Vibration analysis is all about looking for patterns in complex datasets – something artificial intelligence (AI) excels at. In conventional condition monitoring, a baseline vibration signature is recorded when machinery is known to be in good condition. Tolerances can then be set on the overall level of vibration as well as frequency bands known to be characteristic of specific fault conditions. When a vibration level goes out of tolerance, automated condition monitoring software can notify maintenance staff of an issue for diagnosis and repair.
While AI is starting to play a role in diagnostics, it generally requires well-structured problems and large training data sets. For both of these reasons, AI remains a long way from fully replacing the skilled vibration analyst, and is probably better seen as a tool to enhance the ability of the analyst. A human analyst is able to look around and notice things about the environment a machine is operated in and the process it is performing. The ability to apply logical reasoning, thinking laterally, with relatively little data, makes human reasoning superior in this situation.
Rajet Krishnan, CEO of Viking Analytics, states: “If you’re trying to replace or simplify or speed up the activity of diagnosis of a problem with the machine, to some degree, that can be automated. But what we have seen is that the level of performance of many of these diagnostic tools – which operate based on rules or automation of the rules – can be fairly poor, and the performance is about 30-40%, depending on which machine you’re talking about, under which conditions.”
He continues: “The second part, which is where automation or AI is being talked about, is in the area of recommendation or prescriptive analytics. And we have seen that the performance of most of the systems out there is less than 10%, for the simple reason that vibration data alone does not have all the information that is needed to make a useful recommendation. A lot of the information is actually in the context in which the machine is operated, the process under which the machine is operated, kinematic information related to the machine, the way the machine interacts with other components in the machine train, and so on and so forth.”
BIG DATA
Where AI comes into its own is sifting through very large quantities of data that would be simply overwhelming and far too time-consuming for a human. Where there are thousands of machines or millions of signals being monitored, it’s clearly not feasible to have a skilled analyst watching them all.
Traditional condition monitoring solves this issue to some extent through the use of simple thresholds. These involve recording a baseline vibration signature when a machine is in a fault-free condition. When the amplitude of vibration in a particular frequency band goes above a threshold, the machine is flagged for attention by the analyst. AI can provide a half-way-house between simple thresholds and the role of a human, providing additional screening to detect faults earlier and flag them for the attention of an analyst, or direct the analyst more quickly to the area needing attention.
Krishnan adds: “Vibration analysts, machine experts, or maintenance experts, who understand the machine and the process under which they operate, still play a very critical role in terms of deciding what is to be done with the machine and how it should be prioritised and operated. And we expect that this will continue for very many more years to come.
“What we think is a more scalable approach, a more widespread approach is to augment the vibration analysts in a way that accelerates their workflow in today’s world… Vibration analysts are quite critical in the process of root cause analysis, diagnostics, and recommendations. And most of the time, a plant expects the vibration analysts to provide them with concrete recommendations in terms of which machines to focus on and what is to be done with the machines.
He suggests that an AI support tool could help analysts pinpoint problems, rather than having to search through trend data.
This type of more refined screening can be achieved without AI. For example, Schaeffler produces systems that aim to automate vibration analysis, some of which use the simple ‘black box’ threshold approach, while others go further to consider how a machine works. Automated vibration analysis software can be given parametric information for standard machine elements, such as bearings, motors, fans, pulleys and gears. The software can then determine defect frequencies as multiples of shaft speed, and therefore determine low pass filter and resolution requirements, as well as indications of faults in specific components. Where the software isn’t able to diagnose a fault, it can at least produce pre-configured time series and spectral plots that are relevant to the components under consideration.
UNDERSTANDING RELATIONSHIPS
This kind of deterministic approach is fine when the relationship between the kinematics of a machine and the vibration signature of a fault is easily understood. However, there are many cases where complex data from multiple sensors is available and an analytical approach doesn’t easily result in a known fault signature. This can be an excellent problem for AI to solve, provided enough data is available.
AI requires a lot of data, identifying examples of machines in healthy conditions and in fault conditions. Given enough data, correctly labelled, AI should be able to find the patterns that indicate fault conditions, without needing any underlying understanding of the physics of the machine. Obtaining such data is sometimes problematic because of the way vibration data is managed. Companies rarely share vibration data, meaning nobody has a large enough dataset to train algorithms accurately. However, companies such as Rolls-Royce and Caterpillar have demonstrated the advantages of operators making their data available to the machine builder, enabling them to aggregate this data and obtain meaningful insights from these larger training sets.
USE IN WIND TURBINES
Niels Jessen, wind turbine performance analyst at RWE Renewables, comments: “Artificial neural networks are particularly useful when the underlying mathematical relationships of a system are not known in detail, but large amounts of operational data are available. All of this applies to wind turbines. The SCADA system records the operating and environmental conditions of wind turbines, typically in 10-minute intervals. In modern wind turbines, the SCADA data can contain hundreds of signals, including all sorts of temperatures, pressure data, electrical quantities from several components, tower vibration etc. The datasets often contain the average, the minimum, the maximum and the standard deviation values for each quantity and for each 10-minute-interval.”
If you’re looking to replace skilled analysts by automating what they currently do with AI, you’re going to be disappointed by the results that AI gives you. However, if you use AI to enhance the capabilities of humans, then it can be a powerful tool. The best results are achieved when humans do what they do best – thinking laterally to consider everything that might be affecting a machine in its actual operational environment, while AI does what it does best – identifying patterns in large and complex datasets.