Social Sharing block
In high-volume manufacturing, inspection often struggles to keep pace with production. That challenge becomes even greater when the material itself is naturally variable and the defects aren’t always easy to define consistently. In one high-volume manufacturing environment producing about 20,000 unassembled parts per day, that was exactly the problem: Quality teams were expected to identify substrate-related wood and milling-process defects at line speed, even though many of those defects were subtle, visually inconsistent, and open to interpretation.
|
ADVERTISEMENT |
Before AI-based inspection was implemented, the process relied heavily on manual judgment. Some defects were obvious, but many were not. Parts with similar conditions could be classified differently depending on who inspected them, when they were inspected, and how much time was available to make a decision.
Under high-throughput conditions, that inconsistency created two problems at once. Inspection generated too many questionable calls while still allowing a substantial number of true defects to escape downstream. As a result, the operation was dealing with unstable quality decisions, inconsistent data, and higher downstream costs.
…
Comments
Terrific article
There are little clues here and there to suggest that this article was maybe written with the assistance of AI, but it ultimately doesn't bother me. The meat of the article is excellent and exactly what I want to see more of from Quality Digest: an application of a new technology towards an enduring problem, how it went, and what was learnt.
I've worked in environments where, as in this article, the categorisation of defects could be described as a "taxonomy": a lot of things that could be wrong, a lot of locations where they could be wrong, and a lot of ways that those things could be wrong in those locations. Because of the granularity, it is usually the case that a specific kind of defect will be a rare event, which adds some limitations and complexity to the use of SPC in analysing the data. There are tricks for overcoming these limitations and complexity, but after you add in the variability of categorising the defects in the first place, it tends to become extremely unprofitable to invest time in slicing and dicing the data to look for patterns: there's too much noise in the inspection process itself. In my experience, such data can provide critical evidence for unfucking some problem that's already known about, but the real power of process improvement comes from minimising variance and preventing problems in the first place: i.e., discovering that there is something to discover because of a signal of nonhomogeneity in the data stream.
Even if you don't use AI to make sense of the resulting data, simply using it on the front end to eliminate variability in inspections is a worthwhile goal in such an environment. Beyond that, I'll bet that there's enormous utility in using AI to identify which kinds of defects trend together or which are variations on a common issue: something I'd be very hesitant to trust an AI model to do if I had doubts about the reproducibility of the inspection process.
Add new comment