PCA–Mahalanobis Anomaly Detection in CZT Gamma Spectra

Abstract

Reliable detection of radioactive sources in gamma-ray spectra is important for nuclear security, environmental monitoring, and radiation protection. This work presents an experimental proof-of-concept anomaly-detection approach for low-count cadmium zinc telluride (CZT) gamma-ray spectra, combining the Anscombe variance-stabilizing transform, principal component analysis (PCA), and Mahalanobis distance to detect deviations from natural background using full spectral-shape information. Measurements were performed with a compact 0.5cm3 CZT detector and a custom Python-based data acquisition system. Background and 137Cs spectra were acquired as 1024-channel, 10s measurements and combined into moving sums for the principal analysis. The anomaly-detection calculations were performed directly in analog-to-digital-converter (ADC) channel space; the energy calibration was used only for detector characterization. Sensitivity studies showed that aggressive PCA truncation strongly reduced source sensitivity and that increasing the integration duration improved detection at the cost of increased sensitivity to temporal background variation. For the 60 s, 400-component configuration, the mean source-detection probability was 70.4% and the weakest-condition detection probability was 14.5%; the independent-background threshold-crossing frequency was 13.4/day, equal to the calibration-background crossing frequency within the finite resolution of the available data. Requiring at least two consecutive threshold crossings eliminated all false alarms observed in the available independent background sequence. The results demonstrate the feasibility of PCA–Mahalanobis scoring as a background-only proof-of-concept anomaly detector for low-count compact-CZT spectra under the investigated 137Cs measurement conditions, while also showing the importance of PCA dimensionality, integration time, background stability, and temporal persistence.