Skip to content
AI in Operations | Part 4 of 10

Data Readiness in One Hour

Score availability, quality, history, labels, access and ownership per use case, and fix data problems at the source.

“We have lots of data” is the most common answer when a plant considers AI. It is also the least useful. Data readiness is not a property of a company. It is a property of one use case.

Six dimensions

Score each dimension from 0 to 4 for each candidate use case:

  • Availability: does the data exist and is it stored?
  • Quality: is it complete, with correct time stamps and few gaps?
  • History: are there enough examples, including the rare events that matter?
  • Labels: have experts marked the correct answers, such as good and bad parts or failure causes?
  • Access: is there a secure path from the machine to the model?
  • Ownership: is someone responsible for keeping the data right?

The average decides the next step. From 2.5 a pilot makes sense. Between 1.5 and 2.4, close the weakest dimensions first. Below 1.5, collect data before thinking about models.

Two use cases, two answers

A vision inspection of molded housings scores 3, 2, 3, 2, 3 and 3: an average of 2.7. Images can be captured, the history of defects is known, the line control can receive a reject signal. The weak points are quality and labels, because two inspectors judge the same scratch differently. A defect catalog with limit samples fixes that within weeks.

Predictive maintenance on four molding machines scores 1, 2, 0, 1, 2 and 2: an average of 1.3. The machines do not store sensor values, and failures are written as “broken” or “see note”. No algorithm can learn from that. The next step is sensors on two machines and three months of clean history.

Fix data at the source

Most data problems in factories have simple causes:

  • Different clocks. The machine runs ten minutes off from the quality system, so no defect can be matched to its process values. Use one time source for all systems.
  • Free text instead of codes. A short list of failure and defect codes replaces “broken” and “defect”.
  • Missing rare events. Twelve images of a crack in a year are not enough. Collect them deliberately and keep every one.
  • Unclear labels. A defect catalog with photos and limit samples aligns the judgment of inspectors.

Cleaning data after the fact is expensive and never finished. Changing how it is created costs less and lasts.

The data path

Data travels from the sensor or camera through the machine network, across a secured interface, into storage with its context: part, order, shift, machine. Each step has an owner, and every connection between the machine network and the office network is documented and approved.

Score readiness honestly, in one hour, with the process expert and the automation engineer at the table. It is the cheapest way to avoid a pilot that fails in month two for reasons everybody could have seen in week zero.

Newsletter

One Lean idea.
Every week.