Turbofan engine health status prediction with heterogeneous ensemble deep neural networks

Aircraft turbofan engines undergo gradual performance degradation that, if left undetected, can compromise flight safety and increase maintenance costs. In this study, we propose a heterogeneous ensemble deep neural network (HEDNN) for multi-class health status prediction of real-life turbofan engines. Our methodology begins with the collection of 43,492 real operational records (2012–2024) from a dual-spool, variable geometry turbofan, each run sampled at a fixed throttle setting. We preprocess raw sensor outputs using a robust z-score (median absolute deviation) filter to remove outliers, apply a 35-sample moving median imputation to address 1,812 missing temperature entries, and eliminate any non-numeric artifacts. The cleaned dataset consisting of 97 predictive features after FADEC upgrade harmonization is partitioned into training (80%, n = 34 794) and testing (20%, n = 8 698) sets, and labeled into three health classes (no/low degradation, mid-life degradation, and advanced degradation) built upon performance margin (PMAR) and specific fuel consumption (SMAR) indices. We engineer three base learners: bidirectional LSTM (128 hidden units), one-dimensional CNN (11 convolution normalization ELU blocks), and BiGRU (four GRU layers, custom flip/concatenate stages), each trained for up to 60 epochs, batch size 16, using the Adam optimizer with piecewise learning rate decay (initial η = 0.01). Individual networks produce class probability vectors that are fused in a shallow fully connected ensemble, finalized by softmax classification. Performance is assessed via RMSE, MAPE, overall accuracy, precision, recall, F1 score, and multi-class AUC. The HEDNN achieves RMSE = 0.3421 (33–41% lower than any single model), MAPE = 4.78% (60–65% reduction), and accuracy = 95.32%, with F1 = 0.928 and mean AUC = 0.777. In contrast, the best individual learner (BiLSTM) reaches RMSE = 0.5149, MAPE = 12.17%, accuracy = 82.44%, and AUC = 0.717. These results demonstrate that our heterogeneous ensemble markedly outperforms standalone architectures, offering a robust, data-driven prognostic tool for predictive maintenance of turbofan fleets.