Multi-Time Resolution Ensemble LSTMs for Enhanced Feature Extraction in High-Rate Time Series

Area of Science:

• Engineering
• Machine Learning
• Dynamic Systems Analysis

Background:

• High-rate dynamic systems (e.g., airbag systems, hypersonic vehicles) require accurate modeling for critical functions.
• Physical modeling tools struggle with high-rate systems due to uncertainties, non-stationarities, and unmodeled dynamics.
• Fast and accurate predictive models are essential for ensuring performance in rapidly changing environments.

Purpose of the Study:

• To develop and validate a deep learning algorithm for modeling and predicting the response of high-rate dynamic systems.
• To address the limitations of physical modeling in handling complex dynamic system characteristics.
• To enable real-time prediction capabilities for critical high-rate applications.

Main Methods:

• Utilized an ensemble of concurrently trained short-sequence Long Short-Term Memory (LSTM) cells.
• Implemented a multi-rate sampler to individually select input spaces for each LSTM cell based on local dynamics.
• Employed the embedding theorem for extracting local dynamics from non-stationary time series.
• Validated the algorithm using experimental data from a high-rate system.

Main Results:

• The proposed deep learning algorithm demonstrated superior feature extraction from non-stationary time series compared to heuristic methods.
• Significant improvements in step-ahead prediction accuracy and prediction horizon were achieved.
• The algorithm exhibits an average computing time of 25 microseconds, suitable for real-time applications.

Conclusions:

• The deep learning approach, particularly with the multi-rate sampler, effectively models high-rate dynamic systems.
• The algorithm's efficiency and accuracy make it a promising tool for real-time prediction in critical high-rate applications.
• This method offers a viable alternative to traditional physical modeling for complex dynamic systems.