A Resource-Aware Method for Evaluating AI Models in Data Streaming Systems for Predictive Maintenance

Computation, data transmission, and all the activities involved in Artificial Intelligence of Things (AIoT) systems raise growing concerns in terms of resource consumption and environmental impact. Industrial asset monitoring is a data-intensive use case requiring a balance between predictive performance, responsiveness, and resource efficiency. In this context, edge and cloud computing represent complementary paradigms, enabling trade-offs between latency, scalability, and resource availability. However, understanding how these trade-offs affect model behavior and system-level resource demands in data streaming systems remains an open challenge, particularly in the context of resource-efficient and sustainable AI system design. In this work, we present a resource-aware method for evaluating AI models in data streaming systems for predictive maintenance, applied to the NASA IMS Bearing Dataset. The method is based on a controlled streaming setup that emulates a resource-constrained configuration and a resource-rich configuration, enabling a systematic analysis of model behavior under heterogeneous conditions. The evaluation considers predictive performance and operational metrics, including accuracy, latency, CPU utilization, and memory usage, to capture interactions between model characteristics and system-level factors. The proposed method enables a structured analysis of trade-offs between performance and resource utilization, offering insights for the design of resource-efficient and sustainable AI-based predictive maintenance systems.