Fault detection in thermocouples: Unveiling anomalies with machine learning and signal processing

Abstract

Reliable data acquisition from installed sensors is crucial for ensuring operational efficiency and safety in industrial settings. Early detection of sensor anomalies is particularly vital in high-integrity applications such as avionics, nuclear reactors, and associated fuel recycling plants, where data reliability directly impacts process and personnel safety. Thermocouples (TCs) are commonly used in critical temperature measurement applications due to their robustness and long history of dependable performance. This paper proposes a data-driven technique to detect TC sheath failure as part of the Operator Support System (OSS) in operating plants, alarm generation, decision support, and predictive maintenance. Additionally, an accelerated aging setup is proposed to simulate sheath failure in TCs and assess its impact on performance characteristics in a controlled environment mimicking the dissolver stage of the Plutonium Uranium Reduction Extraction (PUREX) process in nuclear fuel reprocessing. Our in-situ failure detection approach introduces an application of Empirical Mode Decomposition (EMD) as a data-driven technique to extract sensor noise from true measurement data. The statistical features of the extracted noise signal are then combined with machine learning (ML) based decision-making for early sheath failure detection. This approach is specifically designed for in-situ detection of sheath failure, a primary cause of TC malfunction in corrosive environments. The effectiveness of the proposed method is demonstrated using experimental data from accelerated testing of faulty TCs in a controlled environment. Results show that K-Nearest Neighbor (KNN) and Random Forest (RF) classifiers achieved over 96% classification accuracy under all experimental conditions.