When building AIoT systems for automotive operations the biggest trap seems to be beginning with the AI model.
It's typically much more beneficial to start with the data flow.
Without relevant, trusted data, no AI system will tell you anything of use about what is happening operationally. In the automotive domain those operational elements are numerous - connected assets, identification systems, location services, production processes, inventory management, test systems, and more.
The job of the engineer is to find a way to link and unify all those discrete signals into an operational representation of what actually is happening at the physical level.
A simplified AIoT architecture
The very high-level architecture of an AIoT system in an automotive context can be drawn as such:
Physical Assets Identification & IoT Data Layer AI/Analytics Operational Application Human Decision
These layers will, to some extent, each be responsible for solving a slightly different problem within the system.
- Physical assets
This first layer, that the system is concerned with monitoring, is simply anything or activity within the physical operation that is desired to be known. Examples include:
Vehicles
Components
Tools
Fixtures
Equipment
Material containers
Production assets
Test Equipment
People
These are just physical objects and not simply records in a database. There are characteristics such as state that change as a physical entity interacts with it's environment and it changes physical state. This aspect of the system is important when describing a physical process.
- Identification and IoT
The physical entity and some identifier must be linked conceptually and factually. An identifying characteristic can include whether the asset has been seen, where it has been seen, and what process context it's part of. Knowing that something exists is valuable in some contexts, but in operational settings we typically need more information than this:
What is this thing?
Where is it?
When was it located here/when was it last known here?
What zone/area am I finding this thing in?
Is this thing in the correct production state and available/committed/out of service etc.?
What production step or order is this item associated with?
Signals at this level form the bedrock of operational visibility. However, the standalone value of an event, from say a sensor, is generally quite low.
- The Data Layer
Once the raw events can begin to be aggregated more meaningful information starts to be generated, typically by combining several related event sources. For example, 'Asset ID + Location + Time + Production Step + Inventory Status' carries far more information than just 'Item at Location XYZ at Time T'.
This is where data modeling begins to get interesting and useful. Instead of relating discrete events, relationships between entities themselves start to become important:
- A vehicle is associated with a specific production stage.
- A component part is associated with a given manufacturing batch.
- A tool is associated with a specific work center.
- A specific material container may be associated with both a material and inventory and a production need.
These relationships form the platform for performing analytics.
- AI and Analytics
Based on the nature of the decision the AIoT system will support the most appropriate analytical approach can then be identified. Analytical methods span from very basic counting and summing through sophisticated pattern recognition, anomaly detection, and machine learning-based forecasting and optimization.
It is often tempting with a new AI project to focus on what we can use AI for, but this usually doesn't solve operational problems effectively. Always begin the effort with the following question:
'What decision can we improve?'
This keeps the systems focused on actual, desired changes at the operational level.
Automotive use case example: Visibility of Work-In-Progress
Consider a simplified production system for cars.
Body Assembly
Paint Assembly
Inspection
Final Processing
A specific car ID can be linked to each step and location.
If a vehicle has been sitting in one production station longer than is normally accepted for that phase, that becomes immediately observable and indicative of potential problems.
Multiple vehicles stuck in that one specific step would indicate a production bottleneck.
If an item of necessary components is missing from that workstation as a vehicle hits a certain stage in it's lifecycle, Production and Material concerns can be immediately coupled.
The helpful information arises from correlating multiple sources: not just that the vehicle is at this station but why is it there and what is happening that is anomalous to desired state/cycle time/etc..
Another data problem where automotive components create yet another link:
Part- Tooling - Machine - Production Batch - Inspection - Shipment - Storage
Component Traceability is generally where an asset is linked back into the production system and becomes critical in situations where parts need to be identified for warranty or failure analysis, etc.. Through the system, tracing links can provide both individual part trace-back and product trace-forwards into how different aspects influenced how the component was fabricated.
Why Contextual Modeling Is Hard
The creation of static snapshots in the form of databases is typically easy; understanding what those snapshots actually mean at an operational level, however, is generally far harder. A specific report might tell us that an item is 'At Location ABC at Time T'. That one data point tells us very little about what we actually want to know.
Was it useful that the item was at 'Location ABC'?
Should it have been there in the first place and what event preceded this one to send it there? Is the current location and state going to have ripple effects downstream? To answer these kinds of questions we start to see the value in integrating a broad array of otherwise distinct event and data sources together in a structured way that helps establish relationships such as:
- Human interacting with the asset
- The specific location it is now at
- The materials with which it is associated
- The process it is currently within
- Any time-based parameters associated with it
- Any specific events that triggered its current location and state
The more deeply we can embed these links within our model, the more useful subsequent analytical efforts will become. This observation is one also reflected throughout all Automotive Group domains, which include vehicle manufacturing, component production, Automotive electronics, Autonomous vehicles, after market and refurbishment efforts.
Automotive processes go beyond assembly
In other automobile specific applications these links and concerns also apply. For autonomous vehicle programs we would be concerned with autonomous vehicles, their testing equipment, personnel that operate and supervise them, storage location for vehicles, availability and use by the programs. If you were to model an aftermarket or refurbishment endeavor you are going to be tracking high volumes of items, components, containers and products that need to be manufactured, shipped and stored appropriately.
And for every application we can imagine within the automobile's product lifecycle, there's always a requirement for traceability and operational representational link management.
The architecture does not just serve the need of figuring out where something is; it needs to maintain knowledge of how something relates to everything else in the universe of that operation.
Engineering Conclusion:
It's really quite simple at the very lowest common denominator when it comes to getting systems to work for operations. It isn't enough to say IoT plus AI will create intelligence. What happens within the enterprise is often represented as a flow, a structure if you will of
Physical operations+Reliable data + Contextual model +Analytics + Human intervention = Operational intelligence
The AI is but one part of the larger picture; without reliable data identification the system isn't useful. Without relating data it is hard to find much out. Without connecting disparate data streams meaningful operational relationships might remain undiscovered. This shows that for the engineer of such systems, that part of the equation that remains challenging is usually not finding out how to use the AI; rather it is constructing the most reliable link from the real physical operation to it's digital shadow. For more info visit: apertureventurestudio.com
Top comments (0)