Prediction of metallic erosion concentration from engine exhaust plume spectra via deep learning

Accurate and quantitative detection of metallic species in the exhaust plumes of liquid rocket engines is essential for assessing component wear and enabling early fault diagnosis. Aiming to establish a deep‑learning‑based inversion model for plume metal concentrations, this paper presents three main contributions. First, a corrected line‑by‑line (LBL) spectral simulation method is developed. The correction model, which accounts for self‑absorption effects, is derived from empirical relationship observed in experimental data. Although calibrated using copper (Cu) measurements, the correction model is experimentally shown to be equally applicable to iron (Fe) and nickel (Ni). Second, by combining the corrected LBL simulations with a minimal set of experimental measurements (e.g., only two concentration levels per metal), the proposed approach generates a comprehensive and accurate training dataset without extensive experimental effort. Third, the inversion model itself is a fully connected feedforward neural network whose hyperparameters are optimized by a multi‑strategy Grey Wolf Optimizer (MS‑GWO). Experimental validation under unseen operating conditions demonstrates that the model achieves prediction accuracies exceeding 90% for Cu, Fe, and Ni, thereby confirming its strong generalization capability. The performance of the proposed MS‑GWO‑FNN model is further benchmarked against a genetic‑algorithm‑optimized back‑propagation (GA‑BP) network, a standard FNN, and a radial basis function neural network (RBFNN). The comparative results demonstrate that the MS‑GWO‑FNN achieves superior prediction accuracy relative to the other three models, confirming its strong potential to meet the stringent requirements of plume‑based health monitoring and diagnostics.