A Similarity-Based Ensemble Framework for Remaining Useful Life Prediction of a Subway Door System

Abstract

Remaining useful life prediction is challenging when only a small number of run-to-failure trajectories are available and the evaluation emphasizes early prognostic accuracy. This paper addresses the PHME 2026 Data Challenge on remaining useful life prediction for a subway door system. We propose a similarity ensemble that characterizes each operating cycle with statistical features, captures position-feedback degradation behavior, and estimates the failure cycle by comparing partial trajectories with historical run-to-failure cases. The resulting estimates are fused and constrained using model-disagreement and operating-condition information, and the final remaining-life sequence is generated as a monotonic countdown from the estimated failure cycle. The method is evaluated using stratified cross-validation on the official training set and further assessed through the challenge leaderboard submission. Results show that the proposed framework provides accurate and stable predictions under limited-data conditions, achieving a validation score of 0.9467 on observable-degradation scenarios and an official leaderboard score of 0.9963. The study demonstrates that constrained similarity-based failure-cycle estimation is an effective and interpretable strategy for small-sample prognostics.