How Few Sensors Are Enough? Toward Efficient Sensor Selection for Remaining Useful Life Prediction

The number of sensors used in machine health prognostics directly affects computational and deployment costs of predictive maintenance (PdM) systems. This work investigates whether the sensor set required for remaining useful life (RUL) prediction can be reduced without compromising performance, and whether a model-derived measure of sensor importance corresponds to empirical sensor ranking, reducing the need for exhaustive ablation in practice. Using the FD001 and FD003 subsets of the NASA C-MAPSS turbofan degradation dataset, sensor sufficiency is assessed via sequential backward elimination against a validation-RMSE tolerance criterion, and compared against a sigmoid feature-gating attention mechanism trained jointly with an LSTM-based RUL predictor. Sensor sets were reduced by $60\%$ and $31\%$ for FD001 and FD003 respectively while preserving test-set RMSE within the variability range of the full-sensor baseline, demonstrating the potential for reduced sensing and deployment costs using only existing time-series data. While the two ranking methods agreed on three of the five most important sensors in both datasets, overall rank agreement remained weak, indicating that attention alone should not yet replace empirical validation as a low-cost sensor-selection strategy. The achievable reduction varied with fault-mode complexity, showing that such strategies must be validated on a system-specific basis. These findings represent an important step toward more accessible PdM for industries with limited sensing infrastructure.