Successful predictive maintenance isn’t about algorithms or assets, it’s about users

Experienced maintenance engineers have a detailed mental picture of the machinery they care for. They know when a rattling valve means a breakdown is imminent or when it’s safe to ignore it until the next scheduled shutdown. If an automated predictive maintenance system can tap into this pool of knowledge, it can learn to provide the best possible support for users who must decide which maintenance activities to prioritize.

In a user-centric approach to predictive maintenance, maintenance teams are alerted only when the system thinks they will find that information useful, based on what they have found helpful in the past.

It’s a bit like when digital content providers such as Netflix or Amazon routinely store information about what each user chooses to watch. They can use that to drive a ‘recommendation engine’ that refines the movies it shows to each user continuously in response to their feedback.

This is the approach we’re taking at Senseye in order to model and understand user behavior in addition to what is happening with the assets that they are monitoring, to help direct their attention where it’s most needed.

High and low context environments for machine data

Figure 1: Environments rich in contextual and high-quality condition monitoring data allow the assets to be perfectly monitored - like a Digital Twin approach. Most factory environments are low-context and so models of user interest can improve predictive maintenance software efficacy 

The power (and limitations) of prediction

Every predictive maintenance initiative – from using an iron rod held to a rattling machine to understand if the gearbox needs a service before its next scheduled maintenance interval to a sophisticated automated software monitoring system - seeks to help operators identify when trouble is brewing. The chief aim is to flag up issues early enough to prevent a breakdown that would otherwise lead to costly unplanned downtime. There are often other benefits too, related to improved productivity and maintenance planning.

The concept of the ‘digital twin’ is particularly popular at the moment – yet the oft-muddled marketing of this mathematical modelling approach makes little mention of the users who are meant to interact with these ‘twins’ and what that means to them. A ‘digital twin’ of each asset also needs the user to be comfortable working in a complex digital environment. But there is another approach that focuses much more on the needs of users.

Many failures leave distinct signals or ‘fingerprints’ in machine data and leading predictive maintenance software on the market right now can learn to spot them by accepting a wide range of inputs – from dedicated condition monitoring data to more general plant data. These systems can understand which patterns or characteristic signals indicate that there may be a problem. In some cases and with sufficient data, they can even calculate the remaining useful life (RUL) of each asset – a technique known as prognostics.

Most of these predictive maintenance systems work with limited data from the factory or plant floor, triggering an alarm when a pre-set threshold is breached. They might alert users when something is heating up or vibrating, for example, but they are unlikely to have enough information to make a detailed diagnosis.

In other words, the system can ‘advise’ by raising the alarm, but only the user has the experience and expert knowledge to decide when to act. Crucially, it becomes much trickier to manage the situation in environments where many machines are being monitored at once, because users can easily become overwhelmed. The challenge with predictive maintenance is not ‘can you spot issues in the data’ but ‘can you spot what is of interest to the users?’

Who cares about the user?

We feel it’s important that when we raise an alert, the user can indicate at the touch of button whether that alert is useful or not. Over time, this teaches the system to direct the operator’s attention towards the most pressing maintenance priorities. This is done automatically, so operators don’t need any expertise in data analytics.

In other words, Senseye’s user-centric approach to data analytics guides attention by modelling the response of the user. This feedback loop between Senseye’s solution and the user means that the number of alerts gradually falls away until it reaches a stable level where almost everything it directs the user to look at is helpful.

Proven results from advanced analytics – for machines and maintainers

Data analytics is about searching for patterns within the mass of incoming plant data. For instance, the proprietary algorithms, or pattern engines, behind Senseye PdM are designed to spot the characteristic behaviors that precede potential breakdowns and to trigger an alert using our proprietary Attention Index. Several types of pattern can trigger an Attention Index alert: anomalies are periods of unstable data, trends are gradual shifts in the baseline and violations exceed thresholds or rules that can be specified by the user.

Yet all this ingenious data analysis means nothing unless it delivers genuine business benefits for users. Thankfully, Senseye PdM has a strong track record of delivering on its productivity promises across a range of industries. Existing customers typically enjoy a 50% reduction in downtime, 55% increased productivity and 85% increase in maintenance accuracy.

What’s more Senseye PdM is also backed by Senseye’s ROI Lock™ guarantee: If deploying Senseye PdM fails to reduce unplanned downtime as agreed upfront, customers can claim a refund on their entire subscription fee.

If you’d like to learn more about how we model the maintenance users as well as the assets they watch, we’ve put together a detailed white paper on how Senseye PdM can help direct your maintenance efforts to where they’ll do the most good. Download the detailed white paper below for more information or get in touch to see a demo and get started!

PdM isn't about algorithms