Deep Learning with Dynamically Weighted Loss Function for Sensor-Based Prognostics and Health Management.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Prognostics and Health Management (PHM)

Background:

• Deep learning is widely used in automotive and aerospace prognostics and health management (PHM).
• Existing research predominantly focuses on deep learning model architectures.
• Improvements in other aspects, like custom loss functions for PHM, are less explored.

Purpose of the Study:

• To enhance deep learning effectiveness for PHM without altering model architectures.
• To investigate the impact of custom dynamically weighted loss functions on PHM tasks.
• To address the scarcity of research on improving deep learning components beyond architecture.

Main Methods:

• Implemented and evaluated two dynamically weighted loss functions: a novel weighting mechanism and a focal loss function.
• Applied these loss functions to four deep learning architectures: feedforward neural networks, 1D CNNs, Bi-GRU, and Bi-LSTM.
• Tested models on NASA's aero-propulsion system data and Scania truck air pressure system failure data.

Main Results:

• Dynamically weighted loss functions demonstrated significant improvements in remaining useful life (RUL) prediction.
• These custom loss functions also led to enhanced fault detection rates compared to standard loss functions.
• The proposed weighting mechanism and focal loss effectively guided model learning towards more accurate predictions.

Conclusions:

• Dynamically weighted loss functions offer a viable method to boost deep learning performance in PHM.
• Customizing loss functions is a promising avenue for advancing PHM capabilities.
• This approach provides performance gains without requiring complex architectural modifications.