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

Accessing operational information across industrial systems (ERP, MES, SCADA, PLC) in concrete plants requires 15-30 min and specialized knowledge. This work addresses this accessibility gap by developing a conversational AI system that democratizes industrial information access through natural language. A five-layer cognitive architecture was implemented integrating the Mistral-7B model quantized in GGUF Q4_0 format (3.82 GB) on a Raspberry Pi 5, Spanish speech recognition/synthesis, and heterogeneous industrial protocols (OPC UA, MQTT, REST API) across all automation pyramid levels. Experimental validation at Frumecar S.L. (Murcia, Spain) characterized performance, thermal stability, and reliability. Results show response times of 14.19 s (simple queries, SD = 7.56 s), 16.45 s (moderate, SD = 6.40 s), and 23.24 s (complex multilevel, SD = 6.59 s), representing 26-77× improvement over manual methods. The system maintained average temperature of 69.3 °C (peak 79.6 °C), preserving 5.4 °C margin below throttling threshold. Communication latencies averaged 8.93 ms across 10,163 readings (<1% of total latency). During 30 min of autonomous operation, 100% reliability was achieved with 39 successful queries. These findings demonstrate the viability of deploying quantized LLMs on low-cost edge hardware, enabling cognitive democratization of industrial information while ensuring data privacy and cloud independence.