Anomaly detection: a game changing approach to predictive maintenance
Image credit: BiancoBlue/Dreamstime
Data analysis and machine learning are revolutionising the way in which the condition of crucial machinery components can be monitored.
The purpose of preventative maintenance is to avoid the costly impact of unscheduled downtime should machine components fail. In many cases the components replaced may still be within specification, but we replace them anyway rather than take the risk – especially where the equipment is critical to output.
An accusation regularly levelled at this approach is that it can lead to excessive maintenance. Operators would much prefer to know when a machine part is going to fail, so they can avoid carrying out the work until it is absolutely necessary. Ideally, this would be the exact remaining lifetime of the component, but predictive maintenance technology is some way off achieving that level of precision.
What we tend to see happening currently, when it comes to critical machinery, is a condition-based approach to maintenance that involves the deployment of condition-monitoring solutions which rely on sensors to detect signs a component could be deteriorating. Once a certain threshold has been reached – be that pressure, temperature, vibration or whatever – an alarm is triggered to inform the engineers to change the machine part.
While this is better than swapping out components periodically, it is still using fairly crude limits to determine when a part should be changed.
The emergence of anomaly detection is changing the game, however, by enabling next-gen predictive maintenance. This technology uses multi-dimensional data and machine learning to evaluate a variety of factors that can impact the condition of a component, and how they will likely impact each other – before determining whether the data indicates a genuine problem with a machine.
The benefit of this multi-dimensional approach is that it adds context to the situation. For example, it understands whether abnormalities are the result of a degrading component condition or normal seasonal changes due to production demand.
It may be that process parameters, sampled from a conveyor electric motor, show abnormal activity signatures that - under a traditional condition-based maintenance approach - could trigger an alarm for an intervention. When other contextual factors such as seasonal workload are taken into consideration, however, the algorithm may conclude that these levels are normal and an intervention is not required.
This does more than just save operators downtime for unnecessary maintenance: it can also help to identify more accurately why components are failing in the first place.
If it is found that a machine component needs to be changed just three years into a ten-year design life, for example, anomaly detection can show what factors are responsible for this. Maybe the logic controlling a motor was configured incorrectly, causing it to be overloaded. By identifying such a problem, the operator then prevents the same problem happening again and again with subsequent replacement components.
This level of understanding also allows companies to carry out fitness for service checks on their ageing machinery, or to perform due diligence on another organisation’s assets before an acquisition. The insights gained from anomaly detection can also help original equipment manufacturers (OEMs) to improve their specifications and designs. This will help ensure future equipment better meets the needs of the operators.
In many cases, without anomaly detection, organisations would need to procure subject matter expertise, at significant cost, to determine an asset’s state of health.
This technology is not just game-changing for companies looking to reduce the downtime of critical equipment. Anomaly detection is a cost-effective solution that can be deployed without disrupting output, so it can be deployed on less critical machinery, too. As long as the machine is adequately rich in data and has good connectivity, anomaly detection can be trained using datasets gathered from healthy components.
Although anomaly detection cannot yet infer what period of time a component has left prior to failure, it is a step closer to that predictive maintenance goal. By allowing us to better understand the chance of failure, it’s helping to reduce scheduled maintenance and operational downtime, improving operational safety, delivering sustainable maintenance practices and saving businesses huge costs.
Sam Burgess is CEO of SamsonVT
Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.