This paper proposes a defect depth inversion method for bimetallic pipelines based on BMA ensemble learning. A feature dataset is constructed by integrating multi-frequency magnetic flux leakage and eddy current signals. The method combines PLO-optimized DHKELM and polynomial regression models as base learners, and fuses their outputs using Bayesian Model Averaging. Experimental results show that the proposed approach effectively improves inversion accuracy by balancing trend modeling and nonlinear feature extraction. The model outperforms single learners in terms of MAPE and RMSE metrics, demonstrating strong applicability in complex electromagnetic inspection scenarios.