Predictive Maintenance: How Will Machine Learning Impact Industrial Manufacturing
With all the technological advancements of the past couple of decades, it is easy to overlook just how reliable weather forecasting has become. I can remember my father laughing at the local weatherman's predictions on the five o'clock news because they were so often just flat wrong. Over time, however, weather forecasting has become so precise they can almost predict it down to the minute when a severe storm is about to arrive, allowing whole communities to be better prepared for the event. Were my father alive, I think he'd say it is pretty spiffy just how accurate weather forecasting is nowadays. What is driving this incredible progress? Machine learning.
Machine learning has become a massive part of our lives, from recommending clothes that suit our personal preferences, to the magic of autonomous vehicles for public transportation. . We can argue how useful some of these algorithms are and whether all of them make our lives better, but one thing we can say is that they are here to stay and will become more ubiquitous.
It may make sense to discuss the advantages of machine learning adoption in manufacturing operations. There are many places where machine learning can be beneficial. Simple examples are taking inputs from upcoming orders and staging the necessary materials before each process. Complex examples include analyzing thousands of data points from sensors built into systems to predict how and, even more importantly, when the breakdown of critical components might occur to identify them before they actually occur. This last example is where machine learning offers enormous advantages to the OEE, or overall equipment efficiency, having the potential to save manufacturers thousands, and potentially hundreds of thousands, of dollars in expenditure due to unplanned downtime.
At nLIGHT, we have recently launched our analytics platform, Cirrus, explicitly targeted at the crucial pain point of unplanned downtime. We utilize the hundreds of sensors designed into our fiber laser systems and run machine learning algorithms on this dataset to identify systems needing maintenance before reaching a failure state.
Today's maintenance model for fiber lasers is driven by the fact that fiber lasers are much more reliable than the systems they are replacing. CO2 lasers require multiple adjustments and tweaks to ensure daily operations and long-term performance. However, even though fiber lasers have an outstanding maintenance record, parts of the system will degrade over time and require maintenance to keep you up and running. Preventative maintenance is not ideal because parts are often replaced too early, costing more money over the long term. Waiting too long between PM intervals results in unplanned downtime. Unplanned downtime is costly not only because the tool is not producing parts the business depends on but also because the downstream processing steps become out of sync and require shifting of priorities for the production floor.
Shown below in figure 1 are three maintenance models.
Figure 1 - Maintenance models for industrial equipment
The best solution is predictive maintenance. In this model, the health of the laser is continuously monitored by comparing its performance against an entire population of similar lasers operating in the field. Early indicators of declining performance can be identified in advance of a fault or failure so a service event can be scheduled at a time that works best for the shop. This model is the goal of our analytics platform, Cirrus. The platform is designed to assess the performance condition of a laser by leveraging the powerful statistics of thousands of other fielded lasers, with operating hours on each of them that can also be in the several thousand. Contemporary machine learning algorithms are perfectly suited to operate on data sets like this for teasing out subtle, yet reliable indicators of component performance variances worthy of concern.
In this predictive maintenance model, maintenance downtime is mitigated, and maintenance can be performed when convenient for the shop or factory, increasing OEE. In the reactive model, customers may have to wait days for repair teams to arrive and then analyze what went wrong. With the predictive model, that repair time can be cut down to a few hours, a truly revolutionary change to the maintenance of your fiber laser.
Predictive maintenance is only the beginning of machine learning for industrial manufacturing. Cirrus provides a digital platform for all types of data analysis, accommodating sensor data streams from fiber laser cutting, welding, and additive manufacturing machine tools. This additional data can provide a view into operating efficiencies of the shop floor or factory and give updated recipes from existing libraries for different material combinations. Additionally, this technology opens new business models, allowing increased capabilities or decreasing as needed. Machine learning offers all of these opportunities and more.
It is essential to remember that improvement happens over time with all new technologies, especially with machine learning. Weather forecasting did not get better overnight. It took many iterations and
improvements, as with any "learning" system. However, data analytics and machine learning are poised to make revolutionary improvements to the industrial markets.
