AI, Anomaly Detection & Asset Management – Keeping it Real

Even if anomaly detection is in place, and functioning as expected, the failure can still occur.  Anomaly detection alone is not enough.  Implementing a holistic asset strategy and work process which leverages time-based activities, condition monitoring, and anomaly detection is essential to truly optimize risk and costs.

Back in the mid-1990’s members of the Itus Digital team were involved in a project to consolidate service intervals of a heavy haul fleet in an open pit coal mine.  Rather than spend ten or twelve million dollars to increase the size of the haul truck fleet, we wanted to see if we could achieve an increase in production capacity from the existing fleet. One of the requirements of the project was to double the engine oil drain interval.  This would allow the service intervals to be consolidated sufficiently to keep the existing haul trucks working for longer and unlock un-utilized production capacity.  However, we needed to have confidence that we could increase these oil drain intervals without increasing the risk of engine failure.  The things we needed twenty-five years ago are the same things we need today:

• Understanding of failure modes
• The protections that needed to be in place to sufficiently reduce the risk of failure
  • Doing the right things
  • Monitoring the right things
• Historical data about:
  • How the lubricant breaks down
  • How the engine can fail due to lubrication-related problems
• Analysis of used oil and filters to justify a PM extension

 

If we’d been able to apply deep neural networks back then, to watch one or multiple conditions, and detect anomalies that indicated potential failure risk, we might have reduced the duration of the project by several months and achieved a lower-cost way to monitor certain failure modes with higher confidence.

Unfortunately, we experienced two engine failures during the project.  These failures were directly attributed to maintenance activities that were not done when they needed to be done.  One of the failures occurred because the lubricant hadn’t been topped up.  The second failure occurred because the engine lubricant filter hadn’t been properly serviced.  In this scenario, anomaly detection would not have detected these failures and highlights the need to treat Artificial Intelligence (AI) and Machine Learning (ML) as a specialized technique to complement foundational asset management practices such as preventative maintenance compliance.

With ChatGPT, AI/ML has reached a new level of hype, promise, and even a dose of fear.  Have we reached a point, like a Star Trek episode, when we can simply ask the ship’s computer to answer questions, provide options, or execute certain tasks?  While ChatGPT is a powerful tool that can quickly generate an answer to your question, it has no mechanism for telling you when the right time is to apply the solution.  An effective asset strategy requires more than just a simple answer to a question.  It must be operationalized, dynamic, and the enabling capability of your work process.

The anomaly detection concept involves picking up an incipient failure signal, further leftward on the IPF curve.  This simple concept implies that the earlier we detect the deviation of the expected signal, the more time we’ll have to do something about it thus, minimizing potential impact to operation.

This isn’t a new technique in the industrial space.  Mathematics and models of failure have been around for decades.  There is a vast amount of data available to feed these models and produce the decision support we need.  The role of AI/ML is to bring down the cost of and increase the accuracy of prediction by efficiently analyzing large volumes of data and advising equipment experts when failure risks emerge.  It should also bring about a more optimal division of labor between humans and machines.

Industrial equipment can have significant variation in risk. Failure modes exist which are not optimally detected with AI/ML, and every organization must work within budget and cost constraints.  Furthermore, when assessed appropriately a run-to-failure strategy could be the optimal answer.  Consideration of all these factors, the classic constructs of asset management, is more important than ever given the level of hype given to AI, as the singular answer to optimizing asset performance.

In summary, Asset Performance Management requires more than just anomaly detection or understanding of when maintenance activities occur.  Asset Performance Management helps organizations balance the risk, cost, and performance of assets.  As such, anomaly detection may be an important part of an operationalized strategy, if it is in service to that outcome.