Ensemble Active Learning by Contextual Bandits for AI Incubation in Manufacturing

Online sensing and computational resources in Industrial Cyber-physical Systems (ICPS) facilitate AI-driven decision-making. Yet, issues with data quality, such as imbalanced classes, hinder AI models trained offline. To address this, AI models are updated online with streaming data for continuous improvement. Supervised learning models, however, face challenges in selecting quality streaming samples for updates due to annotation constraints. Active learning methods in literature offer solutions by focusing on under-represented or well-represented regions. Balancing these strategies in changing manufacturing contexts is challenging. Some acquisition criteria learned by AI dynamically adapt but may not consistently handle frequent changes. We introduce an ensemble active learning method, CBEAL, employing active learning agents specifically for exploration or exploitation. Weights of agents are adjusted based on agent decision effectiveness. CBEAL optimally guides data acquisition, minimizing human annotation. Our theoretical analysis and empirical studies validate CBEAL's efficiency in ICPS manufacturing process modeling.