Scientific Machine Learning for prediction and monitoring of structural integrity of bridges

Problem Context

New AI capabilities are needed enabling AI systems to be 1) flexible: adaptive to a complex and changing world, 2) trustworthy: the user is in control, the system is robust and reliable, human values are respected, and the system is transparent and explainable to the user, and 3) engineered for lifetime validity. We do this by incorporating physics knowledge and models into AI models, resulting in a hybrid AI. This class of methodologies are also referred to as scientific machine learning (SML).

Solution

The main objective is to develop an adaptive prediction system that supports a proactive maintenance approach of bridges and other infrastructure objects by combining physics models with machine learning to enable real-time, accurate and transparent decision support. The aim is to develop generic components that can benefit any application – inside and outside TNO - where physical models and measurement data interact. This can yearly save billions of euros to the Dutch government in costly renovations and operational risks. The newly developed technology will result in a decision-making workflow that differs from the current practice, where only very specialised, inflexible and costly high-fidelity physical models are used. Instead, the envisioned solution aims at adaptive hybrid models and will include continuous monitoring data.

Results

We started by a period of state-of-the-art research by means of recent literature. Based on that, we moved on to successfully create a scientific machine learning model that can localize the damage for a truss bridge and predict damage levels of the damaged elements. The model is based on a convolutional neural network, where additional scientific domain knowledge was incorporated. In addition to the sensor data, an extra physics term in the loss function was added, which evaluates the predicted quantities and verifies the output of an Final Element model. In addition, to increase the model transparency, uncertainty quantification was added. Lastly, The validation of the models was also performed with experimental data of a scale model bridge.