Building An AI-Enabled Automotive Smart Factory: From Shop-Floor Data To Real-Time Intelligence
Building an AI-enabled automotive smart factory is usually a problem of machine-learning approaches.
It's often a problem of data integration before anything else.
An automotive factory has PLCs, robots, sensors, MES, ERP, RTLS, RFID, AGVs, quality systems, and, crucially, the vehicle genealogy databases run concurrently. While each of these systems yields useful information, you might be surprised at how valuable it often is to tie that information together.
This article addresses a practical architecture for tying these technologies together in real-world AI-enabled manufacturing settings.
The Basic Architecture
We've seen a basic architectural design that looks something like this:
┌──────────────────────────────────────────┐
│ Factory Floor │
│ │
│ PLCs │ Robots │ Sensors │ Vision │ AGVs │
└───────────────────┬──────────────────────┘
│
▼
┌──────────────────────────────────────────┐
│ Industrial Connectivity Layer │
│ │
│ IIoT │ RFID │ RTLS │ OPC UA │ Gateways │
└───────────────────┬──────────────────────┘
│
▼
┌──────────────────────────────────────────┐
│ Edge Processing / Events │
│ │
│ Filtering │ Rules │ Alerts │ Local AI │
└───────────────────┬──────────────────────┘
│
▼
┌──────────────────────────────────────────┐
│ Manufacturing Data Platforms │
│ │
│ MES │ ERP │ Quality │ VIN │ Inventory │
└───────────────────┬──────────────────────┘
│
▼
┌──────────────────────────────────────────┐
│ AI & Analytics │
│ │
│ Anomaly Detection │ Prediction │ BI │
└──────────────────────────────────────────┘
How exactly it's implemented depends on the plant, but the principle is simple: AI should be placed within the structure of the manufacturing data analytics, not alongside it.
- The Shop Floor
The shop floor layer starts with the manufacturing environment itself.
Some of the systems you may have on the shop floor are:
PLCs
Robots
CNCs
Sensors
Machine-vision systems
Torque tools
AGVs
Production stations
These systems can offer up highly valuable operational telemetry.
Take a machine, for example, that provides temperature telemetry, cycle time, vibrational data, alarm codes, and the state of production.
The telemetry is valuable, yes, but it becomes much more valuable when correlated with manufacturing context such as:
What machine does this telemetry event originate from?
What production cycle is it taking place in?
What vehicle is being produced?
Did a quality event happen close to the same time this telemetry is coming in?
Context is needed, and that's where the other layers come in.
- Industrial Connectivity
Next comes the problem of getting data about what is going on out of these systems reliably.
Industrial factories may house equipment with an amalgam of protocols and come from multiple vendors. To this end, an integration architecture will have gateways, OPC-UA, MQTT, RFID, RTLS, etc.
It's not about "replacing" systems, but about getting pathways open for consuming relevant information elsewhere.
- Edge
The next step in the design process considers edge computing.
Not all events need to be processed in a central location.
In the past, manufacturers would send telemetry from each machine, and from each sensor, from every production station right up to the central platform.
This is fine, but an edge layer can provide benefits including:
Filtering out noisy telemetry
Aggregating events in a local machine
Detecting local anomalies
Applying rules engines
Caching data due to spotty connectivity
Generating time-sensitive alerts
An example use case is a machine processing its own telemetry, but filtering out local noise and passing only relevant events to an analytics platform.
This isn't about completely doing away with centralized systems.
It's about delineating responsibility, and that's a big benefit in manufacturing environments.
- MES and Enterprise
The information on the shop floor is valuable, but it can often only tell you so much.
MES can give you additional production context, ERP can give you enterprise logistics or inventory information, and a quality system can yield quality-related events.
The problem is linking all of these systems.
Let's say, for example, you have an AI model that's spotted an anomalous event on the factory floor.
If you've got no additional manufacturing context, then your response might be something like:
"An anomaly was detected."
With additional manufacturing context, your response might be something like:
Which production station is this occurring at?
Which vehicle is being built?
Which production order is this?
Is the machine failing to follow a standard cycle?
Is there a quality event occurring at the same time?
Has maintenance been performed lately on this machine?
That second response is significantly more actionable than the first.
- RTLS as A Context Layer
Real-time location systems can be a great addition to your manufacturing context layer.
RTLS can yield the following information about the following (among other):
Vehicles
AGVs
Containers
Tools
Materials
Equipment
Other assets
This information, however, shouldn't be considered in isolation.
Consider the following information:
Asset Location
+
Production State
+
Material Requirement
+
MES Event
+
Time
=
Operational Context
This is how the information you gather about your equipment can be leveraged, and it's all about context.
Instead of "What is the location of this asset?", manufacturers can move to "How does this asset's location correlate with production and logistics?"
- VIN Genealogy
When working with automobiles, there's a uniquely valuable source of contextual information that can tie many events together at once: The VIN.
VIN genealogy allows you to correlate a vehicle's VIN to components, production stations, process events, quality events, manufacturing timestamps, and other manufacturing-related data.
This enables the linking of the physical world (a VIN) to the digital one (the manufacturing process).
From an analytics perspective, this can be invaluable in pinpointing which process events correlate most strongly with quality issues, components, or production stations. If a quality issue is discovered, manufacturers can then dig into which process events, components, or production stations are tied to the affected VINs to further isolate the cause.
- Where Does AI Fit In?
Once you've got your data architecture sorted, you can start thinking about how to apply AI to your problems.
Some applications of AI in manufacturing are self-evident:
Predictive maintenance
AI can be used to analyze historical and current telemetry data to detect patterns which indicate machinery failure.
Anomaly detection
Useful in identifying suspicious combinations of machine events, production events, quality events, or even RTLS data.
Production optimization
AI can optimize cycle-time, machine utilization, and other key production metrics based on events that have taken place.
Material flow analysis
Combine inventory data from ERP with production events from MES, RTLS, and AGVs to determine bottlenecks or logistic issues.
Quality analytics
Link quality events to relevant manufacturing history to detect the root causes of defects.
What's important is that the AI model used fits the task.
Not every issue in manufacturing lends itself to a deep learning approach; statistical modeling, rules engines, time-series analysis, or even just event correlation may be better alternatives in certain circumstances.
- AI Doesn't Replace Other Manufacturing Systems
There's a common fallacy that AI will replace MES or other core manufacturing platforms.
The more realistic approach is to consider AI as an intelligence layer which functions alongside these systems.
For example, MES can yield:
Production events
Work orders
Process state
These can be fed into an AI analytics layer, which can then provide predictions, classifications, or alerts back to the appropriate MES module.
A similar approach can be leveraged for other manufacturing systems as well.
Consider the following example diagram.
┌───────┐
│ AI / Analytics │
└───────┘
MES
│
├── Production Events
│
├── Work Orders
│
└── Process State
│
▼
IIoT
▲
│
┌─────┐ ┌─────┐
│RTLS│ │IIoT │
└─────┘ └─────┘
│ │
Location Telemetry
The AI layer consumes information from these systems, and in turn, it produces predictions, classifications, or alerts which can be fed back into them.
This is beneficial because manufacturers can avoid replacing deeply entrenched core systems with new AI systems.
- Designing For Real World Manufacturing
The requirements that come with an industrial application are not always the same as a web application. You need to think about latency, network reliability, data quality, legacy systems, and time-synchronization.
Even the best model, when implemented incorrectly, will struggle in the real world.
You could find your model performs perfectly in your testing environment, but when you roll it out across thousands of machines and a dynamically shifting production environment, its results are less than optimal.
This is why it's important to think about deployment and observability, too. Your manufacturing environment has specific requirements that may include:
Low latency
High availability
Network reliability
Data quality
Legacy equipment
Cybersecurity
Time synchronization
Scalability and observability
Fail-safe behavior
All of these factors can affect your model's performance. A manufacturing platform that works incredibly well in a controlled setting may have to change considerably when put against the dynamism and scale of the real world. That's why deployment architecture matters just as much as the model itself.
Putting It All Together
The companies which are currently working on industrial AIoT solutions are increasingly looking to tie together different technologies which would traditionally have been used in isolation. OEMNex AI, among others, is seeing a demand across the industry for AI solutions in manufacturing. The OEMNex AI Manufacturing product suite is an example of an AI enabled automotive factory that leverages a combination of industrial IIoT, RTLS, VIN genealogy, MES integration, and manufacturing analytics to achieve its end goals.
The engineering principles that drive those projects are more instructive, however: Create reliable data connections, establish manufacturing context, and apply AI where necessary.
Finally
An AI-enabled automotive smart factory isn't just made of machine-learning models, but is a system. The machine tools provide telemetry, the RTLS yields location data, the MES yields production data and process events, the ERP yields enterprise data, the quality systems provide inspection and quality data, and the VIN genealogy binds many of the aforementioned pieces together. Edge computing is utilized for time-critical workloads, and AI and analytics are applied which turn all of these combined data streams into something manufacturers can use. When these data layers are designed to work together, manufacturers can move their factory data from siloed, disparate sources to real-time, unified intelligence.
An interesting engineering challenge isn't “How do we add AI?”
It's
"How do we design a data architecture to give our AI context so that it can do a truly useful job?"
Top comments (0)