Embedded Files
Problems of end-to-end deep learning methods on Engineering Problems (Target on PHM)
1. Low Generation Performance
The trained DL model is susceptible to noise or external condition change
2. Low Interpretation
DL model is black-box, so the reason why the model gets specific results is not explained
3. Lack of Physical Consistency
The result of the DL model is sometimes inconsistent with physical knowledge since it only focuses on mathematical errors.
Domain Knowledge that can be combined with Deep Learning models
Definition of Domain Knowledge-Informed Neural Networks and Their Benefits
Incorporating domain knowledge into deep learning aims to increase generalization performance, physical consistency, and interpretability than conventional end-to-end deep learning approaches.
1. Increased Generalization Performance
ex) A DL model also trained with generated simulated samples of different conditions, in addition to the real samples of specific conditions, can have better performance on various conditions.
2. Physically Reasonable Results
ex) A DL model trained with additional physical loss terms can adhere more to the physics law(ex. PINN).
3. Increased Interpretability
ex) A DL model with a 2D imaging processing part(ex. time-frequency map) can have better interpretation than just using signal input.
My DINN Research
Page updated
Google Sites
Report abuse