Mechanics-Informed Autoencoder Enables Automated Detection and Localization of Unforeseen Structural Damage

TL;DR

Structural health monitoring needs scalable, passive, inexpensive methods capable of detecting unseen damage without extensive labeled data. MIDAS integrates inexpensive sensors, edge data compression, and a Mechanics-Informed Autoencoder (MIAE) that learns a structure-specific intact baseline from undamaged responses and detects damage by comparing instantaneous sensor reconstructions to the baseline. By encoding pairwise mechanical relations between sensors through a mechanics loss and a weight matrix, MIAE achieves earlier detection and improved localization, outperforming baseline methods with up to 35% gains for minor damages. The approach is validated numerically (gusset plate) and experimentally on gusset-plate and beam-column structures, showing robustness to sensor count, noise, and temperature variations and enabling deploy-and-forget SHM with reduced inspection costs. This work advances practical, scalable, autonomous SHM for large civil infrastructures.

Abstract

Structural health monitoring (SHM) ensures the safety and longevity of structures like buildings and bridges. As the volume and scale of structures and the impact of their failure continue to grow, there is a dire need for SHM techniques that are scalable, inexpensive, can operate passively without human intervention, and are customized for each mechanical structure without the need for complex baseline models. We present MIDAS, a novel "deploy-and-forget" approach for automated detection and localization of damage in structures. It is a synergistic integration of entirely passive measurements from inexpensive sensors, data compression, and a mechanics-informed autoencoder. Once deployed, MIDAS continuously learns and adapts a bespoke baseline model for each structure, learning from its undamaged state's response characteristics. After learning from just 3 hours of data, it can autonomously detect and localize different types of unforeseen damage. Results from numerical simulations and experiments indicate that incorporating the mechanical characteristics into the autoencoder allows for up to a 35% improvement in the detection and localization of minor damage over a standard autoencoder. Our approach holds significant promise for reducing human intervention and inspection costs while enabling proactive and preventive maintenance strategies. This will extend the lifespan, reliability, and sustainability of civil infrastructures.

Figures (7)

• Figure 1: Overview of MIDAS, the automated structural damage detection and localization framework. Raw structural response data from the sensors are compressed, and MIAE is trained purely on the response from the structure's undamaged state. No additional information is leveraged besides the pairwise mechanical relations between the strain responses. Once trained, the distribution of reconstruction errors between the network's input and output on the training data serves as a reference representation of an intact structure's response. After deployment, the trained model processes data from the sensors, and resultant reconstruction errors are compared to the reference error distribution to detect and localize potential damage. An observable shift in reconstruction errors (top right) highlights the detection of damage. The incorporated mechanical knowledge notably amplifies the distribution shift, significantly enhancing damage detection at an early stage. Sensor-wise error comparisons are interpolated (heatmaps at the bottom right) to localize anomalies representing the onset of damage
• Figure 2: Damage detection for a cracked gusset plate.a. Finite element mesh of an intact plate, boundary conditions, and loading. b. Sensor arrangement with labels. c. A typical cracked plate and its meshing. Different crack lengths represent damage progression. d. Distributions of reconstruction errors of the structure from its undamaged reference and damaged states. As the crack progresses (three different crack lengths), the error distribution shifts to the right and becomes more distinct from the undamaged reference. e. Damage detection performance as the crack length increases. MIAE outperforms the baseline autoencoder in all five metrics, especially in the early stages of damage emergence. f. Compared to baseline anomaly detection methods, MIAE exhibits the best detection accuracy in the undamaged scenario and consistently achieves higher damage detection rates across all the evaluated metrics and crack lengths.
• Figure 3: Damage localization for a cracked gusset plate. We consider different crack lengths: intact (0$cm$), medium (2$cm$), large (4$cm$), and very-large (6$cm$). MIAE localizes cracks at an earlier damage stage than prior unsupervised methods. a. Damage score maps for different damage scenarios. A high damage score (peak values in yellow) at one or more sensors near the crack indicates successful localization. MIAE can localize the crack earlier (at a small crack length) than SPIRIT and autoencoder. b. Damage localization accuracy from an extensive analysis of 37 different crack scenarios. The $y$-axis refers to the percentage of cases where damage was successfully localized. Compared to autoencoder and SPIRIT, MIAE has a higher localization accuracy across all crack lengths (e.g., 35% better localization for 2$cm$ long cracks), demonstrating its ability to localize cracks earlier than the baseline approaches.
• Figure 4: Damage detection and localization under sensor and temperature variations.a. Damage detection performance as the number of sensors varies. b. Comparison of localization accuracy between MIAE and autoencoder with four sensors for two different crack scenarios. MIAE’s peak damage score is closer to the true crack location in both cases. c. Comparison of damage localization accuracy with four sensors as crack length increases. MIAE outperforms the baseline approaches. d. Damage detection performance with noisy (0.5% additive Gaussian noise) sensor data. e. Damage detection performance was evaluated at two different temperatures.
• Figure 5: Laboratory experiment on a steel plate structure.a. Two types of damage are introduced sequentially (a crack and boundary condition variation). The crack is located in the middle of the plate, and the second damage was introduced by loosening the bolt connections. Under loading, the connection of the plate loosens, thus mimicking damage progression. b. Crack localization results with 27 and 4 sensors, respectively. When using all 27 sensors, MIAE accurately delineates the crack region with a high damage score (yellow region) around the crack tips, outperforming the autoencoder. SPIRIT fails to localize the damage in both setups. c. Localization for bolt loosening damage under loading. MIAE correctly localizes the damage at the bottom plate connection in the early loading stage (damage progression). d. Localization performance for boundary condition variation. Only MIAE can localize the damage early. As the crack size increases, both the autoencoder and SPIRIT gradually succeed in localizing it. e. Damage differentiation through compressed sensor data $\mu$ and $\sigma$. While $\mu$ is more sensitive to boundary condition changes, $\sigma$ responds more to cracks in the structure.