Remaining useful lifetime estimation for discrete power electronic devices using physics-informed neural network.

Area of Science:

• Electrical Engineering
• Machine Learning
• Reliability Engineering

Background:

• Accurate Remaining Useful Lifetime (RUL) estimation for discrete power electronics is crucial for predictive maintenance and system safety.
• Conventional data-driven neural network approaches often yield unrealistic RUL estimates due to their disregard for underlying physical properties.
• Existing methods can produce erroneous predictions, such as increasing RUL estimates or incorrect terminal predictions.

Purpose of the Study:

• To enhance Recurrent Neural Network (RNN) based RUL estimation by integrating the principles of Physics-Informed Neural Networks (PINNs).
• To improve the realism and accuracy of RUL predictions for power electronic devices.
• To validate the effectiveness of PINN-enhanced RNNs using the NASA IGBT dataset.

Main Methods:

• Applied Physics-Informed Neural Networks (PINNs) to augment standard RNN architectures for RUL estimation.
• Incorporated physical constraints into the loss function of the neural networks during training.
• Evaluated performance using the NASA IGBT dataset, comparing against baseline RNN and Long Short-Term Memory (LSTM) models.

Main Results:

• Physics-informed RNN models demonstrated more realistic training behavior compared to conventional neural networks.
• Significant improvements in estimation accuracy were observed, evidenced by reductions in Mean Squared Error (MSE).
• PINN-enhanced RNN achieved an average MSE improvement of 24.7% in training and 51.3% in testing over vanilla RNN.
• PINN-enhanced LSTM showed an average MSE improvement of 15.3% in training and 13.9% in testing over baseline LSTM.

Conclusions:

• Integrating physical principles via PINNs effectively addresses the limitations of purely data-driven RUL estimation.
• PINN-enhanced RNN and LSTM models provide more accurate and reliable RUL predictions for power electronics.
• This approach enhances predictive maintenance capabilities and contributes to improved system safety and operational efficiency.