Source-linked AI summary
A digital twin framework for civil engineering structures
Matteo Torzoni, Marco Tezzele, Stefano Mariani, Andrea Manzoni, Karen E. Willcox
TL;DR
Managing deteriorating civil structures requires condition-aware approaches that can support health monitoring, maintenance, and management decisions under uncertainty. This paper proposes a probabilistic digital twin that assimilates sensor data with deep-learning models, updates structural state sequentially, and plans actions dynamically; synthetic cantilever-beam and railway-bridge studies demonstrate accurate state tracking and timely control suggestions.
Problem
Deteriorating civil structures need condition-based and predictive maintenance approaches because failures or non-optimized planning can incur safety, economic, and social costs.
Method
A dynamic decision network encodes the asset-twin dynamics and uncertainty, while deep-learning models assimilate vibration data and offline reduced-order simulations support training and control-policy learning.
Results
The framework accurately tracked digital-state evolution with relatively low uncertainty and suggested appropriate control inputs within at most two time steps of the required structural-health response.
Takeaways & Limitations
Health-aware digital twins can support predictive maintenance decision-making for civil structures by connecting real-time health assessment with future-state prediction and control.
Takeaways & Limitations
The approach was assessed using high-fidelity simulation data with additive Gaussian noise rather than experimental data, and future work targets learned transition dynamics and reinforcement-learning planning.
Abstract
from arXiv · showhide
The digital twin concept represents an appealing opportunity to advance condition-based and predictive maintenance paradigms for civil engineering systems, thus allowing reduced lifecycle costs, increased system safety, and increased system availability. This work proposes a predictive digital twin approach to the health monitoring, maintenance, and management planning of civil engineering structures. The asset-twin coupled dynamical system is encoded employing a probabilistic graphical model, which allows all relevant sources of uncertainty to be taken into account. In particular, the time-repeating observations-to-decisions flow is modeled using a dynamic Bayesian network. Real-time structural health diagnostics are provided by assimilating sensed data with deep learning models. The digital twin state is continually updated in a sequential Bayesian inference fashion. This is then exploited to inform the optimal planning of maintenance and management actions within a dynamic decision-making framework. A preliminary offline phase involves the population of training datasets through a reduced-order numerical model and the computation of a health-dependent control policy. The strategy is assessed on two synthetic case studies, involving a cantilever beam and a railway bridge, demonstrating the dynamic decision-making capabilities of health-aware digital twins.
1 Introduction
The paper proposes a probabilistic digital twin for civil structures that links structural observations, health estimation, prediction, and maintenance decisions over time. Its framework combines dynamic Bayesian decision modeling, deep-learning-based data assimilation, and offline training for condition-based management.
- Deteriorating civil structures require optimized management because failures or poor maintenance planning can impose safety, economic, and social costs.
- The proposed digital twin uses a probabilistic graphical model, specifically a dynamic Bayesian network with decision nodes, to encode recurring physical-to-digital and digital-to-physical information flows.
- Structural response data are assimilated with deep-learning models to estimate damage presence, location, and severity from vibration-based measurements.
- The updated digital state supports prediction of future system evolution and uncertainty, enabling predictive decisions about maintenance and management actions.
- An offline phase trains health-identification models using labeled numerical data and learns a health-dependent control policy for online planning.Reduced-order modeling helps assemble training data representing potential damage and operational conditions over the structure’s lifetime.
- The framework incorporates automated feature extraction and uncertainty tracking for high-dimensional multivariate sensor time series.The implementation is made available in the public digital-twin-SHM repository.
2 Predictive digital twins using physics-based models and machine learning
The framework represents the asset–twin system as a dynamic decision network that assimilates sensed structural responses, updates digital-state beliefs, predicts evolution, and selects health-dependent actions. It combines Bayesian inference, deep-learning damage classification, computational models, and offline policy learning within a Markovian formulation.
- 2.1 Probabilistic graphical model for predictive digital twins: The dynamic decision network encodes random variables, actions, objectives, observations, estimates, and their probabilistic or deterministic dependencies.Circle, square, and diamond nodes represent random variables, actions, and the objective function; bold and thin outlines distinguish observed and estimated quantities.
- 2.1 Probabilistic graphical model for predictive digital twins: Observed structural responses are assimilated with deep-learning models to estimate damage presence, location, and magnitude, then Bayesian inference updates the digital-state belief.The digital state is a two-component vector describing damage presence/location and magnitude; the update combines the neural-network estimate with the previous belief and control-dependent transition dynamics.
- 2.1 Probabilistic graphical model for predictive digital twins: The model assumes indirect physical-state observation and Markovian evolution, so each time step depends conditionally only on the preceding step.The graph topology is specified from the first two time steps and can then be unrolled for any number of time steps.
- 2.1 Probabilistic graphical model for predictive digital twins: The framework predicts future digital-state evolution and uncertainty by unrolling the relevant state, quantity-of-interest, reward, and input subgraph over a prediction horizon.During forecasting, data assimilation and actions are omitted, while the predictive graph is extended across future time steps.
- 2.1 Probabilistic graphical model for predictive digital twins: A health-dependent policy maps beliefs about the digital state to control actions and is learned offline by maximizing expected rewards over the planning horizon.The policy is computed under an assumption of sufficiently accurate health sensing, using dynamic-programming value iteration; the reward balances control and health terms through α.
- 2.3 Data assimilation via artificial neural networks: The damage detection and localization network NNCL performs multiclass classification by mapping vibration recordings to one-hot encoded labels for predefined damage scenarios.The output has Ny+1 dimensions, with one active entry identifying the target damage class.
3 Numerical experiments
The methodology is demonstrated on two synthetic structural test cases: an L-shaped cantilever beam and a railway bridge. The numerical, probabilistic-graphical, and neural-network components are implemented using Matlab, Python, and TensorFlow-based Keras.
- 3 Numerical experiments: The proposed methodology is demonstrated on an L-shaped cantilever beam and a railway bridge.These are the two numerical test cases used to assess the approach.
- 3 Numerical experiments: The full-order and reduced-order models use Matlab and redbKIT, while the predictive digital-twin probabilistic graphical model uses Python and pgmpy.The neural-network architectures are implemented with the TensorFlow-based Keras API.
3.1 L-shaped cantilever beam
The cantilever-beam study assembles a reduced-order, noise-corrupted monitoring dataset across damage locations and severities, then evaluates digital-state inference and health-dependent maintenance decisions. The digital twin tracks evolving damage, predicts future degradation, and selects actions under both perfect and imperfect maintenance settings.
- Dataset assembly: The synthetic beam uses eight displacement measurements, seven possible damage regions, and stiffness reductions δ ∈[30%, 80%].Recordings are generated under varied loading and operational conditions for damage-state identification.
- Dataset assembly: A reduced-order model built from 400 finite-element evaluations has order NRB = 56 and generates a training dataset of I = 10,000 instances.The dataset trains the neural models used for digital-state classification and regression.
- Digital twin framework: 93.61% overall classification accuracy is achieved when the deep-learning models classify noisy finite-element observations among 43 digital states.Most errors involve damage in Ω6 or Ω7 because sensors near the clamped side are less sensitive to damage near the free end.
- Results: two available actions: For two available actions, the offline policy recommends operation until δ ∈[65%, 75%] or δ ∈[75%, 80%], depending on damage location, before perfect maintenance.The policy uses γ = 0.95 and α = 2.
- Results: two available actions: The online twin tracks digital-state evolution with relatively low uncertainty and suggests perfect maintenance within one time step of the ground-truth health requirement.Its prediction engine also forecasts degradation over a 20-step horizon to support future intervention planning.
- Results: four available actions: With four actions, the twin usually selects the optimal input, misclassifying two terminal-region cases before recovering correct health tracking within one time step.The errors select minor imperfect maintenance instead of doing nothing because the estimated damage interval is too high.
- Results: four available actions: Twenty-step predictions closely resemble the subsequent online evolution, despite unknown future health parameters and mismatched transition models.The forecast recommends major maintenance followed by two minor imperfect-maintenance actions.
3.2 Railway bridge
The railway-bridge case study uses synthetic sensor data and reduced-order modeling to train and test a digital twin for damage-state estimation and maintenance planning. The framework tracks evolving damage, recommends control actions, and produces forward predictions, while its transition model can underestimate deterioration.
- Synthetic observations: The digital twin uses ten displacement sensors recording 1.5-second histories at 400 Hz, with additive Gaussian noise.These recordings support simulated train-passage monitoring under the specified speed range.
- Case-study model: The synthetic railway bridge model represents damage as localized stiffness reductions across six predefined regions, with damage magnitudes from 30% to 80%.The bridge model includes train-speed and axle-load variability, and the damage remains fixed during each train passage.
- Offline preparation: A reduced-order model built from 400 full-order evaluations uses 133 POD modes, and generates a 10,000-instance training dataset.The reduced basis is selected with error tolerance ϵ = 10^-3.
- Offline classification: The offline classifiers achieve 91.39% overall accuracy in identifying the 37 digital states, with most errors between adjacent damage levels at the same location.The confusion matrix is therefore predominantly tridiagonal.
- Control policy: The learned policy keeps ordinary operation initially, recommends restricted operation for δ ∈[30%, 35%], and recommends repair once δ ≥65%.Restricted operation lowers deterioration but also reduces infrastructure revenue.
- Online decision-making: Online tracking follows the ground truth with at most a two-time-step delay, recommends restriction as estimated damage reaches δ ∈[35%, 65%], and eventually recommends perfect maintenance.Forward predictions are close to online behavior but are too optimistic about deterioration, motivating a more refined transition model.
4 Conclusions
The proposed digital twin integrates uncertain structural-health estimation, prediction, and maintenance decision-making in a probabilistic framework. Simulated cantilever-beam and railway-bridge studies demonstrate state tracking and timely control recommendations, while highlighting directions for broader adaptation and improved transition modeling.
- The framework encodes the asset–twin dynamical system and end-to-end observational and control information flow with quantified uncertainty.Digital states are updated through sequential Bayesian inference and used to predict future physical-system evolution.
- Railway-bridge predictions forecast degradation under doing nothing before assigning relatively high probability to restricting operational conditions after a few time steps.The prediction was close to the online experience but was too optimistic about deterioration and lagged the ground truth by two time steps.
- The offline phase uses physics-based and reduced-order models to populate deep-learning training datasets and learns a health-dependent control policy.The policy maps beliefs over the digital state to actions that feed back to the physical asset.
- The strategy tracked digital-state evolution under varying operational conditions with relatively low uncertainty in simulated cantilever-beam and railway-bridge monitoring.The assessment used high-fidelity simulation data corrupted with additive Gaussian noise.
- The framework suggested an appropriate control input within at most two time steps of when ground-truth structural health demanded it.The result was reported for the simulated monitoring studies, where the ground truth was unknown to the framework.
- The framework can be adapted to other structures and engineering systems by changing dynamic-Bayesian-network components, state-transition models, control inputs, and planning methods.The paper also notes that observational data may be assimilated with methods other than deep neural networks.
- Future work will learn transition dynamics from previous data and investigate reinforcement-learning planning over a finite horizon representing the asset’s design lifetime.
Code availability
The implementation code for the experiments and the trained deep-learning models is publicly available in the digital-twin-SHM repository.
- The public digital-twin-SHM repository contains the experiment implementation and can generate the paper’s digital-state estimation and prediction graphs.The trained deep-learning models from Appendix A are also provided.
A Implementation details
The appendix specifies the architectures, feature extraction design, optimization choices, and training setup for the NNCL and NNRG deep-learning models.
- The architectures and hyperparameters were selected through a preliminary study targeting lower LCL and LRG while retaining NNCL and NNjRG generalization capabilities.
- NNCL and NNRG share the same architecture, so the index for the NNRG models is omitted for notational convenience.
- NNCL and NNRG are both 12-layer deep-learning models with damage-sensitive feature extractors built from three one-dimensional convolutional units.The extractors are designed to be insensitive to input transformations unrelated to damage.
- Table 1 reports the NNCL architecture and selected training options, while Table 2 reports the corresponding NNRG details.
- The models use Xavier weight initialization and Adam optimization for up to 250 epochs, with cosine-decayed learning rates and weight decay equal to 0.05.The initial learning rates are 10^-3 and 10^-4 for the respective models.
- Training and validation use an 80:20 dataset split, with 20% of the data randomly held out for monitoring.