While every consumer is aware of AI from finding it embedded into their daily digital lives, specialist implementations of the technology are beginning to emerge that build on the algorithms and methods developed for consumer AI, to produce solutions honed to industry-specific uses.
In a paper titled “Self-Evolving Multi-Agent Network for Industrial IoT Predictive Maintenance“, the authors introduce SEMAS, described as a specific architectural and algorithmic framework for industrial internet of things operations.
SEMAS is introduced in the context of industrial IoT predictive maintenance, one of the primary uses of AI technology in industry. In industrial environments that deploy thousands of sensors and devices at the edge, real-time detection and adaptive response can present challenges that traditional machine learning and monolithic AI systems fail to meet.
The authors of the paper state that out-of-the-box technology challenges stem from rigid architectures, lack of continuous adaptation, and the computational limits of the typically smaller devices found on the edge. They argue a multi-agent system can address this issue by distributing specialised functionality in the computational tiers hosted on the edge and in the cloud. SEMAS is defined as a hierarchical multi-agent system in which lightweight agents at the edge operate on sensor data and collaborate with agents in the cloud tier, the latter handling collective detection-making and overall policy.
In SEMAS the architecture comprises three coordinated layers. At the edge, minimal feature extraction reduces raw sensor streams into forms suitable for further analysis. At a fog layer, data from multiple edge devices are combined for collaborative anomaly detection through consensus voting. In the cloud, reinforcement learning, specifically Proximal Policy Optimisation (PPO), is used to refine detection policies over time.
In a proposed SEMAS framework, multiple agent types are defined and associated with distinct tasks in each tier. The system includes a feedback loops that lets policies evolve without manual intervention.
SEMAS combines elements of multi-agent systems (MAS) and semantic technologies. Multi-agent systems originated in artificial intelligence, comprising autonomous agents coordinating in order to solve tasks that would exceed the ability of a single agent. In IoT settings, MAS can distribute decision-making in heterogeneous devices. Combining semantic technologies with MAS lets agents refer to shared ontologies and structured data representations.
Large monolithic LLMs, like GPT-4 or similar models, have memory and inference latency requirements that exceed what can be supported on typical edge or fog devices, and any reliance on cloud access for inference can increase latency and violate data governance requirements in typical industrial use.
LLMs known to consumers understand language and can generate text in multiple domains. By contrast, SEMAS does not rely on a general-purpose model other than to use one in a limit way in interactions with operators.
A SEMAS implementation uses smaller language models (with parameter counts substantially lower than typical consumer LLMs) to generate domain-specific explanations, can learn operations adaptively, and evolve overall policies. Reinforcement learning in SEMAS refines agent behaviour based on operational data, addressing the need for systems that adapt. The paper reports empirical data that shows performance improvements and lower latency compared with non-adaptive models.
(Image source: “The sun brightens the fog/clouds in early afternoon” by daveynin is licensed under CC BY 2.0.)
Want to learn more about IoT from industry leaders? Check out IoT Tech Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and co-located with other leading technology events. Click here for more information.
IoT News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
