Author Description

Because of its great toughness and light weight, carbon fiber reinforced polymer (CFRP) is widely utilized in the aircraft sector. The anisotropic nature of CFRP contributes to its longevity, but it also has significant drawbacks, such as the generation of hole delamination from the drilling process. Hole delamination will not only cause the hole wall to become rougher, but it will also diminish the hole's quality, making it less resistant to harm. As a result, hole quality monitoring is quite crucial. In this paper, we measure hole quality using the average surface roughness, $R_a$. Roughness (and other quality indicators) are heavily influenced by the condition of the cutting tool, which is commonly measured by Mechanical Engineers using total wear area (TWL), and flank wear land (FWL). In our data set, the drilling process is performed using one drill bit which becomes progressively more worn over time. Thus TWA and FWL can be estimated knowing the number of holes that have been previously drilled (i.e., the hole number). We develop a machine learning pipeline to predict how many holes have been previously drilled using a given tool (i.e., the hole number), as well as a pipeline to predict surface roughness, $R_a$. The pipeline uses force and torque measurements captured directly during active drilling.The contributions of this paper are twofold. First, we provide efficient methods for segmenting the force and torque signals to find the regions of active drilling. Naive Windowing performs well when compared to manual segmentation and is $16$ times faster than the Gaussian Ramer-Douglas-Peucker algorithm. Second, we provide performance results for a wide array of machine learning classifiers on the hole number and $R_a$ classification tasks. We find a feed-forward neural network with SeLU activation performs very well to predict hole number, and that modest performance can be obtained predicting $R_a$ with a support vector machine. These contributions may help the