TLDR: This research introduces an agent-based framework for quickly deploying Artificial Intelligence models on edge devices within Industry 5.0 settings. It focuses on reducing latency, enhancing data privacy through local processing, and enabling real-time human-AI collaboration for tasks like model recalibration and anomaly detection. Validated in the food industry, the framework significantly improves deployment time, system adaptability, and operational efficiency, offering a modular and cost-effective solution for smart manufacturing.
The landscape of industrial operations is rapidly evolving with Industry 5.0, emphasizing the integration of real-time Artificial Intelligence (AI) at the edge of networks. This shift allows factory operators to supervise automated systems more closely, building upon Industry 4.0’s focus on digital transformation, the Internet of Things (IoT), and AI. A key element in this evolution is Edge AI, which processes data directly at its source, enabling immediate data handling and responsive, human-centered manufacturing processes.
Moving AI tasks from distant cloud servers to local devices offers significant advantages: it reduces delays, lessens network load, and limits the exposure of sensitive data. This is crucial for applications demanding quick responses, where prompt actions are essential. Local data processing also enhances service quality and reduces reliance on network connectivity, making it ideal for latency-sensitive tasks and alleviating network congestion.
However, implementing Edge AI is not without its challenges. Hardware must be capable of supporting AI workloads on devices with limited capacity, distributed resources need efficient management, and interoperability across diverse edge environments must be guaranteed. Security is also a critical concern, as edge devices often operate in less controlled conditions than centralized servers. These requirements call for new toolkits that simplify deployment, optimize resource use, and ensure system reliability.
A Novel Framework for Rapid Edge AI Deployment
In response to these needs, a new framework has been proposed for the fast deployment of Edge AI solutions in industrial scenarios. This framework introduces several innovations around the concept of Collaborative Intelligence (CI), allowing human operators to work jointly with intelligent agents to manage and refine AI models deployed at the edge. The solution has been validated through a use case in the food industry, demonstrating improved deployment time and system adaptability.
The framework is designed for rapid prototyping, testing, and deployment of Edge AI models across various industrial settings. It integrates best practices for Edge AI, ensuring seamless communication between edge devices and centralized servers. It uses real-time protocols like MQTT for efficient, bidirectional data flow and incorporates interactive data visualization to monitor system performance through dynamic charts.
Understanding the Architecture
The framework’s architecture is component-based, with each part handling a specific task:
-
Config Loader: Sets up configurations for data ingestion from static datasets (CSV Reader) and real-time sources (Sensor Streaming).
-
MQTT Broker: Acts as a central hub for data exchange between components.
-
Inference Agent: Subscribes to incoming data, processes it, and publishes predictions.
-
UI Agent: Visualizes raw data and inference outputs in real-time, allowing human operators to recalibrate models and inspect anomalies.
-
GenAI Agent: Integrates with external generative AI models (e.g., GPT-4o) via REST API calls to provide explanations for predictions and assist with labeling.
-
Designer Agent: Manages the deployment of inference pipelines.
This modular design ensures that adding new data sources or visualizations doesn’t require changes elsewhere, and automated deployment pipelines streamline updates to inference logic. On-demand AI explanations appear seamlessly in the user interface, helping users understand model outputs.
Human-AI Collaboration in Action
The UI Agent provides a real-time web interface that displays incoming data streams and model outputs. A key feature is a table where target values are editable, enabling users to correct or adjust them manually. This supports human-in-the-loop (HITL) workflows, allowing domain knowledge to refine results. Visual components, such as time series and bar charts, track predicted versus actual values and categorize entries as OK or Non-OK, updating automatically to help users monitor performance and detect anomalies.
The GenAI Agent plays a crucial role by constructing structured prompts for external large language models when discrepancies are detected. It can explain model decisions in natural language, flag entries for recalibration, and even assist with labeling, supporting both zero-shot and few-shot prompting strategies.
The Inference Component, running on edge devices like ESP32 and Raspberry Pi, applies trained models to incoming data. Its open hardware foundation allows for customization, making it suitable for various industrial settings. The Design Component acts as a pipeline builder, enabling users to develop and deploy AI pipelines for execution on edge or cloud infrastructure, with configurable modules for data processing and model execution.
Real-World Application in the Food Industry
The framework has been successfully adopted by Quescrem, a leading cream cheese producer, with support from Gradiant, a technological center. It gathers real-time data at different production stages, monitoring variables like fat content, pH levels, pressure, and temperature. Human operators and AI agents analyze this data to predict deviations and classify batches as OK or Non-OK. Operators can adjust targets and record explanations for Non-OK batches, ensuring transparent decision-making.
This approach has significantly improved production effectiveness, leading to approximately 65% less downtime due to proactive maintenance and around 20% better energy efficiency through optimized machine utilization. Deployment setup time decreased by about 80% compared to traditional manual procedures, and the framework maintained average end-to-end latencies under 200 milliseconds, with predictive accuracy consistently exceeding 95%.
Also Read:
- Advancing Digital Twin Control with Hybrid Models and AI-Driven Strategies
- Agentsway: A New Software Development Approach for AI Agent Teams
Distinguishing Features
Compared to other systems like AIfES, InfiniEdge AI, and OpenEI, this framework stands out due to its agent-based strategy, which allows specialized tasks to be performed by human operators, machines, or both in collaboration. This unique feature, combined with its focus on low latency via MQTT, modular plug-and-play components, local data processing for enhanced privacy, and live visualization tools, aligns perfectly with the human-centric values of Industry 5.0. It runs on modest edge hardware and pairs with open-source tools to keep costs low.
While the framework offers substantial benefits, challenges remain, including the limited computational and storage capabilities of edge devices, difficulties in connecting with existing legacy systems, and potential issues with data reliability. Future work aims to improve model efficiency through techniques like quantization, extend deployment to more diverse edge environments, and refine the GenAI agent’s interaction through feedback-driven prompt adaptation.
The source code for the Collaborative Intelligence component is publicly available under an open-source license, encouraging reuse and contributions. More details can be found in the research paper.
