Defect detection using an optimized and innovative processing technique of thermography images

The vision technology and in particular the thermography testing (TT) is a rapid developing NDE method within the nuclear industry. This paper proposes a new approach based on TT and anomaly detection algorithms for the detection of defects in nuclear components. The novelty of the approach consists of using a processing method, originally developed for multiand hyperspectral imagery, for the automatic and unsupervised processing of thermography images. Practicability of the lock-in and pulsed thermography (LT and PT, respectively) are experimentally investigated by using reference samples containing different surface and sub-surface anomalies such as open cracks, and closed notches with different sizes and depths. The heating is carried out with Eddy current approach. The generated Eddy-current is launched on the specimen for different time periods and frequencies. Both, heating and cooling parts of the temporal signals are used. Principal component thermography is used to reduce the data space dimension of the acquired thermal sequences, and thus permits the decrease of the processing time. After the reduction of the data space dimension, anomaly detection algorithms are applied on the reduced data cubes. The influence of the size of the reduced data spaces on the anomalies detection is studied using the false alarm rate as evaluation criterion of the detection results obtained from two anomaly detection algorithms: the well-known Reed and Xiaoli Yu detector (RX) and a spatially adaptive version, the regularized adaptive RX (RARX). The investigations show that; in the case where the targets are small and have significant temperature values, their signal spaces are kept after the reduction of the data cube dimensionality; the optimal false alarm rates are obtained when 80 – 86 % of the variance proportion of the projected data is considered. Indeed, with a detection of 80 – 90 % of the anomalies, we obtain optimal false alarm rates with a dimensionality which does not exceed 10 components, for both surface and subsurface defects.