Gated Residual End-of-Life Prediction for Variable-Prefix RUL in the PHME2026 Data Challenge

Abstract

This paper presents a gated residual end-of-life (EOL) method for variable-prefix remaining useful life (RUL) prediction in the PHME 2026 Data Challenge. The task is to estimate the lifetime of partially observed subway-door actuation runs under hidden operating conditions. We frame the problem as EOL estimation: the model first predicts the cycle at which failure occurs and then derives RUL by subtracting the current cycle index. This gives prefixes of different lengths a common lifetime scale. The approach uses engineered prefix features that summarize electrical load, position behavior, shock events, trends, and source-model predictions. Several source models provide an initial EOL prior, while disagreement among them serves as an uncertainty signal. Rather than replacing this prior with one unconstrained predictor, the method refines it through bounded residual corrections. To keep the prediction stable, the method first forms a conservative anchor and calibrates it to the inference setting using released test measurement features. Two bounded specialists then propose residual corrections: one from similar training prefixes and one from shock, position, and current signal evidence. When these specialists disagree, a clipped gate controls how much each correction affects the final EOL estimate. Local diagnostics indicate where the source prior is sufficient, where corrections add value, and where difficult prefixes remain uncertain.