Online polling conducted by CP late last year shed some light on the success of plants in achieving predictive maintenance. While more than a quarter of respondents said their sites had attained high or very high success, a larger number reported only moderate success, and over 15% called success low or very low (Figure 1). Several reasons account for the disparate results, say vendors, who also suggest a number of ways to improve outcomes.
Operating companies generally realize that predictive maintenance promises greater uptime and, thus, increased production capacity, along with lower costs from planned maintenance versus reacting to asset failure or performing unnecessary work, notes Michael Risse, vice president of Seeq (Seattle, Wash.). The main barrier to adoption is supporting the lifecycle, he says. “Chemical companies need to determine what data matters, collect it, find the event/predictive signal that matters, and take action on it in timely fashion. They need to get all the pieces in place from technical to process details of who does what, and often more importantly determine how to use predictions in the most profitable manner.”
This could be as simple as monitoring and analyzing data already collected or as complex as implementing a whole new set of wireless sensors with a cloud structure for data storage and analysis, he notes. Meanwhile, costs of advanced analytics are coming down and the simplicity of plug-and-play deployments is increasing — prompting pressure from executives to leverage these new approaches.
Even so, Risse believes the real question is, given equipment health status and profitability/cost data, how can plant staff decide in the very short term about what to do with these insights (Figure 2)?
“Yes, maintenance needs to be performed at some point, but should they shut down the asset, line or plant now, or should they delay it? Can they work around it for a period of time, or must the work be done now? The data is already overwhelming. More of it because it’s cheaper won’t make it better; what makes data better is knowing what to do about it by quickly finding insights.”
Over the coming years, he foresees technology development leading to faster analytics (likely on the cloud), broader consideration sets of relevant data (contextual, profit, “what if” analysis) and a better execution process (distribution of insights to the right people, on mobile devices, tracked in system, and assisted perhaps by technicians with skills enhanced by virtual reality experience).
He can point to many success stories. One involves a user that wanted to gather more real-time data on valve performance to determine when problems were developing. Seeq now takes the snippets of data produced by the diagnostic platform, historizes them, and looks for trends related to any critical variable. “Rather than waiting for offline test results, valve condition is now evaluated in real time, with corrective action taken quickly,” notes Risse.
Another involves a refinery reactor bed. Here, the user had difficulty predicting fouling due to the variety of conditions that could cause build-ups. Data from various sensors tended to be noisy, making it difficult to identify actual trends and take corrective measures. Seeq used its data cleansing and capsule technologies to eliminate noise on the fly and identify critical operating conditions contributing to accelerated fouling.
“Benchmarks created during steady-state conditions make it easy to determine when problem situations are forming, and to perform predictive maintenance,” he says.
Risse advises chemical companies pondering predictive maintenance applications to start now: “It doesn’t take rocket science, just advanced analytics to investigate historical data and use regression to find what happened, and prediction/monitoring solutions to determine what will happen. Using existing historian data, chemical companies can start now on a trial, and move from there to broader deployments after success is proven.”
One site that can attest to the benefits of predictive maintenance is the Stenungsund, Sweden, plant of Borealis (Vienna), notes Aspen Technology, Bedford. Mass. Failures of the hyper compressor used in its low-density polyethylene process resulted in high maintenance costs and plant shutdowns. Borealis deployed machine learning, via Aspen’s Mtell prescriptive analytics software, and was able to get advance warning of the repeating failures about four weeks before they occurred.