What
This project compares the rejection rates of pressure relief valves from multiple vendors and service mediums and draws a predicted failure rate to inform buyers on how many valves might need to be purchased to guarantee at least one functional vale.
Why
When procuring parts for aerospace, ensuring the reliability of the final deliverable is paramount. In some cases, such as satellites, failure of parts is unacceptable as the time to design and manufacture is long and the amount of final deliverables is low. This means that every mission critical part must work on launch day. Unfortunately, in the case of pressure relief valves, this is not the case. According the ASME study “Reliability Testing of Pressure Relief Valves”, many vendors send valves that fail testing for a variety of reasons from leaking and higher or lower pressure than required, incorrect stamping, or just delivering the wrong valve.
This means that if you need one functioning valve you might need to order several valves to ensure that you receive one working valve. This project uses Bayesian statistics to model failure rates per vendor and service medium and eventually predict the number of valves that one would need to purchase to ensure that at least one working valve is received.
How
I'd like this section to be a bit more thorough that many of my other projects. For now you can look at the graphs I've posted here while I write up an adequate summary of the work that was done.
What Next?
This project was used to further my understanding of Bayesian statistics. Bayesian statistics is hidden in many parts of both modern control systems (like with the Kalman Filter) and in deep learning. While reading papers about SLAM and its evolution over the last 20 years, I found that many of the algorithms relied on statistical knowledge that I had not been introduced to yet. This project gave me a better understanding of how this branch of statistics is used and how to apply it to both control systems and machine learning.