Gas turbine fault probability prediction method based on classification-regression coupled model

Accurate gas-turbine risk assessment requires the error introduced by sparse compressor maps to be linked to the uncertainty of downstream fault decisions. This study proposes a two-stage framework in which a simulated-annealing and adaptive-discovery-probability cuckoo search initializes a back-propagation neural network (SA-ADP-CS-BPNN) for compressor-map reconstruction, and a Bayesian-optimized multilayer perceptron (BMLP) is softly coupled with mode-specific radial-basis-function regressors for conditional fault-risk estimation. Upstream evaluation used 840 points from two published normalized compressor maps; downstream evaluation used 400 independent 70-cycle trajectories generated by a component-level SGT5-4000F MATLAB/Simulink model, while public NASA C-MAPSS records served only to set degradation and sensor-noise ranges. Each base trajectory was rerun under 100 perturbations and split at 60:20:20 before window extraction, with tuning restricted to training and validation groups. Across 30 random seeds, the full upstream model reduces RMSE from 0.135 ± 0.007 to 0.053 ± 0.003 relative to BPNN. The BMLP obtains 96.08 ± 1.12% classification accuracy, and the complete probability pipeline achieves MAE 0.046 ± 0.003, RMSE 0.061 ± 0.004, Brier score 0.117 ± 0.002 and expected calibration error 0.028 ± 0.005. Ablation and end-to-end bridge tests show that both map reconstruction and mode-specific probability mapping contribute to the gain. The framework therefore provides a compact, calibrated solution for condition-based maintenance under limited fault data.