Loading...

Indiana Manufacturers Bring AI Closer to Machines With Edge Computing

Industrial IoT (IIoT)

Published by IOT Indiana Research & Editorial Team

Indiana Manufacturers Bring AI Closer to Machines With Edge Computing

Running an AI model on a server in a distant data center is one thing. Running it next to a press, a welder or a packaging line, where it must respond in milliseconds and keep working when the network does not, is a different job. That second job is edge AI, and it is where much of the practical value of artificial intelligence in manufacturing is now being captured.

This article is about the lifecycle of an edge AI model in an Indiana plant: choosing the use case, sizing the hardware, deploying and updating models, watching for drift, and keeping it all secure. It assumes you have already decided that some processing belongs near the machine. If you are still deciding where processing belongs, our earlier article on edge versus cloud placement covers that choice.

Why the conversation is moving to the plant floor

Two trends meet here. First, industrial AI has moved beyond analyzing data after the fact toward acting in real time: rejecting a part, adjusting a setpoint, stopping a machine. Second, the research and workforce ecosystem around Indiana is leaning in. Purdue's Institute for Physical Artificial Intelligence, which covers areas from robotics to self-driving cars, has a multiyear partnership with Google Public Sector announced in January 2026. Purdue and Google are hosting a summit in Indianapolis on October 29 and 30, 2026, focused on moving physical AI from pilots to scaled results. Starting with this fall's freshmen, Purdue also requires an AI working competency for graduation, so the pipeline of graduates who understand these tools should grow.

A 2023 survey of Indiana advanced manufacturers, published in 2024 by the Central Indiana Corporate Partnership with Conexus Indiana, concluded that AI adoption in manufacturing lags other Indiana sectors, though manufacturers are making progress. That fits what many plant managers report: interest is high, but production-grade deployments are still the exception.

Use cases that justify edge hardware

Use caseWhat the model doesWhy it needs to be local
Visual inspectionClassifies surface defects, missing features or assembly errors from camera imagesImages are large; decisions must come in milliseconds at line speed
Anomaly detection on motors and pumpsFlags unusual vibration or current patternsContinuous high-rate signals; alerts needed even if the network is down
Process control assistanceSuggests or applies setpoint adjustmentsClosed-loop timing and safety limits
Safety monitoringDetects people or vehicles in hazardous zonesReliability and latency are critical
Robot guidanceLocates parts for picking or positioningTight loop between vision and motion
Sound and acoustic checksListens for abnormal noises at end-of-line testsRaw audio is bulky; quick pass or fail is needed

If your use case does not need fast response, high-volume raw data or independence from the network, a cloud or central server approach is usually simpler and cheaper.

Sizing hardware without overbuying

Edge devices range from small industrial computers that handle a few sensor streams to ruggedized servers with graphics processors that run several camera models. Buying too small causes dropped frames and missed defects. Buying too large wastes money and adds heat and maintenance burdens in a dusty plant.

  1. Start with the workload. How many cameras or signals, at what rate, and how many models run at once? A single 5 megapixel camera at 30 frames per second needs far more than a vibration sensor sampled at a few kilohertz.
  2. Benchmark on real data. Test candidate hardware with your actual images or signals and measure latency under load, not vendor peak numbers.
  3. Leave headroom. Plan for models to grow and for lighting or product changes that make the task harder. A margin of 30 to 50 percent is common practice.
  4. Check the environment. Heat, vibration, dust, humidity and power quality kill electronics. Match enclosures, cooling and power conditioning to the location.
  5. Plan for failure. Decide what happens if the box dies: a spare unit on the shelf, a fallback to manual inspection, or a safe machine state.

Deployment: from a laptop demo to a production line

Many pilots succeed in a lab and fail on the floor. The gap is usually operational, not mathematical.

Data collection

Models learn from examples. For defects, collect images across shifts, lighting conditions, suppliers, materials and operators. Include rare defects, even if you must stage or collect them over months. Label carefully, using a written definition of each defect class so different people label consistently.

Validation against reality

Measure the model on data it has never seen, and report what matters operationally: missed defects, false rejects and how both change by product variant. A model that is 98 percent accurate overall may still miss the one defect that causes customer returns.

Packaging and rollout

Run models in containers or a similar packaging method so the same build can be tested and installed on multiple devices. Use a staged rollout: run in shadow mode first, comparing model decisions with human decisions without acting on them, then move to advisory alerts, and finally to automatic action where safe.

Integration with controls

Connect results to the PLC or manufacturing system through standard interfaces such as OPC UA or MQTT, and keep safety functions in certified safety hardware rather than in an AI model.

Model drift: the problem nobody mentions in the demo

Models trained on yesterday's conditions can degrade when conditions change. A new supplier's material has a slightly different color. A camera gets bumped. A lamp ages and dims. The model's performance quietly declines.

  • Monitor inputs. Track basic statistics such as image brightness or sensor ranges, and alert when they move outside the training range.
  • Monitor outputs. A sudden change in reject rate often means a process change or a model problem. Compare with manual checks regularly.
  • Keep a review queue. Send a sample of uncertain or rejected items to a person, and feed confirmed examples back into training.
  • Version everything. Record which model version, trained on which data, made each decision. This matters for traceability and for troubleshooting.
  • Schedule retraining. Treat it as maintenance, with an owner and a calendar, not as a surprise project.

Security and governance at the edge

Every edge box is a computer on your plant network. Treat it accordingly: restrict inbound connections, use encrypted communication, change default credentials, keep software patched on a tested schedule and log access. Separate the network the AI device lives on from office IT, following the same zone thinking used in standards such as IEC 62443. For vendors, ask how updates are signed and how models are protected from tampering.

Governance matters too. Decide who can approve a new model, who can override it, and how you will document changes for customers who audit your quality system. Automotive, aerospace and medical customers will ask.

An illustrative example: inspecting stamped brackets

These numbers are invented to illustrate the reasoning. A Midwest stamping supplier checks small steel brackets for burrs and cracks with human inspectors at the end of the line. Inspectors catch most problems, but fatigue causes some escapes, and the customer has asked for better containment.

  • The plant installs two cameras with controlled lighting and an industrial computer with a GPU at a cost of about $28,000 including integration.
  • For six weeks it collects images and inspector decisions without acting on them, capturing 120,000 parts.
  • The model is tested on a held-out set and found to flag nearly all cracks, with a false reject rate of about 1 percent. Operations accepts that level because rejected parts are re-inspected by a person.
  • After a shadow period, the system begins diverting flagged parts to a review bin. Inspectors now check a small fraction of output and focus on the uncertain cases.

Whether this saves money depends on the cost of escapes and inspection labor, which differ by plant. The structure of the project is the reusable part: collect, shadow, validate, then act.

Skills and partners

You do not need a team of data scientists to start, but you do need a few roles: a process engineer who understands the defect or fault, someone who can handle data and networking, and a maintenance owner for the hardware. Many plants combine an internal champion with an integrator or vendor for model development. Purdue Manufacturing Extension Partnership advisers can help scope projects, and Purdue's industry partnership offices can connect companies with research expertise. State grant support has changed over time; the Manufacturing Readiness Grants were paused in 2025, so verify what is available.

Mistakes to avoid

  • Chasing the model before fixing the process. If lighting, fixturing or part presentation vary wildly, no model will be reliable.
  • Skipping the shadow period. Trust must be earned with side-by-side evidence.
  • Letting AI own safety. Keep safety functions in dedicated, certified systems.
  • Forgetting the exit plan. Know how to run without the system, and how to move data and models if you change vendors.

The bottom line

Bringing AI closer to machines is less a technology leap than an operations discipline: define the decision, collect honest data, size the hardware, prove the model beside people, watch it for drift and secure it like any other plant system. Indiana manufacturers that treat edge AI as a lifecycle, not a demo, will get durable value from it, and they will be well placed as the research and talent base around the state continues to grow.

Frequently asked questions

It means running AI models on devices near the machines, such as an industrial computer at a production line, so decisions like rejecting a part or flagging an anomaly happen in milliseconds and continue even if the network connection fails.

Edge is a good fit when the task needs very fast response, produces large volumes of raw data such as video, or must keep working without a network connection. If speed and data volume are not issues, a central server or cloud approach is often simpler and cheaper.

Model drift is the gradual decline in a model's accuracy when real conditions change from those it was trained on, for example new materials, aging lights or shifted cameras. Monitoring inputs and outputs, sampling decisions for human review and scheduling retraining help catch it.

Collect representative data, validate on unseen examples, run in shadow mode beside human inspectors, move to advisory alerts and only then automatic action. Keep safety functions in certified safety hardware, not in the AI model.

A 2023 survey published in 2024 by the Central Indiana Corporate Partnership with Conexus Indiana found that AI adoption in manufacturing lags other Indiana sectors, though manufacturers are making progress. Interest is high, and production-grade deployments are still developing.

Purdue Manufacturing Extension Partnership advisers, Purdue industry partnership offices, Conexus Indiana and integrators with machine vision experience can help. State grants change over time, and the Manufacturing Readiness Grants were paused in 2025, so confirm current options.

Related articles

Top