Partially Supervised Learning for Data-Driven Structural Health Monitoring

The cost of labelling data by engineer inspections remains a significant issue for performance and health monitoring. In many cases, this is because the actual data annotation process is expensive (e.g. non-destructive testing) or it is simply infeasible to label all the measurements (e.g. lack of access). Often, however, it is possible to provide a small number of budget-restricted labels, to describe the measurements. In these scenarios, methods for partially supervised learning are proposed. Active learning, semi-supervised learning, and transfer learning are summarised here—demonstrated with simulated monitoring examples. Each family of algorithms is shown to significantly improve conventional methods for data-driven monitoring.