Making Factories Smarter Through Machine Learning
Figure 4: Reducing the Big Data to Critical Metadata to Get to the Smart Data
One example of an application is when the tool tip is monitored for process deviations. Note, there are many types of process deviations possible depending on the nature of the product line operation and the required cycle time – there are many types of tools: A typical CNC machine has multiple tools with many different types of tool tips( laser, milling, drilling, grinding, tapping, etc.).
Temperature fluctuations cause most common deviations. In these cases, variables like position and temperature in different points has to be measured. If 14 measuring points are selected( mainly different temperatures and tool tip position), using a sample rate of 5 minutes can give enough data to produce 11.2 KB / day( 4 MB / year). This data with a supervised learning algorithm can increase the machine availability by 20 %, which is nearly 3 hours / day.
Another example is the operation behavior of servomotors. In this case, variables like torque, power, temperature, vibration and angular speed can be measured. The number of variables could be up to 15,000. However, using feature subset selection, the reduction could leave only 50 variables and 1 TB / year. Figure 5 shows the results of using Unsupervised Learning to identify patterns relative to Power Consumption, Torque and other variables generating a“ fingerprint” of the servo drive, where servomotor behavior is represented in the X-axis.
- 36- January 2017