The truth that emerges from new data science techniques facilitates fundamental improvements. Such is the case in maintenance practices. Break-fix, calendar, usage, condition, and reliability-centred maintenance (RCM) techniques have been the basis of 40 years of improvements. The reliability successes came from a fixation on inspection, with more complex ways to determine the inspection intervals; including assessments from RCM, which requires exceptional excessive skills, time and costs. Indeed, one VP at a refinery declared about RCM, “We cannot afford to keep feeding the beast!” Machines are more reliable: but they still break down.
Historically, improvements were fuelled by word of mouth and experimentation. They were severely limited and customary inspections generally only took care of wear-n-tear issues. In 2015, ARC Advisory Group highlighted that 82 per cent of assets suffer from breakdowns caused by errant process behaviour in ‘seemingly’ random failure patterns. So, we’re spending time and money looking for 18 per cent of the problems and ignoring the biggest cause: chance events. Such events include pumps losing feed or cavitation, an incorrect setpoint entry takes equipment outside the design and safety envelope, and compressors suffering almost imperceptible liquid carry-over that over months can lead to catastrophic failure.
Advanced technology & the future of maintenance
Predictive Maintenance and Prescriptive Maintenance (where the AI software offers advice on how to correct errant conditions) improves maintenance by supplementing it with continuous monitoring in real-time. It can also remove periodic inspection with very early detection techniques that distinguish both damage-causing process behaviour and wear-n-tear in machines. Consequent early intervention can adjust the errant process behaviour to avoid the degradation and damage – whereby no maintenance service occurs. Or, early activation of a maintenance service request can do two things. First, it may involve minor service to fix an issue that avoids a catastrophic failure. For example, early correction of a minor lubrication condition that would lead to complete bearing failure and a major overhaul. Or second, an extremely early warning, in weeks and months not just days, allows the time to plan a safe, orderly, and environmental incident-free repair with the least disruption to operational goals; clearly avoiding the scramble that often occurs after a breakdown.
Manufacturing companies recognise that breakdowns and unplanned maintenance drain profitability. That’s why scores of AI/ML (machine-learning) platforms are available for exercising ideas around Predictive Maintenance. With tools such as Python, Google Tensor-Flow, Microsoft Azure, or Amazon Web Services personnel can work diligently trying to spot the data patterns suggesting they intervene to prevent failures. Unfortunately, these tools are for experts and likely not the current staff employed in the industry.
In-house or out-of-the-box solutions?
Historically, large manufacturing companies prepared in-house solutions saving and presenting process data, for advanced control automation, resource planning, etc. Recognition of the distracted focus from company core competence along with the appearance of more capable commercial out-of-the box applications resulted in most companies ceasing these in-house developments. For example, one time working for a world-scale refiner I proposed to senior executives we should not build a solution. I proposed our core competence was to “boil oil” not to develop software. I reasoned we should articulate the problems and get the best guys to fix them; suggesting that the best software personnel do not work for oil companies, they work for Microsoft, Google, Facebook, Apple and so on. But under-developed software skills are not the only reasons for considering commercial software.
Large companies build software projects – software companies build software products. There’s a big, often overlooked, difference. Projects can be very focused on intense domain knowledge on a very specific problem belonging to the company. However, history shows us that individual projects often result in fragile code, a lack of robustness and errors. Product companies broaden the application to cover more use cases serving a wider audience. Robustness in commercial application comes from a development process to obtain the right use cases, generate traceable requirements and quality assurance and testing plan with all the rigors to assure the software performs well in diverse situations. Such techniques assure solidity and scalability exceeding a single project. All companies want applications to execute faster, easier with less cost and better support. But that’s not what they get from internal software. Consequently, developing the new brand of software applications incorporating data science does not fundamentally change the internal versus external debate.
Users must be able to deploy and maintain software fast and easily, sustain and maintain it over its lifecycle. An application must rapidly scale to blanket a whole operation for universal coverage; not just the expensive bits. To do that, products hide intense technology inside the software. Knowledge of expert engineering or data science plays no part because advanced tech mimics the iPhone model; where anyone can use it because the technology is hidden on the inside with a simple, easy user interface. In the process industry, this is a breakthrough; as other products need experts and time to implement and support.
The benefits of an advanced maintenance solution
Domain knowledge and real science rather than anecdotal references drive the way we use machine learning – because it’s not about the fact you use machine learning, it’s about “what” you do with it.
Machine learning-enabled digital agents in predictive maintenance solutions do the work so that you do not have to. After a very short training time you can build agents in minutes. Agents encompass all the smarts to sidestep engineering, modelling and statistical solution disadvantages. This provides precise pattern matching to recognise normal machine behaviour and the deviations that indicate impending failures. This process manages the amalgamation of multiple data streams to extract the multi-variate and temporal patterns describing deviations from normal to build and enable the digital agents. The agents detect impending failures and also automatically adjust to correct for changes in the manufacturing process. By measuring the exact patterns that lead to specific failures and root cause conditions agents take detection to new levels of accuracy and early warnings. Also, because agents deploy pattern detection, they are equally useful on any asset, for any industry and any failure mode: rotating and static machines, mobile vehicles, and process equipment such as heat exchangers, furnaces, etc.
The ability to use prescriptive analytics helps to improve the earning potential of the businesses using the solution by making assets more available. Typically, operational throughput increases are about four times higher magnitude than the savings in maintenance. Manufacturing facilities can expect to see 1 per cent to 3 per cent or more plant throughput due to increased asset availability. In contrast, we see maintenance savings of the order of 5 per cent - 10 per cent or more. Also, superior maintenance service management requires less planned maintenance as your organisation becomes comfortable with the accuracy of predictive alerts. Chemicals, oil & gas, mining, etc., will also see less safety and environmental event when extremely early warnings allow the time to plan interventions.
AI can improve maintenance. However, internally developing or using commercial products comes down to the assertion I made to a senior executive at a manufacturing company. We can argue about the prowess of an internally developed solution and may prove that out-of-the-box solutions work technically better. That is not the issue. It is always rooted in cost:
- the skills, time and money it takes your company to actually build it
- the resources, time, skills, and cost to deploy a solution
- the capability to scale to “blanket” a site, multiple sites, and a whole corporation
- the effort needed to sustain a solution for perhaps the 20 years you will own it.
A one-time project rarely considers such very important requirements. So, what’s the real cost of an internal solution and will it really give you a competitive edge?
Mike Brooks, Global Director APM Solutions, AspenTech