Industrial AI Reference Architecture for Manufacturing

Data integration and processing

Industrial systems (ERP, sensors, equipment) feed raw data through Industrial DataOps platforms into a cloud-based medallion architecture, where it’s cleaned and enriched through Bronze, Silver and Gold layers for different use cases.

• Analytics and intelligence: Processed data powers business intelligence dashboards for production metrics (OEE, compliance, real-time counts) and real-time alerts and trains AI models for production scheduling, maintenance management, and quality inspection.
• AI agents and edge deployment: AI agents orchestrate field service recommendations, production control recommendations and facilitate root cause analysis, with models trained in the cloud and deployed to edge devices near equipment for real-time, low-latency decision-making and safety controls.

Data flows

Following are descriptions of the data flows shown in the industrial AI architecture diagram.

1. The ingestion patterns fall into two categories:
   1. ERP, EAM, market data, product manuals provide context about the configuration of the asset or product
   2. Industrial DataOps platforms connect to data sources in Level 3 and below within Purdue Model to the cloud for visibility within Databricks
2. Transactional data is ingested via Lakeflow Connect managed connectors, while streaming data is ingested with Zerobus into the medallion architecture bronze layer alongside metadata (unified namespace), leveraging efficient incremental reads and writes to make data ingestion faster, scalable, and more cost-efficient, while your data remains fresh for downstream consumption.
3. Clean and enrich heterogeneous data scalably using Spark Declarative Pipelines for both batch and streaming data pipelines into Silver tables (manufacturing data models). Silver tables are often used as training inputs to AI models for predictive maintenance, quality inspection, etc., thereby improving asset performance and product quality across complex production processes.
4. For business intelligence and reporting, data can be aggregated within Gold tables to support real-time analysis of production performance (OEE), including compliance (product passport, carbon certification) and real-time production counts. Additionally, natural language interfaces like AI/BI democratize access to global factory network performance, statistical process control decision rules, and production alerts.
5. Agent Bricks can train, serve, and audit AI applications like:
   1. Autonomous Production Control that simulate production scenarios and override production schedules based on economic conditions, feedstocks, or ambient conditions.
   2. Field Service Assistant which guides technicians in natural language on spare parts, repair instructions, and safety guidance based on asset history and real-time operating conditions.
   3. Machine Vision applications to assess product quality and informing machine set points to eliminate waste and improve production yield.
6. AI Agents and machine learning models deployed to the edge to run locally near equipment for low-latency control and safety purposes.

Benefits

Benefits of using the Databricks Platform for industrial AI architecture include the following.

• Establish best-practices architecture for Industrial AI use cases and IT/OT Convergence
• Managed connectors and full ecosystem integrations with governance to facilitate agentic process control