Damage detection in guided wave structural health monitoring using Gaussian process regression

Damage detection using permanently installed guided wave sensors typically involves identifying changes in A-scan data from a nominally defect-free state. A well-recognised challenge is the influence of changing operational and environmental conditions which can lead to complex evolution of the A-scan data which may mask the presence of damage. This paper is based on the observation that the amplitude at each location within the A-scan has a unique functional dependence with temperature and a unique associated uncertainty in the functional dependence. The uncertainty in the functional dependence is based on both the repeatability of the measurements and the amount of data available for characterisation. The fact that each location in an A-scan has a different uncertainty leads to the concept of an adaptive threshold. In contrast to typical damage detection that utilises a flat threshold for the whole A-scan, an adaptive threshold will be lower in locations of greater predictability and rise in locations of poorer repeatability. To achieve the aim of an adaptive threshold, the amplitude–temperature dependence and associated uncertainty is characterised for each data-point in an A-scan based on a training set of defect-free data. The inverse process can then be performed to generate a new reference A-Scan with associated confidence bounds for a given temperature. The confidence bounds act as the adaptive threshold; if a new measurement passes outside the confidence bounds then it indicates statistically significant difference between the new measurement and the training set. To demonstrate the method, pitch-catch data between two sensors of a pipe-bend over 5 months and a temperature range of approximately 50°C is used. To demonstrate the damage detection capability in a controlled way synthetic damage is added to the A-scans to produce test data. The performance is compared to Optimal baseline subtraction (OBS). In this example, the performance of the Gaussian process regression adaptive threshold method exceeds OBS when the training set exceeds around 40 measurements.