video frames can improve the output of the network. This thesis investigated the use of a neural network in order to perform frame forecasting on simulated videos of reinforced concrete walls being subjected to impacts from high velocity projectiles.
The inputs were based on simulated data where each video consists of 88 frames where all frames are in greyscale. Two cost functions were be evaluated, Mean Squared Error (MSE) and Structural Similarity Index Measure (SSIM) with three different models being evaluated, with various kernel size and number of hidden state tensor channels.
The architecture evaluated use a Convolutional LSTM autoencoder structure, which produces forecasts of the impact on the concrete wall based on 44 input frames. Using this method, the neural network managed to produce accurate forecasts of size and shape of the crater when MSE is used. While SSIM failed to produce accurate forecasts of crater shape and size, it better retained information in deeper forecasts. However, results indicated that the approach is not suitable for fracture analysis as cracks were not retained in deep forecast, SAAB should explore other approaches or architectures in order to improve predictions. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9063227
- author
- Fang, David LU
- supervisor
-
- Dmytro Orlov LU
- Lars Wadsö LU
- organization
- course
- FKMM01 20211
- year
- 2021
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Video Prediction, Frame forecasting, ConvLSTM, MSE, SSIM, Projectile Resistivity, Reinforced Concrete
- ISBN
- LUTFD2/TFMT--21/5066—SE
- language
- English
- id
- 9063227
- date added to LUP
- 2021-08-26 15:33:12
- date last changed
- 2021-08-26 15:33:22
@misc{9063227,
abstract = {{SAAB Dynamics produces and develops warheads and uses computer simulations to evaluate its products. Recently, SAAB has begun using neural networks to quickly obtain accurate outputs, based on simulated data, from a set of input variables with success. This thesis investigated whether interim data in the form of video frames can improve the output of the network. This thesis investigated the use of a neural network in order to perform frame forecasting on simulated videos of reinforced concrete walls being subjected to impacts from high velocity projectiles.
The inputs were based on simulated data where each video consists of 88 frames where all frames are in greyscale. Two cost functions were be evaluated, Mean Squared Error (MSE) and Structural Similarity Index Measure (SSIM) with three different models being evaluated, with various kernel size and number of hidden state tensor channels.
The architecture evaluated use a Convolutional LSTM autoencoder structure, which produces forecasts of the impact on the concrete wall based on 44 input frames. Using this method, the neural network managed to produce accurate forecasts of size and shape of the crater when MSE is used. While SSIM failed to produce accurate forecasts of crater shape and size, it better retained information in deeper forecasts. However, results indicated that the approach is not suitable for fracture analysis as cracks were not retained in deep forecast, SAAB should explore other approaches or architectures in order to improve predictions.}},
author = {{Fang, David}},
isbn = {{LUTFD2/TFMT--21/5066—SE}},
language = {{eng}},
note = {{Student Paper}},
title = {{Building a neural network for the analysis of fracture in concrete}},
year = {{2021}},
}