Continual Deep Learning for Time Series Modeling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep Learning (DL) models excel at feature abstraction but struggle with non-stationary data distributions.
• Catastrophic forgetting, the abrupt loss of learned knowledge, is a key challenge for DL in dynamic environments.
• Sensor time series data often exhibits non-stationarity, necessitating robust learning approaches.

Purpose of the Study:

• To systematically review Deep Learning applications in sensor time series analysis.
• To highlight the necessity of advanced preprocessing techniques for sensor data.
• To summarize methods for deploying DL in time series modeling while mitigating catastrophic forgetting using continual learning.

Main Methods:

• Systematic literature review of Deep Learning applications in sensor time series.
• Analysis of preprocessing techniques tailored for sensor data challenges.
• Exploration of continual learning strategies to address catastrophic forgetting in time series models.

Main Results:

• Identified diverse applications of DL in sensor time series across various domains.
• Emphasized the importance of domain-specific preprocessing for optimal model performance.
• Demonstrated the efficacy of continual learning methods in preserving knowledge during model updates.

Conclusions:

• Deep Learning, combined with advanced preprocessing and continual learning, offers powerful solutions for sensor time series analysis.
• Tailored DL approaches are crucial for handling the complexities of real-world sensor data.
• Continual learning is essential for building adaptive and persistent DL models in dynamic environments.