New method tracks engine wear using physics and data
A physics-constrained autoencoder helps identify specific component degradation in turbofan engines.
Researchers developed a physics-constrained autoencoder to track health in turbofan engines. This method uses Health Indices (HIs) to summarize degradation from operational data.
The system uses a decoder with differentiable physical equations. These equations receive latent variables and operating conditions as inputs. Latent variables are used to represent predefined physical quantities.
The model uses system knowledge as an inductive bias in the decoder architecture. It also uses a learning bias in the loss function. This allows for component-level analysis before creating a system-level HI.
The method was tested on the N-CMAPSS dataset. It reached a reconstruction of 0.9993 and an MAE of 0.0036. The aggregated HI had a Spearman correlation of 0.9791 with the ground truth.
One model trained across four datasets could distinguish compression-side from expansion-side degradation. This was done without diagnostic labels. It also allows degradation trends to be visualized on component maps.
Key facts
- The model achieved a reconstruction of 0.9993 and an MAE of 0.0036.
- The aggregated HI shows a Spearman correlation of 0.9791 with the ground truth.
- The framework can distinguish between compression-side and expansion-side degradation.