How To Run a Predictive Maintenance Proof of Concept
In this blog, we share some of the lessons we’ve learned from both successful and unsuccessful PoCs. We’d love to be able to say that all our PoCs have achieved their end goals of scaling, but this would be ignoring valuable insights that we can now share to help you to avoid making the same mistake.
AI and analytics are hot topics and almost every vendor claims an ability “to do” Predictive Maintenance (PdM) with their predictive analytics platform. However, the relationship with a Predictive Maintenance software vendor is typically lengthy and requires a large amount of trust. With a large amount of noise in the market, it therefore makes sense to perform an exercise to down-select potential vendors. As with choosing your doctor, you need to be certain a vendor can do what they say they’re capable of.
It often makes sense for an organization to run an exercise to ensure they not only select the right vendor, but that the organization itself is sufficiently prepared to ensure a PdM project has everything it needs to succeed, and that it is able to change and adapt to take advantage of the benefits on offer. The PdM proof of concept (PoC) can help to ensure this – if done the right way.
In our experience, many companies have tried and failed multiple PdM PoC’s before they finally achieve the results they expect. Some of our customers have tried three other solutions before achieving what they expected. The lessons learned from the failures were that often the vendor wasn’t entirely at fault - the problem actually lay within. It was only their awareness of success in other industry projects that gave them the confidence to make changes internally and persist to achieve success.
Predictive analytics ≠ Predictive maintenance
It’s important that we address a key and common misunderstanding: predictive analytics tools can be used as part of a PdM program, but predictive analytics and PdM are far from the same thing.
Data scientists can work on machine data and identify anomalies and trends – and produce some convincing screenshots. But it’s another thing entirely to have a deep understanding of what those things mean to the health of machinery, and to be able to have an in-depth discussion about what maintenance strategy to adopt with this new information. Pure data scientists usually lack the experience and knowledge needed to act as maintenance engineers. As a result, the custom algorithms they devise often perform poorly in real-world industrial conditions.
Spotting a vendor more used to predictive analytics than predictive maintenance
Special machines may have special and unique failure modes, but the failure modes and the kinds of information needed to detect them in most common machinery, such as motors, gearboxes, and robots are very well understood from a condition monitoring perspective. Failure often “is what it is” so, if a vendor is asking basic questions about machine failure modes and asking you to define everything, it’s clear they don’t have a background in condition monitoring or machine maintenance and don’t know what they’re doing. Chances of success are very slim.
Asking you to label everything at an early stage – and being unable to understand what’s shown in your maintenance logs and make correlations - can mean you’re dealing with a vendor who’s going to be taking a bespoke modelling approach. This can have very good results for up to tens of machines but will struggle with scalability and cost for anything more.
Maintenance is the ultimate practical ‘hands-on’ discipline, while data analytics is rooted in application of theory and advanced mathematics. The PoC should be the point at which those two worlds converge to deliver measurable business benefits through PdM.
Finding the right PdM supplier is tough – they need to understand your maintenance practices, your machines and what business outcomes you are looking to achieve and they need to be as committed as you are to achieving those. They need to do this through application of the right theories and mathematics – and do so in a way that is economically and organizationally scalable.
For more on best practices and common pitfalls to avoid, download our full white paper "How to Run a Predictive Maintenance Proof of Concept" below.