Predicting success with your predictive-maintenance programs
By Aaron Merkin, Fluke Reliability chief technology officer
Since the Industrial Revolution, maintenance managers have struggled with equipment failures that lead to costly downtime. Perhaps you are one of them. And while the problems of generations past are still common today, Industry 4.0 techniques are providing fresh hope for cost-effective, scalable solutions to enable maintenance managers to predict problems and mitigate them before they lead to unplanned downtime. Perhaps you are one of those maintenance managers.
The growing streams of condition-monitoring data from wireless sensors, handheld tools, supervisory control and data acquisition (SCADA) systems, and more, combined with modern software and analytics, are setting the stage for a ground shift. Even today, predictive maintenance lets maintenance leaders monitor real-time asset condition data to prevent failures before they occur.
Yet the artificial intelligence and machine-learning technologies on the horizon promise to deliver an even greater era of connected reliability. In the past, data collection and data analysis have been two distinct processes. While data collection has been scalable, the cost and limited availability of experts to perform analysis has made effective coverage feasible for only a subset of critical assets. In the future, sensors and software will not only work together to predict machine failures before they happen—they’ll also help teams diagnose the problems and recommend solutions. This practice is being called prescriptive maintenance or prescriptive analytics.
Machine learning vs. business intelligence
Today, a number of maintenance leaders are already applying business-intelligence (BI) techniques to enable analysis and improve decision-making. BI includes visualization tools that make it easier to tap into existing data, and is an attractive approach to many maintenance organizations because their companies are already using a BI tool and the systems can be relatively user friendly.
While business-intelligence tools and techniques are useful for collecting, visualizing and analyzing data, a human expert is still required to perform the analysis, limiting the breadth and depth of coverage. Incorporating machine-learning (ML) techniques addresses the scalability limitations of BI-based approaches by relying on algorithms to perform the analysis, rather than human experts. These algorithms can be developed through a variety of approaches, including unsupervised learning techniques that dramatically expand the scenarios that can be analyzed and the accuracy and precision of predictions.
Utilizing machine learning in today’s maintenance world means identifying the right datasets that can help predict equipment failure in the first place. Only then is it possible to create a repeatable machine-learning process that can capture, analyze, diagnose and address business problems.
Dealing with big data
There are innumerable data sources available that provide information relevant to analyzing asset health and enabling predictive maintenance. Examples include route-based data collected by technicians using handheld tools, integrated and third-party sensors, process data, and more. More teams are dealing with more data from more sources than ever before. Often, a team’s existing data is not suited for machine-learning efforts—and teams shouldn’t simply try to make their data fit.
Fortunately, there are a number of companies are now making sophisticated condition-monitoring sensors that can help get organizations started on this journey. With today’s sophisticated condition-monitoring sensors, teams can have around-the-clock measurements providing a clear picture of asset health and performance—without expending extra labor. Vibration sensors and other condition-monitoring devices make it simple for teams to troubleshoot assets quickly. Measurements from wireless condition-monitoring sensors can be automatically sent to the cloud, which reduces both errors and the need for hands-on labor.
How machine learning changes asset reliability
Vibration sensors are a common starting point for new condition-monitoring programs, but other possibilities include ultrasound, oil analysis, thermography, and motor testing. Knowing how an asset fails—its primary failure mode—tells you which indicators to watch and therefore which measurement approach is most appropriate for the asset.
With machine learning, assets have their data collected, analyzed, and diagnosed before a person ever needs to engage. An engineer verifies the diagnosis before the work order is assigned by the system. After completing the repair, a technician can input their findings. The algorithm can learn from the inputs and validation it receives.
But how do teams get there from where they are today?
Getting started with machine learning
Instead of rolling out an organization-wide machine learning program, starting small and scaling up is likelier to lead to success. No two organizations will follow identical paths, but these general steps will benefit anyone, no matter their starting point:
1. Complete an asset-criticality analysis—Prioritizing asset health and maintenance on a hierarchy of importance gives your team focus and direction. Grade each asset by its use within your specific organization, including the business impact when it fails. An asset-criticality analysis generates insights into which assets are prime candidates for condition monitoring.
2. Plan a pilot program—Start with a small set of assets identified above. Identify the most common failure modes and the appropriate condition-monitoring technology to detect them. You will need a sustainable, repeatable process for data collection. Identify the condition monitoring sensors and strategy (route-based or continuous) that suit your needs, scale, and budget. Determine your model creation approach. Options include onboarding data science resources, contracting out development to a third party, or adopting off-the-shelf solutions.
3. Establish a framework and evaluation criteria—Create a framework for engineering analysis that will enable you to test asset diagnoses for accuracy and build confidence in the solution. Evaluate what steps may be needed to incorporate this validation into your organization’s standard processes and workflow.
Define the criteria for a successful pilot. This may include the duration of the pilot, number of assets covered, number of correctly identified faults, thresholds for incorrectly diagnosed faults, thresholds for missed faults, etc.
4. Initiate a pilot phase—The timeline for a successful pilot will vary greatly. It is dependent on lead time to deploy condition-monitoring technologies, frequency of data collection, volume of instrumented assets, frequency of target failure modes, and many other factors. A clear project timeline with agreed-upon milestones is critical to maintaining support in the organization for running the pilot to its conclusion. Ensure open communication throughout the pilot so your team understands the ultimate goals and progress toward them.
5. Review results with stakeholders—Once you have data from your pilot program in hand and your team has some experience, you can make a case for further program expansion to leadership. With your team, review successes and opportunities for improvement and determine if there are ways to refine your setups.
6. Roll out a larger program—Return to your asset-criticality analysis to determine how to widen your condition monitoring program. You can grow your program within one facility or among multiple facilities, depending on the needs and priorities of your organization.
The future of asset management
It’s critical for organizations to understand the changing landscape of maintenance and reliability. Realizing the benefits of emerging tools and technologies, including machine learning, is how successful organizations will survive and thrive in an era of rapid changes. Laying the groundwork now will ensure you have the tools you need to integrate new technologies.
The ultimate goal of maintenance—to keep things up and running—has not changed. But these technologies will allow maintenance teams to work smarter, not harder.
