Applying Machine Learning for Milling to Prevent Chatter

Machining technology continues to advance in response to industry needs. Despite continuing increases in productivity, however, challenges remain. For example, while the digital transformation of part design and path planning for CNC machining is widespread, CAM software generally treats machining as a geometric activity. Provided the cylindrical tool follows the commanded path through the stock model, it is assumed that the machining process is acceptable and the desired geometry is obtained. This approach does not consider the inherent constraints imposed by machining dynamics.

It is known that some spindle speed-axial depth of cut combinations exhibit chatter, which produces large forces, large vibrations and poor surface finishes, while others do not. Machining dynamics models are therefore implemented to select spindle speed-axial depth combinations that avoid chatter, while meeting design tolerances. Milling stability maps, which separate stable spindle speed-axial depth combinations from those that produce chatter, can be produced using a measurement of the tool tip vibration response and a cutting force model. See Fig. 1.

Stability maps demonstrate large stable zones at high spindle speeds. When tool wear constrains the maximum cutting speed, however, lower spindle speeds must be selected. At low spindle speeds, process damping can increase the stability limit. Because the inputs to the stability map are not perfectly known, the predicted stability boundary also has uncertainty.

Applying machine learning enables the uncertainty to be reduced as more data is collected. One machine learning algorithm is k-nearest neighbors, or KNN. It is applied here to model milling stability with process damping effects. A KNN classifier uses proximity to predict the behavior at selected data point based on the surrounding points. The assumption is that similar points are found near one another. For the milling stability classification problem, a label (stable or unstable) is assigned based on a majority vote. In other words, the label that is most frequently represented around a given spindle speed-axial depth combination is used to predict the behavior at that point.

The Fig. 1 stability map was produced using the vibration response for a selected tool-toolholder-spindle combination and 1018 steel workpiece. It was used to generate a training dataset for the KNN classifier, where each point in a grid was labeled as stable or unstable depending on whether it was above or below the stability boundary. See Fig. 2.

The process damping behavior was next established using a “stair step” approach with a minimum number of cutting tests. A low spindle speed of 800 rpm was selected for the first test. A conservative axial depth of 1 mm was chosen given the predicted stability map (with no process damping) shown in Fig. 1. Based on the test result, a progression was followed:

• If the cut was stable, the spindle speed was maintained and the axial depth was increased by 0.5 mm (50% of the original axial depth). This was repeated until an unstable result was achieved.
• When an unstable cut was obtained, the spindle speed was reduced by 200 rpm (25% of the original spindle speed) and the previous unstable cutting depth was selected.
• The sequence was repeated.

The test results are shown in Fig. 3. Next, knowledge about process damping was included with the test results to update the grid of points in Fig. 2. For a stable cut at a selected spindle speed-axial depth, all points with an equal or smaller axial depth and equal or lower spindle speed were also labeled as stable. This provided many updating points from a single test and contributed to the “stair step” updating.

The re-labeled low spindle speed points from Fig. 2 are shown in Fig. 4. The KNN classifier with process damping effects is displayed in Fig. 5. It is seen that the stability boundary is now increased at low spindle speeds. This demonstrates the value in combining machine learning and limited test data with physics-based milling models to increase prediction accuracy.