8 Fast probabilistic damage detection using inverse surrogate models
-
Michael D. Todd
Abstract
Bayesian inference has been widely employed in structural health monitoring (SHM) to evaluate structural integrity by inferring damage indices based on physics- based computational simulations and or field observations; this is a fundamentally inverse problem. To overcome the challenge caused by the computationally intensive physics-based simulations, machine learning (ML) models have been enabled as forward emulators to supplant the conventional physics-based simulations in Bayesian inference. While the forward ML models provide a viable solution to the computational challenge in probabilistic damage detection using Bayesian inference, it requires coordination with Bayesian inference algorithms, followed by evaluation via likelihood functions. For complex stochastic systems, the likelihood functions sometimes are analytically intractable even though the forward model has been replaced with a computationally cheap ML model. This poses challenges to model-based probabilistic damage detection in SHM. This chapter presents two inverse surrogate modeling methods for two scenarios of problems to enable for fast probabilistic damage detection without the need of evaluating the likelihood function. The first method is for problems where a one-to-one mapping exists between the observations and damage indices. A variational Bayesian neural network is developed to directly map observations to damage indices and to account for uncertainty in the observations. The second method adopts the idea of normalizing the flow and is suited to problems with highly nonlinear behavior when there is no clear one-to-one mapping between the observations and the damage indices. A summary and a conditional invertible neutral network (cINN) are employed to extract features for damage detection and to perform the nonlinear mapping between observations and posterior distributions of damage indices. Two engineering application examples, including a miter gate and a concrete building frame, are used to demonstrate the presented inverse surrogate models for probabilistic damage detection in SHM.
Abstract
Bayesian inference has been widely employed in structural health monitoring (SHM) to evaluate structural integrity by inferring damage indices based on physics- based computational simulations and or field observations; this is a fundamentally inverse problem. To overcome the challenge caused by the computationally intensive physics-based simulations, machine learning (ML) models have been enabled as forward emulators to supplant the conventional physics-based simulations in Bayesian inference. While the forward ML models provide a viable solution to the computational challenge in probabilistic damage detection using Bayesian inference, it requires coordination with Bayesian inference algorithms, followed by evaluation via likelihood functions. For complex stochastic systems, the likelihood functions sometimes are analytically intractable even though the forward model has been replaced with a computationally cheap ML model. This poses challenges to model-based probabilistic damage detection in SHM. This chapter presents two inverse surrogate modeling methods for two scenarios of problems to enable for fast probabilistic damage detection without the need of evaluating the likelihood function. The first method is for problems where a one-to-one mapping exists between the observations and damage indices. A variational Bayesian neural network is developed to directly map observations to damage indices and to account for uncertainty in the observations. The second method adopts the idea of normalizing the flow and is suited to problems with highly nonlinear behavior when there is no clear one-to-one mapping between the observations and the damage indices. A summary and a conditional invertible neutral network (cINN) are employed to extract features for damage detection and to perform the nonlinear mapping between observations and posterior distributions of damage indices. Two engineering application examples, including a miter gate and a concrete building frame, are used to demonstrate the presented inverse surrogate models for probabilistic damage detection in SHM.
Chapters in this book
- Frontmatter I
- Preface V
- Contents VII
- Editors’ short biography IX
- 1 Seismic site investigation and structural amplification based on geotechnical and structural health monitoring 1
- 2 Impact of industry 4.0 technologies on structural health monitoring 29
- 3 Health monitoring of recycled aggregates-reinforced concrete beams retrofitted by concrete jacket using piezoelectric transducers 63
- 4 Identification of critical response of bilinear-hysteretic SDOF model with tuned inertial mass damper under long-duration ground motion through internal simulation monitoring 79
- 5 Damage identification using physics-based datasets: From convolutional to metric-informed damage-sensitive feature extractors 101
- 6 Structural health monitoring of steel plates using modified modal strain energy indicator and optimization algorithms 125
- 7 Vibration-based damage detection using a novel hybrid CNN-SVM approach 137
- 8 Fast probabilistic damage detection using inverse surrogate models 159
- 9 Remote sensing techniques for post-disaster infrastructure health monitoring 197
- 10 Recent developments in the building information modeling-based programs used for structural and architectural purposes 215
- 11 Evaluating the current state of digitalisation of the UK construction industry 237
- Index 259
Chapters in this book
- Frontmatter I
- Preface V
- Contents VII
- Editors’ short biography IX
- 1 Seismic site investigation and structural amplification based on geotechnical and structural health monitoring 1
- 2 Impact of industry 4.0 technologies on structural health monitoring 29
- 3 Health monitoring of recycled aggregates-reinforced concrete beams retrofitted by concrete jacket using piezoelectric transducers 63
- 4 Identification of critical response of bilinear-hysteretic SDOF model with tuned inertial mass damper under long-duration ground motion through internal simulation monitoring 79
- 5 Damage identification using physics-based datasets: From convolutional to metric-informed damage-sensitive feature extractors 101
- 6 Structural health monitoring of steel plates using modified modal strain energy indicator and optimization algorithms 125
- 7 Vibration-based damage detection using a novel hybrid CNN-SVM approach 137
- 8 Fast probabilistic damage detection using inverse surrogate models 159
- 9 Remote sensing techniques for post-disaster infrastructure health monitoring 197
- 10 Recent developments in the building information modeling-based programs used for structural and architectural purposes 215
- 11 Evaluating the current state of digitalisation of the UK construction industry 237
- Index 259
