A dual-task prognosis framework for thermocouple sensor remaining useful life prediction

Thermocouples are widely used temperature sensors in industrial applications, and their performance directly affects the accuracy of temperature measurements. Accurate remaining useful life (RUL) prediction of thermocouples is essential for reliable maintenance and replacement scheduling. However, real-world thermocouple degradation data often exhibit low sampling frequency and limited observable variables. To address these challenges, this paper proposes a dual-task learning framework that integrates empirical mode decomposition with a novel dynamic interactive convolution (DIC) module for multi-scale temporal feature extraction. A multi-gate mixture-of-experts (MMoE) network is adopted to jointly perform first prediction time (FPT) detection and RUL estimation, enabling both shared and task-specific representation learning. Additionally, an automated FPT labeling method is introduced to replace traditional manual thresholding. To validate the approach, this study collects run-to-failure lifecycle data through accelerated degradation experiments. Experimental results indicate that the proposed DIC-MMoE framework reduces RUL prediction error by 16.37% and improves FPT detection accuracy by 0.19% compared to existing single-task and multi-task baselines, demonstrating its effectiveness.