Seeing How Every Machine Is Used Across 400+ Plants with AI

Marmon Holdings, Inc., part of Berkshire Hathaway Inc., is a global industrial organization. It comprises more than 100 autonomous manufacturing and service businesses. The company operates over 400 facilities and employs more than 20,000 people worldwide. Each Marmon business runs independently within a group structure that provides shared access to expertise and common markets.

01 The Challenge

Machines reconfigured for higher output, but no data to measure whether it worked

U.S. manufacturing capacity utilization sits at roughly 75.5%, below the long-term average of 78.2%. For discrete manufacturers, idle time is the largest single source of lost productivity. The gap between potential output and actual output is invisible until someone measures it.

Marmon’s operations team at the McKenzie Valve Plant had already taken a step toward closing that gap. They reconfigured machines so a single operator could run an entire cell of four machines. Output per operator improved. The reconfiguration worked, but the team wanted to know exactly how well. Were all four machines running when they should be? How much time was lost to idle periods between jobs?

The plant did not have data on run versus idle time for individual machines. Without it, the operations team was making workflow decisions based on observation rather than evidence. They could see that production was better after the reconfiguration. They could not see by how much, or where the remaining time was going.

Marmon’s leadership wanted an automated way to collect and analyze machine utilization data in real time. The system needed to be simple to deploy, cost-efficient to run, and easy to extend to other facilities. Provectus, an AI-first systems integrator and solutions provider, joined the project.

02 The Approach

Start with the cameras already on the floor, process video at the edge, prove it at one plant

Provectus began with an on-site investigation: floor layout, CCTV camera placement, connectivity, and video streams of machine operations. The question was how to detect machine status without adding sensors to every piece of equipment.

The answer was already there. Each machine has indicator lights that show its current operating status. Solid green, solid yellow, solid red, or light off. A camera pointed at those lights, paired with an ML model, could detect status automatically. No new sensors. No integration with PLCs or SCADA systems. Just a camera, an edge device, and a model.

Provectus chose an edge-based approach: processing video feeds directly at the plant on an on-premises ML appliance. This kept response times fast, bandwidth costs low, and the system independent of network conditions on the floor.

03 The Build

Video annotation pipeline, indicator-light detection model, and real-time status notifications

Provectus collected video fragments from the McKenzie Valve Plant to build a training dataset. The team created a video annotation pipeline and labeled the indicator light states. They built automated pipelines for training, evaluation, deployment, and continuous improvement using Amazon SageMaker.

The model reads indicator light colors from each machine and sends real-time notifications when status changes. Green means running. Yellow means standby. Red means stopped. Light off means the machine is powered down. The operations team sees a live feed of machine states across the cell. Historical data shows utilization patterns throughout the day.

The system requires two pieces of equipment per facility: a standard camera and an edge computing device. That simplicity is deliberate. Marmon operates over 400 facilities worldwide. Any monitoring system that required custom sensor installations or heavy IT setup at each site would never scale. This one can be deployed at a new plant in days.

04 The Results

From guessing at machine utilization to seeing it in real time

Provectus delivered the complete system in five weeks from concept to production deployment. The model was tested in a real-world production setting at the McKenzie Valve Plant. It accurately detected machine on/off status based on indicator light colors throughout operating hours.

5 Weeks

From concept to production-ready deployment

Across 400+ eligible facilities

For the first time, the operations team has continuous, real-time data on how each machine is used. Run time, idle time, and downtime are no longer estimates. They are numbers, broken down by machine, by shift, by day.

With that data, the team can measure the actual impact of the cell reconfiguration. They can see which machines sit idle while the operator works another. They can see how utilization shifts between morning and afternoon. Decisions that were previously based on experience are now backed by evidence.

The pilot’s success opened the path to broader rollout. The system runs on standard cameras and a single edge device. Deployment at any Marmon facility takes minimal setup. Marmon now plans to extend the system across its network of 400+ facilities worldwide.

05 What’s Next

A machine utilization system built to scale across a 400-facility network

The McKenzie Valve Plant pilot proved that video analytics can deliver accurate utilization data without sensor installations. Provectus works with Marmon on scaling the system to additional facilities and extending its capabilities.