Deep learning for structural health monitoring: An application to heritage structures

arXiv:2211.10351 · eess.SP, cs.LG · Submitted 2022-11-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Deep learning for structural health monitoring".

Dev: Here's the gist of the paper. Deep learning for structural health monitoring: An application to heritage structures employs recent deep learning techniques for time-series forecasting to inspect and detect anomalies in large datasets recorded during a long-term monitoring campaign on the San Frediano bell tower in Lucca The data analysis shows promising results, highlighting anomalous events like the Amatrice earthquake that happened on 24 August 2016 or the Santa Croce celebrations on 13 September 2016.

Rosa: Problem Context Heritage structures are threatened worldwide by aging, material deterioration, environmental actions, and extreme meteorological events due to climate changes, and therefore they need maintenance and restoration In Italy, this need is exacerbated by the seismic vulnerability of the territory, which has made structural health monitoring (SHM) increasingly attractive because it provides a relatively inexpensive tool for promptly assessing the structural conditions, planning, and controlling maintenance interventions The availability of high-sensitive instrumentation and extensive data sets from long-term monitoring protocols opened new issues about data analysis, particularly regarding the implementation of automatic anomaly detection and early warning procedures Exploiting artificial intelligence (AI) represents a significant opportunity in this data-driven context.

Dev: Model Architecture The proposed approach belongs to a set of techniques that pertain to the domain of unsupervised anomaly detection Specifically, a network is trained to understand the normality patterns present in some data and then, at inference time, the more significant deviations from the learned normality patterns are considered anomalies In this work, we use a state-of-the-art deep neural network for learning the structure’s dynamics directly from data called Temporal Fusion Transformer (TFT).

Rosa: The TFT can also estimate the uncertainty of the prediction, which is fundamental during the anomaly detection phase.

Paper summary: Rosa: We've covered how the paper uses time-series forecasting with a Temporal Fusion Transformer to find anomalies in structural data from a tower in Lucca, and they've mentioned that it works for things like earthquakes and celebrations. Now let's look at what the authors are actually claiming regarding this method.

Dev: The thesis is that by framing the problem as an unsupervised anomaly detection task, they can train a model to understand the normal behavior of these old buildings, and then any significant deviation from that learned norm is considered an anomaly.

Rosa: They use the TFT model specifically because it can estimate prediction uncertainty, which lets them quantify exactly how far away an observed measurement is from the predicted confidence intervals.

Dev: So, if a new measurement falls outside of those defined margins, it's flagged as anomalous based on a threshold they set, and they test this against training and testing data to verify its performance.

Taro: It sounds like the system is designed to be sensitive enough to catch events that are genuinely different from the learned baseline behavior of the structure. But what happens if the damage itself changes how those normal patterns look?

Rosa: The paper doesn't explicitly detail how it handles true structural degradation, but by detecting events like earthquakes and mechanical vibrations, they show its capability in identifying significant external disturbances.

Dev: The paper is about applying recent deep learning techniques to process sequential data for building service life and to localize the onset of damage when data comes from repeated or long-term measurements, which is what they are trying to do.

Taro: So it's not just about sudden shocks; it's also about continuous monitoring over time. That suggests the model should be good at catching gradual changes too, provided those changes are still somewhat predictable within the learned structure.

Rosa: That’s right. The goal is to use this data-driven approach to move beyond just looking at static measurements and instead get insights into how the structure behaves over a long period of time.

Conclusion: Dev: So, looking at the paper "Deep learning for structural health monitoring: An application to heritage structures," the authors successfully applied this deep learning method to monitor the San Frediano bell tower in Lucca, and they showed it could correctly identify events like the Amatrice earthquake and those specific Santa Croce celebrations.

Rosa: The main implication is that we can use these sophisticated AI techniques to get early warning signals for structural issues in old buildings using long-term monitoring data.

Dev: It moves the field toward using data-driven methods to assess structural conditions without needing constant, expensive manual inspection of every single measurement point.

Taro: For someone who only listens to this show, it means that complex models can sift through huge amounts of vibration and environmental data to tell you when something important is happening in an old building.

Rosa: Precisely. The work suggests that integrating deep learning into structural health monitoring provides a way to find those specific moments—the earthquakes or the mechanical stresses—that might otherwise be missed by traditional methods.

Institute of Information Science and Technologies 膌-CNR

eess.SP, cs.LG

Submitted: 2022-11-04

Updated: 2026-11-09

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: Deep learning for structural health monitoring: An application to heritage structures employs recent deep learning techniques for time-series forecasting to inspect and detect anomalies in large…

Key concepts

Structural Health Monitoring (SHM)
This is using sensors to constantly check the condition of a building. It helps identify problems like cracks or deterioration early, which is crucial for maintaining old structures that face risks from age and earthquakes.
Temporal Fusion Transformer (TFT)
This is a sophisticated deep learning model designed to understand sequences of data over time. It learns the normal behavior of the tower's vibrations and can predict future states while estimating prediction uncertainty, which is key for spotting anomalies.
Anomaly Detection
This method involves training a system on 'normal' data so it can flag anything that significantly deviates from those learned patterns as an anomaly. In this study, deviations from expected vibration levels indicate potential structural issues or unusual events.

Terminology

Summary

Deep learning for structural health monitoring: An application to heritage structures employs recent deep learning techniques for time-series forecasting to inspect and detect anomalies in large datasets recorded during a long-term monitoring campaign on the San Frediano bell tower in Lucca The data analysis shows promising results, highlighting anomalous events like the Amatrice earthquake that happened on 24 August 2016 or the Santa Croce celebrations on 13 September 2016. p. 0

Problem Context

Heritage structures are threatened worldwide by aging, material deterioration, environmental actions, and extreme meteorological events due to climate changes, and therefore they need maintenance and restoration In Italy, this need is exacerbated by the seismic vulnerability of the territory, which has made structural health monitoring (SHM) increasingly attractive because it provides a relatively inexpensive tool for promptly assessing the structural conditions, planning, and controlling maintenance interventions The availability of high-sensitive instrumentation and extensive data sets from long-term monitoring protocols opened new issues about data analysis, particularly regarding the implementation of automatic anomaly detection and early warning procedures Exploiting artificial intelligence (AI) represents a significant opportunity in this data-driven context.

Model Architecture

The proposed approach belongs to a set of techniques that pertain to the domain of unsupervised anomaly detection Specifically, a network is trained to understand the normality patterns present in some data and then, at inference time, the more significant deviations from the learned normality patterns are considered anomalies In this work, we use a state-of-the-art deep neural network for learning the structure’s dynamics directly from data called Temporal Fusion Transformer (TFT). The TFT can also estimate the uncertainty of the prediction, which is fundamental during the anomaly detection phase. p. 2

Anomaly Detection Methodology

The system that we develop can be formalized as X(1) = TFT(x1, x2, …, xT; u1, u2, …, uT), where xt is a vector encoding the values measured from the instrumentation in the structure at time t. The model characterizes this output random variable by predicting its mean value and 1, 10, 25, 50, 75, 90, 99 percentiles. p. 2 An anomalous sample is defined by checking if the following condition is NOT satisfied: π, (2) <= xT+1 ≤ π p (X), where πp (X) represents the p-percentile of the random variable X. p. 2 In our experiments, we use p = 99. p. 2

Case Study Findings

The San Frediano bell tower in Lucca has been subjected to a long-term vibration monitoring campaign from October 2015 to November 2017 The training dataset was built by processing data via the covariance-driven Stochastic Subspace Identification technique to extract the first five natural frequencies from the tower’s response. Anomalous events detected by the TFT procedure include:

  1. The Amatrice earthquake that struck central Italy with a 6. p. 30 magnitude. p. 3

  2. The Santa Croce celebrations and their effect on tower vibrations. p. 3

  3. Vibrations induced by the swinging bells, involving the second frequency. p. 3

Conclusion and Future Work

This preliminary work showed promising results in the case study of the San Frediano bell tower in Lucca where we were able to correctly identify some important events like the Amatrice earthquake and the Santa Croce celebrations. Further investigations are necessary to confirm the good performance and reliability of the adopted method. p. 3 A comparison between anomaly detection techniques used in the literature and the TFT model will be carried out in a future work where the performance of the Transformer network will be tested on artificial damage scenarios. p. 3 It would also be interesting to characterize found anomalies by inspecting and clustering hidden representations learned by the model. p. 3

Acknowledgments

This research has been supported by the Fondazione Cassa di Risparmio di Lucca (SOUL project, 2019-2022), and by the INAROS project (CNR4C program, CUP B53D21008060008).

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Deep learning for structural health monitoring: An application to heritage structures can be improved by using a Temporal Fusion Transformer (TFT) to predict structure dynamics given sensor and environmental data from the recent past. This system can detect anomalies by checking if actual observed values fall outside the predicted confidence interval, defined as xT+1 is anomalous if it is an outlier with respect to the estimated distribution of the random variable XT+1.

  2. The improved AI system can be specifically trained to predict main natural frequencies of heritage structures starting from environmental data and recent past frequencies. This allows for proactive identification of structural issues, such as detecting anomalous events like the Amatrice earthquake that happened on 24 August 2016 or the Santa Croce celebrations on 13 September 2016.

  3. The TFT model can be conditioned on temporal fingerprints, specifically the hour (1-24), the day (1-30), and the month (1-12), to better capture periodical patterns. This enhances the system's ability to distinguish between normal operational variations and genuine structural anomalies like the vibrations induced by the swinging bells.

Related papers