Frequency-Aware Virtual Instrument for Remaining Useful Life Estimation: Soft-Sensing with Allan Deviation-Based Stability and Noise Robustness

Accurate remaining useful life (RUL) estimation from multivariate industrial sensor streams can be undermined by high-frequency measurement noise and short-term fluctuations. This paper proposes a frequency-aware virtual instrument (soft sensor) that applies lightweight spectral low-pass conditioning as a front-end before a compact LSTM regressor. Experiments on NASA CMAPSS FD001 show that conditioning improves accuracy over an LSTM trained on raw signals (RMSE 13.64 vs. 14.72; MAE 9.76 vs. 10.36) and outperforms CNN and CNN-LSTM baselines under the same preprocessing and windowing protocol. Beyond point accuracy, we evaluate measurement-oriented reliability using Allan deviation (ADEV) on an example sensor channel (s11) and on prediction residuals, indicating improved short-term stability after conditioning. A noise-injection study (Gaussian noise added before conditioning, σ=0.12) further shows markedly smaller error degradation for the conditioned model (ΔRMSE +0.02 vs. +0.23). Overall, frequency-aware conditioning provides a simple and effective layer for more accurate, stable, and noise-robust RUL soft sensing.