Incremental learning algorithm for anomaly detection applied to computed tomography scans in nuclear industry

Abstract

During routine nuclear power plant (NPP) inspection, each maintenance tool is inspected manually before and after use on a nuclear reactor. This could result in long inspection duration (up to months), time, and resource wastage. To address this, an automated tool inspection process using a classification-based supervised anomaly detection technique is employed to categorize the CT scan of the NPP tool as defective (with missing tool parts) or not (defect-free). Furthermore, the incremental learning (IL) concept has been introduced for supervised anomaly detection and is suitable for data-restricted applications. Existing IL approaches have employed ML techniques such as Naive Bayes or proximity measures such as nearest neighbors on numeric 1D datasets and for intrusion detection. In this research, a new soft thresholding-based algorithm that can enhance model prediction in existing IL frameworks and ensure stable training towards the desired prediction accuracy for supervised anomaly detection on 2D data is proposed.