AI case study 4: machine learning in the manufacturing process

AI case study 4: machine learning in the manufacturing process

How Micron Technology uses machine vision, thermal imaging and acoustic listening to detect defects and breakdowns early, and what you need to do the same.

by Tradecloud Marketing Team

September 6, 2021

Illustration for Tradecloud's AI case study series

The business problem

Making memory chips is a delicate job. It takes about 1,500 steps, all carried out in sterile conditions to keep dust away from the wafers (see how it works). Even so, damage happens. Scratches and holes are often microscopic and almost invisible.

On top of that, machines, pipes and parts wear out, break down or start leaking. Detecting this early is essential. Engineers take care of maintenance, but even skilled engineers can miss the early warning signs.

The process leaves a lot of room for error. Relying on human vigilance was costing Micron Technology an average of about 250,000 US dollars per hour of downtime (Micron Technology, 2021).

The solution

This problem is a good fit for AI. It is clearly defined, it is measurable and there is enough in-house data to apply machine learning on several fronts. Smaller datasets also work, but with lower accuracy. The more data you collect, the more accurate the system becomes.

Intel also used machine vision and machine learning in its wafer production. Its whitepaper notes:

"Similar technology can be used in many different industries."

The machine learning algorithms detect anomalies earlier, more precisely and more often than people can. People are still needed to interpret the alarms and act on them. Micron uses three techniques.

Machine vision

Micron added machine vision to the photolithography cameras that etch circuitry into the wafers. The system scans for frequently occurring flaws and alerts engineers. Depending on the type of flaw, the alert arrives between 15 seconds and 15 minutes after detection.

An Auto-Defect-Classification (ADC) system uses deep learning to classify and sort millions of flaws automatically. Intel's whitepaper describes the ADC approach in more detail.

Thermal imaging

Micron also uses thermal imaging to monitor production. Heat maps of the factory under normal operating conditions are overlaid on a digital twin, a digital copy of the factory. Together they form a baseline. Real-time infrared images are compared with this baseline, and irregular temperatures trigger an alarm.

Acoustic listening

Abnormal machine sounds often signal trouble. Micron's AI system is trained to detect irregular sound frequencies by converting sound into visual data points. Audio sensors are placed close to machinery and pumps, so individual sounds can be isolated in a loud environment. Engineers then categorise the sounds and determine the likely causes.

Benefits and impact

Micron reports clearly higher manufacturing efficiency and accuracy. Worker safety improved, because people are less often exposed to extreme temperatures and harmful substances. Engineers gained time for other work. AI also spread to other business processes, such as product demand forecasting, where accuracy improved by 10 to 20%.

Key benefits reported by Micron:

  • 10% increase in manufacturing output
  • 35% fewer quality issues
  • 25% faster time to yield maturity
  • millions of US dollars saved through early detection of breakdowns and quality issues
  • more engineer time for other work
  • improved worker safety
  • a foundation for AI in other business processes

Accessibility and requirements

Data is the core of this solution. Micron collects petabytes of manufacturing data from more than 8,000 sources and over 500 servers worldwide. The data flows into two environments that run Apache Hadoop for data mining. Hadoop processes large datasets in parallel. More than 2,000,000 images for machine vision are stored in the Hadoop environment.

Acoustic data goes to a GPU system, which handles the complex machine learning workload quickly. A GPU splits tasks across many processors that work at the same time.

To build a comparable AI solution, you need:

  • enough images and data on flaws, normal sound frequencies and temperatures
  • a deep learning algorithm (machine vision flaw classification)
  • a digital twin (thermal imaging)
  • infrared cameras (thermal imaging)
  • audio sensors (acoustic listening)
  • Apache Hadoop for data mining
  • GPU systems
  • machine learning algorithms

Also read AI case study 1: demand forecasting using artificial intelligence.

Sources

Want to know what AI and better data can do for your supply chain? Book a demo.