Advances in Data-Driven Risk-Based Performance Assessment of Structures and Infrastructure Systems

The special collection on Advances in Data-Driven Risk-Based Performance Assessment of Structures and Infrastructure Systems is available in the ASCE Library (https://ascelibrary.org/jsendh/advances_data-driven_risk-based_assessment).

Structures and infrastructure systems are under the threat of extreme hazards, which may cause damage and even failure of the structures and systems. Robust and efficient performance assessment methods are essential for quantifying the risk and system capacity levels and their uncertainties to ensure structural safety. Traditional mechanics-based approaches are capable but face several challenges vis-à-vis computational efficiency in analyzing complex structures and distributed or interconnected infrastructure systems and high-dimensional uncertainty quantification, among others. In recent years, the advent of artificial intelligence algorithms and the near ubiquity of big data have paved the way for data-driven approaches for practical and highly granular performance assessment and risk analysis of civil infrastructure systems. This special collection contains one review paper, one technical note, and nine technical papers. It aims to bring together studies in this new frontier and demonstrate how the data-driven method can assist and improve the performance assessment of structures and infrastructures under various hazards, e.g., earthquakes, winds, and/or environmental loads.

The collection begins with a state-of-the-art review by Wang et al. (2022), which comprehensively reviews the application of machine learning (ML) methods for risk and resilience assessment in structural engineering. It summarizes the related papers concerning six attributes—i.e., method, task type, data source, analysis scale, event type, and topic area—and focused on four types of infrastructure—namely, buildings, bridges, pipelines, and electric power systems. The current progress in the ML-based performance assessment of these four infrastructure types is reviewed in detail, and it is found that the utilization of ML indeed enhanced the assessment accuracy as well as efficiency. Meanwhile, the accessible database is also summarized for further research and/or education purposes in structural engineering. Three major trends and challenges are highlighted for future studies: data scarcity and sparsity, interdependence and regional-level effects, and the consideration of climate change factors.

Four papers are focused on the assessment of structures under earthquakes. Chen et al. (2022b) develop a probabilistic ML model based on natural gradient boosting (NBG). Unlike the traditional ML methods, which offer a deterministic prediction of the target, the proposed method outputs not only the mean value but also the probability density function (PDF) of the target, so it can help estimate the influence of uncertainties. Two illustrative examples are given to show the superiority of the model: one regression problem involving shear strength prediction of shear walls, and one classification problem related to seismic damage state prediction of bridges.

To avoid the complicated preprocessing of signal data and/or using highly compressed features of the input for an ML model, Sajedi and Liang (2022) employ customized filter banks for feature extraction, which is inspired by speech recognition technology, to directly feed the raw data to ML models. In addition, they also develop a multiheaded neural network architecture that enables using ground motions with variable lengths and different sampling rates as inputs. The proposed method greatly simplifies the preprocessing of the data and is implemented in the probabilistic seismic performance assessment. A case study of the reinforced concrete bridges is conducted, and how the method can be integrated into fragility analysis is shown.

Sun et al. (2022) propose a cross-building response reconstruction model with the recorded data from 188 buildings under 25 earthquakes. Peak story drift ratios (PSDRs) and peak floor accelerations (PFAs) are the key responses of interest, and the model is formulated by extending the modern ground motion model to add a building response term. Meanwhile, a kriging model is adopted to estimate the within-event residuals from spatial interpolation of structural response recording. The model can be used for a rapid post-earthquake assessment of the building portfolios and has proven to have a good performance for events that have a reasonable number of recordings.

Chen et al. (2022a) explore ML techniques in the seismic assessment of bridge networks. By setting the survival or failure of the individual bridges as input and the survival or failure of the whole network as output, seven ML algorithms are used to train the prediction model for the estimation of the seismic reliability of a bridge network. CatBoost, random forest, and decision tree methods obtained the best prediction performance. The method can also be used to calculate the seismic fragility of the network and then identify the important ranking of the individual bridges in the network, which can achieve fast decision-making for optimal retrofit plans.

Two papers investigate the data-driven reliability and resilience assessment methods for structures under strong wind. Dong et al. (2022) present an approach that establishes a relation between hurricane scenario selection for community resilience assessment and the building regulatory process. The hurricane scenarios are the dominant contributors to coastal community risk with a specific return period. The approach provides a basis for simulating multiple coupled hazards with distinct characteristics while accounting for the uncertainties in these coupled storm surges, waves, and coastal flooding hazards.

Song et al. (2022) use the probabilistic support vector machine (SVM) to establish the surrogate model in high-dimensional reliability analysis problems. They propose a new learning function that can also take advantage of the information of wrongly classified samples and their distance to the nearest training points for each candidate sample in the sampling pool. Also, they develop an error-guided stop criterion to enhance the numerical efficiency of the method. This method can overcome the limitations of traditional SVM approaches. It is successfully applied to a real engineering problem involving the wind-reliability analysis of a transmission tower, and comparable performance with the reference Monte Carlo method is obtained.

The last four papers present studies on structural performance assessment using monitored data under environmental loads. Zhu et al. (2022) utilize the relationship between the increments of temperature and the associated bridge responses for assessing the health of bridge structures. Four typical mappings are achieved via a gated recurrent unit (GRU) neural network with different inputs of temperature increments for both displacement increments and strain increments. Principal component analysis is also used for input dimension reduction. Data recorded on a steel truss bridge during a 30-day period are used for validation. This case study revealed that the increments in temperature-induced responses could be effectively mapped when the full spatial temperature distribution within the bridge is considered in the input.

Similarly, Huang et al. (2022) utilize the relationship between temperature and displacement for assessing the condition of bridge bearings through structural health monitoring data. Due to the data sparsity and the uncertainties within recorded temperature differences and associated bearing displacements, a sparse Bayesian learning–based modeling approach is proposed for temperature-induced bearing displacement. Moreover, on the basis of the reliability and anomaly analysis principles, the corresponding index for the bearing is defined for assessment and warning. Data monitored from an in-service long-span railway bridge are used for validating the feasibility of the approach. The results demonstrate the effectiveness of the method and the devised index in assessing the health of the bearings and providing early damage warning.

The global navigation satellite system (GNSS) sensor network gradually becomes a cost-effective alternative for bridge displacement monitoring. Manzini et al. (2022) propose an anomaly indicator for damage detection of bridges on the basis of the GNSS-monitored response time series. The main idea of this study is also the usage of the relationship between the responses of each sensor and the one between the environmental effect and the response. Various ML algorithms, such as neural networks and support vector machines, are used for modeling these relationships. The residuals between the actual observations and ML predictions are used to indicate anomalies. Subsequently, based on an idea that is similar to bagging, the more frequently the models identified a specific time period as abnormal for a particular sensor, the more likely that sensor would be associated with an anomaly event. A cable-stayed bridge equipped with GNSS is used for validation in this study, and the real and virtual anomalies are successfully isolated and located via the proposed approach.

Normally, a closed-form expression of a monitored dynamic system is desired for various tasks such as controlling and assessing. However, a typical deep neural network can only provide a black box model, with no explicit formulas from measured responses. Thus, Chen et al. (2022c) designed a symbolic neural network consisting of predefined operators and specific network structures with an adaptive pruning strategy for efficiently identifying and generating the equation of motion of the measured response from a dynamic system. After an examination involving numerical and experimental studies on dynamic systems, the potential of this method for extracting hidden mechanisms for real-world applications is revealed. The weaknesses and necessary improvements of the proposed approach are also summarized.