Designing an AIoT Architecture for Automotive Manufacturing

Automotive factories are becoming increasingly data-intensive.

A single production facility has PLCs, sensors, robots, SCADA, MES, ERP, RFID readers, RTLS infrastructure, quality systems, and material-handling equipment all working in parallel.

The challenge is rarely about obtaining more data.

Challenging is the task of tying operational data from various systems together and using that information without hampering production.

That is what makes AIoT (Artificial Intelligence of Things) such a compelling proposition for industrial environments.

Start from the factory problem, not the AI model

A common mistake in industrial AI projects is to start with the wrong question.

"Where can we use AI?" is not the right way to phrase it.

Instead, companies should be asking:

Which manufacturing problem do we need to understand or improve?"

Possible examples may include:

Unexpected equipment downtime

Poor visibility into WIP

Difficulties tracing components

Material-flow bottlenecks

Inefficient asset utilization

Production cycle deviations

Fragmented manufacturing data

With the problem well defined, the required data sources and architecture will become apparent.

A simplified AIoT architecture

An automotive AIoT environment can be seen as comprising several layers:

Physical layer → Connectivity → Edge processing → Data integration → Analytics → Applications

On the physical layer are machines, sensors, PLCs, RFID readers, RTLS infrastructure, etc., all generating information.

That information is then transported over the connectivity layer, using appropriate industrial protocols and network technologies.

Edge systems can be used to process selected information closer to its point of origin when there is a need to be mindful of latency, bandwidth, availability, or local decision-making.

The data-integration layer then collates information from multiple operational sources.

Finally, analytics and applications turn data into information that production, quality, maintenance, logistics, and engineering teams can act upon.

Why it matters to have interoperability

Automotive factories tend not to use equipment from a single vendor, and/or a single generation of equipment.

A modern machine may expose its data via a standard interface whereas an older system may require a different approach to integration.

Protocols such as OPC UA and MQTT can be valuable components of an industrial connectivity architecture, depending on the equipment and requirements.

The point is that protocol selection should be informed by architecture and requirements rather than viewed as a technology decision in its own right.

The role of RTLS in manufacturing

RTLS data represents another potential source of operational data.

An automotive plant may be interested in the location or movement of:

Tools

Containers

Components

WIP

Material-handling assets

Vehicles

RTLS data becomes more valuable when it can be correlated with production context; knowing the location of an asset is good but knowing its location relative to a production order, station, or material-flow process is worth far more in terms of operational insight.

Tying production data to vehicle genealogy

Vehicle manufacturing is another area that lends itself to traceability.

A vehicle goes through numerous production operations and receives many components.

Tieing production events and component information to the VIN can create a far richer manufacturing genealogy.

This can allow organizations to associate relevant production information with a given vehicle, as opposed to treating all production events as self-contained items.

An example of how these concepts can be combined for automotive OEM environments is available in this overview of OEM vehicle manufacturing AIoT.

AI is only as good as the data and context around it

AI-based analytics are only as good as the data and context around them.

Before embarking on a predictive model, manufacturers should think through questions such as:

Is the data accurate?

Are timestamps aligned?

Are equipment identifiers standardized?

Is historical data available?

Can production context be applied?

How will abnormal predictions be validated?

Who is responsible for acting on the result?

Such questions can have far more impact on an AI initiative than the choice of machine-learning algorithm.

Security cannot be an afterthought

Interconnectivity adds to the number of systems that need to be considered from a cybersecurity perspective.

An AIoT architecture must therefore be mindful of network segmentation, authentication, authorization, secure communications, device management, monitoring, and appropriate access controls.

The exact controls will vary depending on the environment and risk appetite, but security should be taken into account at the architectural level, not as an addition at the end.

A practical approach to implementation

A gradual approach is best in avoiding unnecessary complications.

A possible sequence would be:

1. Identify a measurable manufacturing problem

2. Map relevant equipment and data sources

3. Create reliable connectivity

4. Normalize and contextualize data

5. Process time-sensitive information at the edge where appropriate

6. Apply analytics for the use-case

7. Tie insights to an operational workflow

8. Measure the outcome before scaling up

This serves to keep the AIoT implementation firmly focused on manufacturing goals.

The end result is not a factory with more connected devices, but a factory where industrial data can flow between systems reliably, with operational context maintained, and where engineers and production teams can use that information to make decisions that have a positive impact.
(OEM vehicle manufacturing)