Predictive analytics is the catch-all term for any technologies which can help you to predict the future by analysing past historical events as well as real-time data feeds. It’s had some particularly successful case studies (‘Other customers bought XYZ’ on Amazon) as well as some potentially unnerving ones (Target predicting a girl’s pregnancy before her family knew).
It’s generally an application of some statistical analysis, perhaps with machine learning to particular data sets. Whilst it arguably has its greatest successes in areas like spotting credit card fraud and understanding behaviour for marketing analysis, an increasing number of predictive analytics platforms are getting vocal about applications in predictive maintenance.
Predictive analytics does not equal predictive maintenance, in fact most analytics platforms are poorly suited to forecasting machine failure due to the non-linear way in which failure happens.
Not all failures are created equal
Predictive analytics is well suited to linear problems such as inventory flow – as an example and in simplified terms you can identify and calculate a formula for stock reduction over a given period and use that to extrapolate when your stock will hit a particular reorder threshold. In an ideal world, machine failure would be linear, sadly we don’t live in an ideal world..
Let’s take a very example of some very simplified condition monitoring data showing bearing failure:
Spotting the point of failure (for a human) is quite obvious – the signal (in this case vibration) has quickly risen and sharply fallen as the part we’re looking at disintegrates. What’s particularly interesting (and troubling) for most predictive analytics techniques is what happens before:
Whilst there is an initial dip in the signal (orange), the signal quickly starts to normalise even return to a ‘healthy’ (green) pattern similar to the signal that we’d seen before. This green point is where predictive analytics alone would fall down – the trajectory of the failure has recovered – so everything must be OK:
Those of you versed with condition monitoring will understand that what the data has shown is in all probability a metal shaving or some other part separating from the bearing as it starts to degrade – as the shaving separates, the signal normalises. Predictive analytics alone cannot pick this failure mode up as it has no understanding of the failure mode and cannot understand the context in which it occurs.
This is where prognostics comes in. Prognostics is about forecasting the risk and type of failure before there is an operational impact (the bearing disintegrating and causing downtime of a machine). Prognostics combines condition monitoring with machine-learning and statistical techniques, crucially tied together with real-world failure models and signatures – context.
Prognostics can be easy to confuse with condition monitoring but it’s a step beyond – rather than focusing on what is going on ‘now’, it uses condition indicators and signatures of past failure to project what will happen in the future and calculate a ‘Remaining Useful Life’ of your machinery – allowing you to accurately implement predictive maintenance!
Free prognostics white paper
Our automated condition monitoring and prognostics software helps you to enable predictive maintenance and reduce unplanned downtime without having to be an expert in condition monitoring and prognostics. Prognostics really is the future of condition monitoring; if you’d like to learn more we’ve put together a free white paper, grab yours here: