CD-PINODE code and dataset

This repository contains the source code and synthetic dataset supporting the study "CD-PINODE: A Physics-Informed Neural Ordinary Differential Equation Framework for Joint Fault Classification and Remaining Useful Life Estimation in Turbofan Engines." Five prognostics models are implemented and benchmarked on NASA C-MAPSS synthetic sensor data across two co-dependent tasks: (1) binary fault classification (HPC compressor degradation and FAN blade erosion) and (2) continuous Remaining Useful Life (RUL) regression.

Each model is evaluated on both an in-distribution test set and a harder out-of-distribution (OOD) test set simulating unseen degradation regimes. The proposed model, CD-PINODE v4, integrates a Neural Ordinary Differential Equation backbone with Brayton-cycle physical constraints, a distance correlation regularizer (dcorr: 0.0471 → 0.0034), and a joint multi-task loss. It achieves in-distribution RUL RMSE of 22.10 and OOD RMSE of 35.05, with OOD HPC balanced accuracy of 84.06% and FAN balanced accuracy of 71.82% using only 61,286 trainable parameters.