DLformer: A Dynamic Length Transformer-Based Network for Efficient Feature Representation in Remaining Useful Life...

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Remaining useful life (RUL) prediction is vital for system maintenance and cost reduction.
• Deep learning models for RUL prediction face computational challenges, limiting deployment on resource-constrained platforms.
• Current methods often use fixed sequence lengths, which may not be optimal for varying data characteristics.

Purpose of the Study:

• To develop a more efficient representation learning method for RUL prediction.
• To address the high computational cost associated with processing long sequences in RUL models.
• To enable adaptive sequence length processing for improved efficiency and accuracy.

Main Methods:

• Proposing a dynamic length transformer (DLformer) that adaptively learns sequence representations of varying lengths.
• Implementing a feature reuse mechanism to reduce redundant computations.
• Designing a confidence strategy for dynamic feature representation and result interpretation.

Main Results:

• DLformer achieves up to a 90% increase in inference speed.
• Model accuracy degradation is less than 5% compared to traditional methods.
• The dynamic architecture enhances model interpretability by highlighting activated components.

Conclusions:

• The DLformer offers a computationally efficient solution for RUL prediction, suitable for low-compute platforms.
• Adaptive sequence length processing is a viable strategy for improving the performance of RUL prediction models.
• The proposed method balances inference speed and prediction accuracy effectively.