Multimodal Cognitive Architecture with Local Generative AI for Industrial Control of Concrete Plants on Edge Devices

Area of Science:

• Industrial Automation
• Artificial Intelligence
• Edge Computing

Background:

• Accessing industrial data in concrete plants is time-consuming (15-30 min) and requires specialized knowledge.
• Current systems like ERP, MES, SCADA, and PLC present accessibility challenges.
• There is a need for democratized, real-time access to operational information.

Purpose of the Study:

• To develop a conversational AI system for natural language access to industrial information.
• To address the information accessibility gap in concrete plant operations.
• To demonstrate the feasibility of deploying AI on low-cost edge hardware for industrial applications.

Main Methods:

• Implemented a five-layer cognitive architecture on a Raspberry Pi 5.
• Integrated a quantized Mistral-7B Large Language Model (LLM) in GGUF Q4_0 format.
• Utilized Spanish speech recognition/synthesis and industrial protocols (OPC UA, MQTT, REST API).

Main Results:

• Achieved significant response time improvements: 14.19s (simple), 16.45s (moderate), 23.24s (complex) queries.
• Demonstrated system reliability with 100% success rate over 30 minutes of autonomous operation.
• Maintained thermal stability with average temperature at 69.3°C, below the throttling threshold.

Conclusions:

• The developed AI system democratizes industrial information access via natural language.
• Quantized LLMs are viable on low-cost edge hardware for industrial settings.
• The system ensures data privacy and cloud independence, enhancing operational efficiency.