Structural health monitoring in a population of similar structures with self-supervised learning: a two-stage approach for enhanced damage detection and model tuning

This article introduces a novel self-supervised learning framework for structural health monitoring (SHM) that utilizes power spectral density (PSD) data to detect structural damage. The approach leverages PSD data to monitor structural health comprehensively, without relying on engineered features or domain-specific expertise. Designed to accommodate data from multiple similar structures, the framework trains a model using data from a population of similar structures, rather than focusing on domain-specific adaptation. This method effectively handles environmental variability and the complexity inherent in high-dimensional PSD data. A key innovation in this work is the introduction of a method for evaluating and tuning the model without needing actual damage data. The framework follows a two-stage training and evaluation process: in the first stage, a neural network is trained to classify PSD data based on the structures from which the data originates. During inference, the model’s classification confidence serves as an anomaly index, which forms the foundation of the proposed damage detection system (DDS). In the second stage, virtual anomalies (VAs) are introduced as data augmentation, modifying healthy data to evaluate the model’s ability to detect deviations that suggest potential damage. A key finding is that the DDS’s ability to detect VAs correlates with its ability to detect real structural damage, allowing for future validation without real damage data. Heatmap visualizations derived from VAs highlight the frequency regions where the model is most sensitive. The DDS is applicable in scenarios where a single accelerometer is used on similar structures (e.g., wind turbines, transmission towers) or multiple accelerometers are deployed on a single structure. The framework’s effectiveness is demonstrated in two case studies: a simulated population of 8-degree-of-freedom mechanical systems and real-world data from a transmission tower instrumented with four accelerometers under controlled damage conditions.