Adaptive Neural Network Architectures for Predictive Maintenance and Remaining Useful Life Estimation in Smart Manufacturing
Keywords:
Predictive Maintenance, Remaining Useful Life, Adaptive Neural Networks, Operating Conditions, Uncertainty CalibrationAbstract
Adaptive prognostic models remain difficult to trust when operating conditions and sensor quality change together. This study examines a compact neural architecture for remaining useful life estimation that separates conditional computation from supervised model updating. Three neural experts process causal sensor summaries over 8, 16, and 32 cycles, while a context gate assigns their weights from load and measurement availability. A restricted update changes only the gate and expert output layers, and independent split conformal calibration supplies prediction intervals. The architecture is evaluated through an executed, reproducible synthetic experiment comprising five independently generated fleets, each containing 900 assets. Four configurations isolate the effects of gating and updating; no industrial measurements or public benchmark results are presented as experimental evidence. Under the simulated target shift, the updated gated model achieves a mean root mean squared error of 24.23 cycles, compared with 31.94 for its frozen counterpart and 25.29 for the updated uniform-weight configuration. Its 4.2% advantage over the latter is modest and not consistent across all repetitions. Nominal 90% intervals attain 90.9% mean target coverage, but coverage falls to 73.6% after additional sensor corruption. The uniform-weight updated model also performs better under this stress. Only 174 of 2,550 parameters require updating, although adaptation and calibration together still consume 160 labelled target assets. These results support controlled adaptation as a useful engineering option while exposing its dependence on representative labels and calibration conditions. The contribution is a transparent architectural experiment and validation protocol, rather than evidence of deployment readiness or superiority over established prognostic models.