Source-linked AI summary
Deep Learning for Accelerated Reliability Analysis of Infrastructure Networks
Mohammad Amin Nabian, Hadi Meidani
TL;DR
Evaluating natural-disaster impacts on infrastructure is important, and the framework targets computational-time reduction. It uses classifier and end-to-end surrogates for reliability analysis and reports fast, high-accuracy infrastructure-response computation.
Problem
Evaluating natural-disaster impacts on infrastructure systems is important, and computational time is a target for reduction.
Source: §1 · pp. 2–3Method
The framework uses classifier predictions to reduce computation and an end-to-end surrogate to replace topology realization, connectivity determination, and connectivity averaging.
Source: §1 · p. 3, §3.2 · p. 10Results
The surrogates offer fast computation of infrastructure response with high accuracy.
Source: §5 · p. 23Takeaways & Limitations
Fast, sufficiently accurate surrogates can replace one or more computational modules in infrastructure systems.
Source: §5 · p. 23The analysis assumes bridges are the only transportation-network components vulnerable to and impacted by seismic hazards.
Source: §2.1 · p. 4
Abstract
from arXiv · showhide
Natural disasters can have catastrophic impacts on the functionality of infrastructure systems and cause severe physical and socio-economic losses. Given budget constraints, it is crucial to optimize decisions regarding mitigation, preparedness, response, and recovery practices for these systems. This requires accurate and efficient means to evaluate the infrastructure system reliability. While numerous research efforts have addressed and quantified the impact of natural disasters on infrastructure systems, typically using the Monte Carlo approach, they still suffer from high computational cost and, thus, are of limited applicability to large systems. This paper presents a deep learning framework for accelerating infrastructure system reliability analysis. In particular, two distinct deep neural network surrogates are constructed and studied: (1) A classifier surrogate which speeds up the connectivity determination of networks, and (2) An end-to-end surrogate that replaces a number of components such as roadway status realization, connectivity determination, and connectivity averaging. The proposed approach is applied to a simulation-based study of the two-terminal connectivity of a California transportation network subject to extreme probabilistic earthquake events. Numerical results highlight the effectiveness of the proposed approach in accelerating the transportation system two-terminal reliability analysis with extremely high prediction accuracy.
1. Introduction
Disaster-impact reliability assessment supports infrastructure planning, but simulation-based assessment of large systems is computationally expensive. The introduction proposes two deep neural network surrogates and reports a projected acceleration exceeding four orders of magnitude.
1. Introduction
- Disaster-impact assessment informs four planning objectives: mitigation, preparedness, emergency response, and post-disaster recovery.The introduction frames these objectives as relevant to infrastructure systems affected by natural disasters.Source: §1 · p. 2
- Large infrastructure reliability simulations are computationally expensive because of component counts, complex topology, dependent failures, and hazard uncertainty, limiting design optimization and sensitivity analysis.Approximate surrogates are presented as a more efficient alternative for response assessment.Source: §1 · p. 2
- A classifier speeds topology-specific connectivity evaluation, while an end-to-end surrogate replaces Monte Carlo simulation to calculate expected two-terminal connectivity from component failure probabilities.The two models address different parts of reliability analysis: connectivity determination and expected-connectivity evaluation.Source: §1 · p. 3
- The proposed end-to-end training approach uses classifier predictions instead of exact training data to reduce computational time.The paper also identifies multilayer neural networks as a way to enhance surrogate-based reliability analysis.Source: §1 · p. 3
- More than four orders of magnitude of acceleration is expected from the end-to-end surrogate for transportation-network two-terminal reliability analysis.The introduction says numerical results will show this acceleration and that it can facilitate sensitivity analysis and potentially other planning procedures for large networks.Source: §1 · p. 3
2. Two-Terminal Reliability Analysis
This section lays out a simulation-based workflow for estimating earthquake-related transportation network reliability, from ground-motion and bridge-fragility modeling to Monte Carlo connectivity evaluation.
2.1. Two-Terminal Connectivity
- Two-terminal connectivity means a surviving path between stakeholder-chosen source and terminal nodes; the model treats bridges as the network’s only earthquake-vulnerable components.Source: §2.1 · p. 4
2.2. Ground Motion Prediction
- GK15 predicts ground-motion intensity measures across specified earthquake and site conditions; its residual variability significantly affects San Jose–Mountain View two-terminal connectivity.Source: §2.2 · p. 5
2.3. Bridge Fragility Analysis
- HAZUS-MH fragility curves estimate bridge damage probabilities from shaking, site, and bridge attributes; this study removes bridges once extensive damage begins.Source: §2.3 · p. 6
2.4. Two-Terminal Reliability Analysis
- Monte Carlo draws roadway states from bridge survival probabilities and averages binary source–terminal connectivity across realizations; DFS checks connectivity in O(|V|+|E|).Source: §2.4 · p. 7
- Monte Carlo samples are repeated until the quantity of interest converges; connectivity calculations are parallelized, and the next section replaces exact DFS calculations with surrogates.Source: §2.4 · pp. 7–8
3. Surrogate Model for Two-Terminal Connectivity
The section introduces deep neural networks and two surrogate designs for two-terminal reliability analysis: one replaces connectivity checks for each sample, while the other replaces multiple Monte Carlo stages. It also defines measures for assessing surrogate predictions.
3.1. Deep Neural Networks
- A deep network feeds activated outputs through new weights and biases across hidden layers, and capacity can also increase by adding layers or units.Source: §3.1 · p. 8
- Regression uses Euclidean loss, while binary classification uses binary cross-entropy; backpropagation minimizes the selected loss to calibrate model parameters.Source: §3.1 · pp. 8–9
3.2. Deep Neural Networks for Two-Terminal Reliability Analysis
- The classifier surrogate replaces DFS for each Monte Carlo sample, mapping a binary roadway-state vector to a binary source-to-terminal connection result.Source: §3.2 · p. 9
- The end-to-end surrogate replaces topology realization, connectivity determination, and connectivity averaging, bypassing roadway-status realizations when estimating expected connectivity from failure probabilities.Source: §3.2 · p. 10
3.3. Surrogate Performance Measures
- Binary assessment uses classification accuracy, sensitivity or TPR, and specificity or TNR; the latter three measures apply only to binary classification.Source: §3.3 · pp. 10, 13
- QoI prediction accuracy compares surrogate two-terminal connectivity with the exact DFS value.Source: §3.3 · p. 10
4. Case Study for the San Jose-Mountain View Transportation Network
The case study evaluates classifier-based and end-to-end surrogate approaches for two-terminal connectivity analysis of the San Jose–Mountain View network under earthquake scenarios, including expected-connectivity estimation and bridge sensitivity analysis.
4. Case Study for the San Jose-Mountain View Transportation Network
- The San Jose–Mountain View transportation network consists of 39 bridges, 12 nodes, and 18 roadway links; numerical examples consider impact from the 1989 Loma Prieta earthquake, assume bridges alone are earthquake-vulnerable, and model the network as undirected.Source: §4 · p. 13
4.1. Classifier Surrogate Training and Prediction
- Across five fixed earthquake magnitudes, classifier predictions closely matched DFS, were about one order of magnitude faster, and achieved accuracies above 99.9%.The comparison used 100,000 network realizations for each earthquake event.Source: §4.1 · p. 16, Figure 8 · p. 17
- For probabilistic earthquakes, DFS and classifier surrogate connectivity estimates were 0.9002 and 0.9001, while runtimes were 603.97 and 53.89 seconds, respectively.The surrogate's αbinary, TPR, and TNR were 0.9997, 0.9998, and 0.9989, and it was about one order of magnitude faster.Source: §4.1 · pp. 18–19, Figure 9 · p. 19
4.2. Uncertainty-Aware Two-Terminal Reliability Analysis Using Classifier Surrogate
- DFS and classifier surrogate estimated expected two-terminal connectivity at 0.6853 each; runtime was 574.48 vs. 53.08 seconds, with αbinary, TPR, and TNR of 0.9990, 0.9993, and 0.9985.The evaluation included residual variability in ground-motion intensity measures.Source: §4.2 · p. 19, Figure 10 · p. 20
- Accounting for ground-motion intensity uncertainty yielded reliability of 0.69 rather than 0.90, indicating that omitting this uncertainty can overestimate network reliability.The uncertainties include aPGA and Sa and are often ignored in reliability studies.Source: §4.2 · p. 20
4.3. End-to-End Surrogate Training and Prediction
- The end-to-end surrogate is intended to accelerate two-terminal connectivity calculations by replacing the entire Monte Carlo simulation process.The passage introduces its training as the next step but provides no end-to-end results in this part.Source: §4.3 · p. 20
- The end-to-end surrogate maps roadway failure probabilities to expected two-terminal connectivity and uses classifier-surrogate-generated training data.Training and evaluation data include 3,000 magnitude samples and 100,000 topology samples per magnitude sample; the architecture has five hidden layers.Source: §4.3 · pp. 20–21, Figure 11 · p. 21
4.4. One-at-a-time Sensitivity Analysis Using End-to-End Surrogate
- One-at-a-time sensitivity analysis varies each bridge’s survival probability while holding the others at nominal values, using a 10% amplification to represent retrofit plans.The goal is to identify influential bridges for repair prioritization under budget constraints.Source: §4.4 · pp. 21, 23
4.3. End-to-End Surrogate Training and Prediction
- Both DFS and the end-to-end surrogate estimated expected two-terminal connectivity at 0.9001; their computational times were 7,857.92 seconds and 0.71 seconds, respectively.Source: §4.3 · p. 21
4.4. One-at-a-time Sensitivity Analysis Using End-to-End Surrogate
- The OAT comparison uses an earthquake magnitude range of 7.3–7.9 Mw and presents the three most and least sensitive components.Expected connectivity for nominal and retrofitted networks is calculated with both DFS and the end-to-end surrogate.Source: §4.4 · p. 23
- Under the probabilistic earthquake event, the network’s two-terminal connectivity probability is 0.7641 using DFS and 0.7643 using the end-to-end surrogate; the surrogate offers substantial computational savings in repetitive analyses.Source: §4.4 · p. 23
5. Conclusion
The conclusion presents deep-learning surrogates as a way to accelerate infrastructure reliability analysis while retaining high accuracy. It also outlines data augmentation and GPU use as possible extensions to improve accuracy and computational efficiency.
5. Conclusion
- Surrogates can replace one or more reliability-analysis modules, potentially facilitating optimal planning and management of infrastructure systems exposed to natural hazards.The conclusion frames this as a potential application of surrogate-based analysis.Source: §5 · p. 23
- The case study highlights fast computation of infrastructure responses with high accuracy using deep-learning surrogates.The conclusion presents this as a demonstrated capability of the studied surrogates.Source: §5 · p. 23
- Automatic feature engineering and detection can reduce the need to identify features manually and support broad applicability to nonlinear responses.Source: §5 · p. 24
- Training-data augmentation is proposed as an extension for improving prediction accuracy.For TNR, additional no-connectivity realizations can be generated by randomly failing surviving roadways, with no extra computational burden because their connectivity is already known.Source: §5 · p. 24
- GPU use is proposed to improve computational efficiency because deep-learning matrix multiplications are parallelizable on GPUs.Source: §5 · p. 24