In a recent post by ARC Advisory Group, Peter Reynolds notes that 80% of assets fail randomly despite being supported by programs designed for asset maintenance and reliability. Only 3-5% of maintenance performed is predictive. The vast majority of maintenance is either break-fix or executed based on the OEM’s asset maintenance schedule – needed or not.
A broad set of factors drive asset performance, including variabilities in process conditions/flow outside the asset itself, which previously may not have been considered relevant to determining asset condition. With advanced analytics, the compute power is available to combine asset health, asset condition, and process variables to determine the asset’s true risk of failure.
More importantly, machine learning will provide a means to see beyond a conventionally-understood state leading to asset failure. These machine learning models require an understanding of the operating and failure mode states of these assets. As Reynolds points out, this probably means working with operating personnel, not maintenance personnel, to develop the models. This marks a change from condition-based maintenance and less sophisticated predictive models.
Using sophisticated machine learning models, asset managers can know that a given asset will continue through a rough spot, not fail as might have been predicted by condition monitoring or prognostic models, and will in fact go on to a longer operation. This suggests that the P-F curve in ARC’s post could look more like a sine wave than a gradual drop off. The key is to have confidence in the algorithm’s prediction that failure is actually not imminent. Only the right set of machine learning analytics can predict into the future without a loss of confidence.
Predictive and prescriptive analytics will indeed drive the next wave of improvements in asset performance. But only the right algorithms will provide the highest return on investment for those seeking lasting improvements in asset performance.