Robust Distribution-Aware Ensemble Learning for Multi-Sensor Systems.

Area of Science:

• Machine Learning
• Data Science
• Sensor Networks

Background:

• Distribution and domain shifts in sensor data pose significant challenges for industrial monitoring systems.
• Adaptability to unnoticed shifts is crucial for reliable decision-making in critical applications.

Purpose of the Study:

• To introduce a novel, robust multi-sensor ensemble framework integrating automated machine learning (AutoML) to address domain shifts and variability in sensor data.
• To enhance adaptability to unnoticed distribution shifts and reduce the training cost of ensemble models.

Main Methods:

• A multi-sensor ensemble framework leveraging diverse model architectures, hyperparameters, and decision aggregation strategies.
• Integration of hyperparameter optimization and model selection for efficient ensemble training.
• Evaluation across five publicly available datasets for supervised and unsupervised shift detection.

Main Results:

• The framework demonstrates enhanced adaptability to unnoticed distribution shifts across various data properties.
• Significant improvement in common evaluation metrics compared to single-model baselines.
• Near-perfect test accuracy for classification tasks and effective distribution shift identification (90% AUROC, 20% FPR95).

Conclusions:

• The proposed AutoML-integrated framework offers a practical, distribution-aware solution for industrial applications facing real-world sensor data challenges.
• The method significantly enhances performance in both supervised and unsupervised distribution shift detection scenarios.
• This approach represents a novel step towards more resilient and adaptive industrial monitoring systems.