Reducing Manufacturing Scrap Through Root Cause Analysis

The Opportunity

A discrete manufacturer produces a technology product which has narrow measurement tolerances. At several steps in the long production process, these tolerances can be exceeded, resulting in scrap product. This can cost them tens of thousands of dollars per batch.

Alerts to possible issues are generated, but not all of them result in scrap or rework decisions. The alerts are sent to technicians who, either through experience, or by consulting a lengthy reference manual, make decisions about the particular lot, based on possible root causes.

This can be time consuming and slows production. The manufacturer theorized that machine learning could help automate decision making.

Our Approach

The manufacturer had very good metrology, sensor, and image data. Importantly, they have kept track of the decisions that technicians had made to resolve alerts.

With this data we explored five different machine learning and deep learning approaches to see how well these could emulate the technicians’ decisions.

The Impact

The best approach, which was an ensemble of classifiers, was able to correctly predict technician decisions 85.9% of the time. Furthermore, we used “explainable AI” methods to rank and quantify the most likely root causes for the decision.

The company had done a financial analysis estimating the cost of the technician time in resolving the issues, as well as the impact on productivity. Globally, this was costing them tens of millions of dollar annually.

By implementing deep learning-based root-cause analysis, they are now able to make technicians more efficient, targeting those tens of millions of cost savings.

In no case was it ever suggested that AI would replace technicians, but rather, their skills would be augmented to make them more efficient.