AttnPINN for RUL Estimation

This repository includes the code and data for the paper "Remaining useful life with self-attention assisted physics-informed neural network"

Dataset

This paper just evaluated the FD004 sub-dataset and the whole dataset can be found here.

Abstract

Remaining useful life (RUL) prediction as the key technique of prognostics and health management (PHM) has been extensively investigated. The application of data-driven methods in RUL prediction has advanced greatly in recent years. However, a large number of model parameters, low prediction accuracy, and lack of interpretability of prediction results are common problems of current data-driven methods. In this paper, we propose a Physics-Informed Neural Networks (PINNs) with Self-Attention mechanism-based hybrid framework for aircraft engine RUL prognostics. Specifically, the self-attention mechanism is employed to learn the differences and interactions between features, and reasonably map high-dimensional features to low-dimensional spaces. Subsequently, PINN is utilized to regularize the end-to-end prediction network, which maps features to RUL. The RUL prediction framework termed AttnPINN has verified its superiority on the Commercial Modular AeroPropulsion System Simulation (C-MAPSS) dataset. It achieves state-of-the-art prediction performance with a small number of parameters, resulting in computation-light features. Furthermore, its prediction results are highly interpretable and can accurately predict failure modes, thereby enabling precise predictive maintenance.

Configuration

• matplotlib==3.3.2
• numpy==1.21.6
• scikit_learn==1.0.2
• torch==1.11.0
• torchsummary==1.5.1

If you want to install the required environments one by one, you can copy the following codes:

pip install matplotlib==3.3.2
pip install numpy==1.21.6
pip install scikit_learn==1.0.2
pip install torch==1.11.0
pip install torchsummary==1.5.1

or use this:

pip install -r requirements.txt

Quick Start

Running the project with the following code:

python main.py

In main.py, it includes training, predicting and drawing functions.

By default, only predicting function will run and the output will be:

Test_RMSE: 18.37,   Score: 2058.5

If you want to train the model by yourself:hammer::hammer:, you can uncomment the train function in main.py.

#pinn.train(1000) => pinn.train(1000)

And then, the output will be:

It: 0,   Valid_RUL_RMSE: 100.92
It: 1,   Valid_RUL_RMSE: 99.87
It: 2,   Valid_RUL_RMSE: 40.89
It: 3,   Valid_RUL_RMSE: 40.76
It: 4,   Valid_RUL_RMSE: 40.74
It: 5,   Valid_RUL_RMSE: 40.74
It: 6,   Valid_RUL_RMSE: 35.48
It: 7,   Valid_RUL_RMSE: 20.47
It: 8,   Valid_RUL_RMSE: 18.70
It: 9,   Valid_RUL_RMSE: 18.27
···

Comparisons with State-of-the-art Methods

Method	RMSE	Score	Parameters	FLOPs
DCNN(Li et al., 2018)	23.31	12466	35.05k	174.76k
RNN-Autoencoder(Yu et al.. 2020)	22.15	2901	378.0K	N/A
GCU-Transformer(Mo et al.,2021)	24.86	N/A	399.7K	393.39
Double attention-Transformer(Liu et al., 2022)	19.86	1741	N/A	N/A
e-RULENet(Natsumeda, 2022)	20.80	1554	32.3K	N/A
CNN-BiLSTM-3DAttention(You et al., 2023)	20.24	1710	151.9k	170.3k
MTSTAN(Li et al., 2023)	18.85	1446	N/A	N/A
PINN-PHM(Cofre-Martel et al., 2021)	25.58	N/A	1,066	N/A
AttnPINN(proposed framework)	18.37	2059	2,260	1728

Star History

Cite Repository

Please, cite the paper using:

@article{liao2023remaining,
  title={Remaining useful life with self-attention assisted physics-informed neural network},
  author={Liao, Xinyuan and Chen, Shaowei and Wen, Pengfei and Zhao, Shuai},
  journal={Advanced Engineering Informatics},
  volume={58},
  pages={102195},
  year={2023},
  publisher={Elsevier}
}