Evaluating nonconformances for escalation using Bayesian methods?

[Tagin]

When a nonconforming product comes back from a customer or is found in production, a question is always looming: is this occurrence a symptom of a trending problem with this product?
E.g., if a circuit board comes back from a customer and has a failed transistor, does this one occurrence suggest that there may be a trend requiring escalation, or does the board just get reworked and sent back to the customer, and the nonconformance logged?

Clearly, if the same issue recurs multiple times we think we see a 'trend', and our suspicions are increasingly raised that these are not random component failures. That is, our estimation of probability that this issue calls for more investigation, escalation, etc. increases with the number (or percentage) of these nonconformance occurrences. How do we know when to act?

So it seems then that this should be a place where Bayesian methods could be applied to assist intuition and experience in determining the likelihood of the need for additional action with increasing evidence. Are there 'standard' or typical Bayesian methods for doing that with nonconformances specifically?

 

[Jen Kirley]

When a nonconforming product comes back from a customer or is found in production, a question is always looming: is this occurrence a symptom of a trending problem with this product?
E.g., if a circuit board comes back from a customer and has a failed transistor, does this one occurrence suggest that there may be a trend requiring escalation, or does the board just get reworked and sent back to the customer, and the nonconformance logged?

Clearly, if the same issue recurs multiple times we think we see a 'trend', and our suspicions are increasingly raised that these are not random component failures. That is, our estimation of probability that this issue calls for more investigation, escalation, etc. increases with the number (or percentage) of these nonconformance occurrences. How do we know when to act?

So it seems then that this should be a place where Bayesian methods could be applied to assist intuition and experience in determining the likelihood of the need for additional action with increasing evidence. Are there 'standard' or typical Bayesian methods for doing that with nonconformances specifically?

I can imagine some are using Bayesian methods to define a trend, but others decide based on what the failure would mean to the organization. Some types of products are so highly visible and critical to safety that even a second failure might not be tolerable. The organization decides.

 

[John Predmore]

The definition from Wikipedia states "Bayesian inference is a method used to update the probability for a hypothesis as more evidence or information becomes available. One example I thought of, in line with your question, is the case where customers reported 10 cases of non-conforming product. Let's imagine your company manufactured 10,000 widgets, but 4000 are still in a warehouse and not in service. Is the failure rate 1:1000 or 1:600? New information is learned that a shipment of 500 suspect parts was the cause of the problem, before the the problem was corrected at the supplier, but you don't know how many of the 500 suspect parts were actually bad. You know the date code when 500 suspect parts were used in the assemblies you made. If you knew what percent of the product in the field was comprised of those date codes, and what percent in the warehouse, you would have a more accurate estimate of the true failure rate in the field. This example is not a simple problem, but illustrates how Bayesian analysis of additional information can improve your decisions. With a better estimate of the true failure rate, your firm could decide based on the cost/benefit of a product replacement campaign versus deal with customer complaints as they arise.