I’d like you to imagine that you’ve been asked to provide a reliability prediction, but you don’t have a final product, or any data. What would you do? Chances are, you will end up using a prediction standard or a generic database of failure rates, this is totally acceptable given your situation, but do you know how your choice of data limits the utility of your results?
In this article I will be focussing on where your failure rate data might be coming from, and I will outline some considerations to keep in mind when you perform reliability predictions.
It is important to note the two key aspects of any reliability prediction; the model, and the data you feed into that model. When I say model, this could be as simple as summing your failure rates together (assuming they are in series), or a complex RBD (shown below). Regardless of how your model looks, if you are aiming to get a quantitative reliability prediction,there will always be data flowing into your model somehow.
We must acknowledge that there are many sources of failure rate data out there, in practice however they may not accessible, relevant or appropriate to your product or analysis. For this reason,I’ll kick things off by listing a hierarchy of data sources for failure rate data:
Using In-house field/test data:
- Here, failure data is collected on the same (or similar) products, as it is used in the field, for known operating conditions.
- This is the ideal scenario, but this is usually not the case, and is almost impossible for a brand-new product.
Using Supplier field/test data:
- This is similar to “In-house” above, but the data is collected by the equipment supplier.
- This is slightly less preferable as you do not have full visibility of the data sources, and may be a little bias.
Physics of failure(PoF) methods:
- Failure rates are made from mathematical models of the physics surrounding failure.
- If the operating conditions are well known, then this method can yield a reasonable approximation of what can be expected in the field. You can imagine this as simulating failure through some complex modelling.
Standards based prediction
- Failure data is first collected from various sources and then modified based on calculations defined in a standard. Examples of this include MIL-HBK-217F, Telcordia and FIDES.
- The reliability community is divided on the use of these methods as they are well known for producing inaccurate (but not useless) results.
Generic failure rate database
- Failure data is collected from a variety of sources and collated into a searchable database, examples include NPRD, OREDA, FARADIP.
So that is our general hierarchy of failure data sources. But depending on various factors, some may rank higher than others, and this hierarchy may start to shift. For example, if you have some in-house data (top of our list) but it is outdated and possibly unreliable, then is it any better than a generic database (at the bottom of our list)? To help answer this question I have outlined some points to keep in mind when you sourcing your failure rate data.
1. Data sources
- Where has the data come from?
2. Type of data and integration with current systems
- Is the data relevant and sufficient for our analysis?
3. Industry specific data
- Is the data relevant to the application/industry of the prediction?
4. Maintenance and accuracy
- Is the data maintained regularly/at all (the database could be outdated)?
- Is the data accurate?
5. Security and trust
- Do we have confidence in the data we are receiving?
By combining this list with our hierarchy of sources we can be more informed when selecting a data source. As a demonstration, I have populated the table below, let’s imagine I’m working on an electronics product, and the bottom two rows represent MIL-HBK-217F and NPRD-2016 respectively.
Ok, so back to the problem, you have no data, your suppliers have no data, and let’s assume you don’t have the tools to perform PoF, then from our table above, it looks like Standards Based Prediction is our best option.
That’s it settled then; we can continue with our reliability prediction! Well, yes, but let’s proceed with caution. The main point here being, what will your prediction tell you? Considering our table screams ‘low confidence’, I wouldn’t recommend you use this result for estimating warranty or maintenance periods. However, besides ticking off a requirement* you could use this result to compare various designs of the product, and flag up any areas that show excessively high contributions to unreliability.
*Reliability should never be a tick box exercise but the reality is that often requirements kick start
Reliability activities…it’s almost as if this is why reliability has become a requirement.
My aim from this post was to get you thinking about where your failure data comes from, and to then consider how this impacts how you use your reliability prediction results.I hope this has inspired you to start collecting your own data to help you to climb the hierarchy of data sources so that your reliability predictions can be more insightful in future.
Contact us on +44 (0) 333 996 9930 or email: email@example.com to discuss your specific needs.