Predictive maintenance and big-data insight
In our earlier blog post, we discussed how machine learning overcomes the limitations of human expertise and more traditional analytics to enable machine intelligence. Today, we consider what that means for the evolution of predictive maintenance.
Reactive maintenance: corrective, preventive
In many companies, corrective maintenance is still current practice: You fix things once they break, but try to get to them before they do. If you add regular scheduling of maintenance tasks based on the history of your machines, you may be able to forestall some disruption otherwise caused by machine downtime. In the case of equipment and machines that are not essential to your business and its customers, this kind of corrective maintenance might even be good enough. One challenge that many reactive-maintenance companies grapple with is getting what they need, when they need it, from the spare parts supply chain.
At a more advanced stage, you make use of machine sensors to keep tabs on the wear and tear of certain machine parts or deploy them to record variances in output quality that might indicate that certain machine parts or components will soon fail. Condition-based maintenance in varying degrees of preventiveness is the norm in many manufacturing and assembly operations today.
A full view of all data that matters
True predictive maintenance only becomes possible with machine learning that incorporates wide-ranging analytics. You harvest and analyze the data streams from the IoT-connected sensors on your machines. You combine that data with historical records that show how your machines responded to changing environmental conditions as well as shifting workloads and production schedules.
When you enable machine learning to benefit from systems integration, it can bring together data findings from various sources to deliver meaningful insight that matters in the moment. You can add data insight from your other business systems, such as ERP, PLM, or Enterprise Asset Management (EAM). At that point, you can correlate machine performance and product development and design specifications. Or, include the operational and maintenance impact and the cost of purchasing raw materials and assembly parts from different suppliers. Or, see the outcomes that differently trained and certified machine operators and maintenance technicians achieve by working with and on your machinery.
Predictable and seemingly random failure patterns
You may well have seen one of the many articles that recycle, usually without giving the source, a claim that only 18 percent of industrial assets show a predictable failure pattern, whereas 82 percent display a random failure pattern. These numbers come from a chart created by NASA and the U.S. Navy whose source data were gathered several decades ago.
However, those percentages can illustrate a valid point: Preventive maintenance might be able to take care of the 18 percent of machine assets whose behavior follows a discernable rhythm. But you need machine learning and advanced analytics to understand and support machines that age, change, and break down unpredictably. Meaning that for most of us it’s too difficult to navigate masses of data to determine the contributing factors and events that cause performance to deteriorate and a machine to break down. We cannot understand what happens: It looks random to us. Machine learning does not experience the same feelings of fatigue and being overwhelmed when faced with vast data volumes from multiple sources as we do. It can make sense of the high volumes of time-series data to help companies improve their machines’ performance, quality, and uptime.
Algorithms enable machine intelligence
Machine learning depends greatly on algorithms whose calculation models determine how it responds to data findings and arrives at decisions. Most of the algorithms in use today fall into one of three types:
- Supervised learning: Algorithms make predictions based on examples. A set of independent variables predicts a target or outcome variable.
- Unsupervised learning: Algorithms need to organize the data or describe its structure. There is no target or outcome variable. Unsupervised learning is widely used to segment entities such as customers into groups that receive different treatment.
- Reinforcement learning: A machine is trained by means of trial and error to make the right, specific decisions based on experience. Reinforcement learning algorithms include the choice of an action in response to data, followed by a feedback signal that indicates the validity of the decision.
In our innius development and ongoing R&D, we make use of some proven off-the-shelf machine learning algorithms. We also rely on our findings in working with different companies and our conversations with our peers to optimize those algorithms for our specific purpose of predictive maintenance. What’s also critical in making predictive maintenance effective is that the solution gathers the right input for its algorithms from the machines, the ERP system, and other sources. Finally, companies need to be ready to adopt a more interactive, event-driven way of working instead of their more procedural approach when they adopt machine learning. As they gain the capability of reliable predictive maintenance, for instance, there will be an impact on how technicians are hired, trained, assigned, scheduled, and rewarded.
Looking beyond predictive maintenance
Machine learning to enable predictive maintenance is a great application, and many companies that run machines in business-critical processes will adopt it. However, machine learning can help companies accomplish more than predictive maintenance. Technologists are discussing and working on machine learning applications that could, for instance, produce spare parts by means of on-demand, 3D-printing or offer recommendations to optimize product configurations.
If you are interested in discussing predictive maintenance or other machine learning applications, or have questions and feedback, please get in touch.