Machine Learning for Predictive Maintenance

March 8, 2018
How embedded system-on-chip technology is combining realtime data acquisition, sensor fusion, data filtering and analysis, pattern detection and cloud communication in an integrated system to enable highly effective machine learning.

As a manufacturing executive, plant floor manager or operations engineer, your first thought when hearing the word system-on-chip (SoC) is to think this is a technology that is purely of interest to device developers. Historically, your need to understand this type of embedded technology would be very limited. But just as everything else about automation technology is changing in light of the Internet of Things (IIoT), so too is the need for you to better understand certain aspects of embedded, chip-level technology.

At the Embedded World 2018 event in Nuremberg, Germany, Christoph Fritsch, director of Industrial, Scientific & Medical at Xilinx, told me that he sees the future of embedded technology in automation evolving very quickly. “For device manufacturers, these changes allow them to leverage the technology to build their devices more efficiently, but it’s also opening up the embedded sandboxes for system integrators and end users to drive applications,” he said.

Explaining this recent change, Fritsch said that, in the past, Xilinx only worked with control device builders, such as Rockwell Automation, Siemens, Schneider Electric, etc. But since the advent of automation supplier IIoT platforms, like Siemens’ Mindsphere or GE’s Predix, “we’re project managing the development of systems with OEMs and end users to help them connect from the edge to the cloud,” he said.

In response to these developments, Fritsch pointed out that Xilinx has been working in depth with system integrators for about two years to show how SoC technology can be applied. One example he mentioned in our meeting is Xilinx’s work with system integrator Aingura IIoT for a project at CNC machine builder Etxe-Tar in Elgoibar, Spain. The aim of this partnership between Xilinx and Aingura IIoT is to support Etxe-Tar’s plan to implement machine learning on its CNC machines to enable predictive maintenance.

This project at Etxe-Tar, which sells its CNC machines to the automotive industry for use in building powertrain parts, has been detailed in the Industrial Internet Consortium’s white paper, “Making Factories Smarter with Machine Learning.” In the automotive industry that Etxe-Tar principally serves, operational failure of a CNC machines’ spindles can result in hundreds of thousands of dollars of damages. When spindle internal bearings fail, they effectively create a chain reaction that can destroy any linked device in close proximity. Such occurrences can shut down a production line for weeks, depending on the severity of the failure and spare parts availability. Accounting for all related aspects, including an idled workforce, the total cost impact of such a failure can easily reach millions of dollars per week.

The predictive maintenance system Aingura IIoT built to leverage machine learning and address the CNC failure issue is called Oberon. The Oberon system gathers data from the machines to which it is connected to deliver information about machine behavior.

A key component of the Oberon system is its intelligent gateway, designed and manufactured by System-on-Chip Engineering. The gateway uses the Xilinx Zynq SoC, which combines ARM processing and programmable logic fabric in a reconfigurable SoC device to perform realtime acquisition, sensor fusion (aggregation and connection of data from multiple sensors), data filtering and analysis, and pattern detection.

In terms of Oberon’s data acquisition using SoC technology, its main objective is to gather data coming from sensors. However, it also pre-processes the data to reduce the overall volume being transmitted. For example, vibration is sampled at least twice as often as vibration frequency. In this predictive maintenance application, a fast Fourier transform is performed and only the frequency of interest is stored.

For machine learning to be effective, it is critical to identify the relevant data variables so that noise and bandwidth use is reduced. This technique is called “feature subset selection.”

One example of how this can be applied in a CNC machine involves the machine’s servomotors, where variables like torque, power, temperature, vibration and angular speed can be measured. The number of trackable variables in a servomotor can be as high as 15,000. However, using feature subset selection, the reduction could result in identifying the need to track and store data from only 50 variables.

To illustrate how this works, the “Making Factories Smarter with Machine Learning” paper shows how the workflow with data reduction (Figure 1) would be used to build the CNC machine learning system. In this example, data is taken from the manufacturing system and sent to a machine learning algorithm that uses the new data and other information, such as mathematical models, to produce the predictive system. While data is traveling within this process, a summarization is performed, helping to only move data that is needed. This helps to reduce the bandwidth utilization and increase the response speed.

Based on historical data acquired during typical operation, machine learning algorithms use this and other real-time operational data to identify and learn system behavior patterns during the machining process. The data is analyzed in real time on the intelligent gateway and compared to typical operation data to identify anomalous operation and predict degradation—down to the component level—prior to any system failure.

In a CNC application, a machine learning-based monitoring system could detect the first signs of failure, providing enough time to stop the line in a controlled manner. Production and workforce teams could then be reassigned to reduce the failure’s impact on line productivity.

As an example, the paper highlights how the acceleration level related to the shaft is plotted against servomotor power (Figure 2). Here, the clustering technique distinguishes between idle, acceleration and deceleration and maximum power in terms of acceleration levels. The acceleration level is independent from the power consumption (power levels can be distinguished more clearly than in the data shown in Figure 3).

Based on this data analysis for the predictive maintenance application, it is expected that a servomotor should maintain this fingerprint acceleration level at all power consumption levels. Since the acceleration is related to the shaft angular speed, a malfunction of the servomotor would be detected whenever anomalies appeared outside the clusters; for example, anomalous vibration levels at a given acceleration state.

Companies in this Article

Sponsored Recommendations

MSD-SLC16G

CLICK industrial memory card, 16GB microSD. For use with all products with microSD memory card slot.

C0-12DRE-D

CLICK Ethernet Analog PLC, 24 VDC required, Ethernet and serial ports, Discrete Input: 4-point, DC, Analog Input: 2-channel, current/voltage, Discrete Output: 4-point, relay, ...

C2-FILL

CLICK PLUS option slot cover.

USB-CBL-AMICB6

AutomationDirect programming cable, USB A to microB-USB, 6ft cable length. For use with Productivity1000 and Productivity2000 CPUs and most USB devices.