Introduction

Drilling seems simple. You twist a pair of cutting edges around a shaft, attach it to something that spins really fast and violá, you can put holes in just about anything. This seemingly simple process has been around for thousands of years, and still accounts for some 40% of manufacturing efforts today. Lots of holes are always needed in lots of things. They usually have definite quality requirements, but maintaining that quality is sometimes difficult because the drill bits themselves wear and deform in minute increments with every hole drilled. At some point, the drill passes from “usable” to “worn,” and ceases to produce satisfactory holes.

In a large-scale manufacturing setting, there may be hundreds or thousands of drill machines operating at once. When an individual bit wears beyond tolerance thresholds, or even breaks entirely, often no one will realize it until somewhere down the assembly parts start appearing with poor quality or non-existent holes. The line must be stopped, the broken or worn drill found and replaced, and the ruined products scrapped. Needless to say, this is expensive.
The only solution right now is to replace drills at conservatively estimated preset intervals. This isn’t a very good solution for several reasons. Drilling is a complicated process, and many factors determine the life and wear of a drill. Especially with non-homogeneous materials, like cast iron, the variances in the material randomly affect the life of a drill. In order to avoid as many drill failures as possible, the drills usually get replaced long before the end of their useful lives. Even with these expensive precautions, there is no way to predict, avoid, or even detect a sudden drill failure.

What we’re trying to do is find a reliable way to monitor drill wear automatically, indirectly, and in-process. By “indirectly” we mean that we are looking for some outside indication of drill wear. Direct wear measuring involves taking the bit out and directly (optically) measuring the wear with a microscope. We do this to check our experimental results, but the end goal is to avoid the time-consuming hassle entirely by finding an automated method of detecting drill wear.
We’re certainly not the first ones to study drill wear monitoring, but we are trying a new method of doing it. Dozens of studies have focused on measuring torque, power consumption, and other indicators of how hard the drill machine is working. The premise is that a worn drill must work harder than a sharp drill to achieve the same material-removing result. None of these studies has yet yielded a sufficiently robust or accurate wear detection method.
In contrast, we are measuring the vibrations of the drill-work piece system. Few studies have dealt with vibration even in part, let alone entirely. Our goal is to find vibration signatures characteristic of sharp and worn drills, respectively. If we are successful, we could then implement and refine a system that can quickly and automatically monitor these vibrations and send out an alert when worn- or broken-drill vibrations are detected.

Feeling the vibes

Vibrations can be categorized by frequency range. Simple “vibration” refers to frequencies of 1 – 10 Hz. “Sound” is defined by the limits of human hearing, 20 Hz – 20 kHz. “Ultrasonic” vibrations range from 20 kHz – 80kHz, and “acoustic emission” generally encompasses everything from 80 kHz – 1 MHz. One additional distinction is that sound vibration is airborne, while the other types are mechanical. Each category has its own types of sensors and its own advantages and disadvantages. Sound, for example, is useful to study both because previous studies have found signature vibrations in this range, and because it is the natural signal that most machine-operators will rely on. Furthermore, microphones are relatively cheap and easy to get. At the same time, because sound is airborne it is susceptible to background noise and other random distortions.

Fast and Fourierous, reclusive Russians

The difficulty in studying vibration is in extracting useful information from the mess of data that monitoring yields. Drilling tends to cause a lot of outlying amplitudes and otherwise uninformative vibrations. It takes some work to find patterns amidst the jumble, but there are ways:
The Fast Fourier Transform (FFT) is often a reliable means to find out the frequency content of a measured signal. FFT is a mathematical tool based on the theory that any vibration can be decomposed into an infinite series of sine and cosine waves. The FFT method analyzes signals by relating frequency and time, and is more reliable than the relation between wavelength and time. FFT has been used regularly in tool condition monitoring, but because of the theory behind it is only useful for stationary objects. This is fine for a lathe, where the tool doesn’t move, but has already proven unhelpful in our experiments.
Higher order spectral analysis and cepstrum analysis are functions based on FFT. Higher order spectral analysis examines the cross spectrum, cross relation, and frequency response as well as some multi-signal frequencies. The cepstrum analysis identifies harmonic bands or side bands and is calculated by taking the inverse of FFT.

The wavelet transform method is extremely useful in analyzing the time-domain at different frequencies. Both continuous and discrete wavelet transforms are used for tool breakage detection using spindle feed and current signals. The test signals can also be shown in time domain and wavelet transform.
The advantage of Artificial Neural Networks (ANNs) lies in their resilience against distortions in the input data and their capability of learning. They are often good at solving problems that are too complex for conventional technologies (e.g., problems that do not have an algorithmic solution or for which an algorithmic solution is too complex to be found).
The most promising method, and the one we are currently using, involves hidden Markov models (HMMs). These powerful analysis models are frequently used in voice recognition technology. Thankfully, they are easy to incorporate into our experimental setup, using a MatLab module off of the main LabVIEW program we use.

There’s probably no single best analysis method. In his paper A summary of methods applied to tool condition monitoring in drilling Erkki Jantunen points out that “the most effective and reliable methods for tool wear monitoring are so slow in practice” that they are not practical for detecting sudden failures. If he is referring to processing time, his assessment is probably still relevant, as the paper was published only half a year ago.

Conclusions

We’re probably on the best track in terms of analysis. HMMs seem to be powerful and applicable to our research. There are a lot of papers about HMMs, and it could be very helpful to learn more about what other people, in any field, are doing to augment their effectiveness or reliability.
We recently added a microphone to our setup, and this is something I would have suggested. In 1995, a team of researchers did essentially what we are doing, but with smaller diameter drills (3mm and 6mm). They identified amplitude peaks in several frequency ranges corresponding to specific types of drill wear. All of these were well within the sound region, ranging from 2.4 to 5.8 kHz. While we could measure those frequencies mechanically, it may be more convenient for manufacturers if we develop a microphone based system. We currently have plans to install an AE sensor, and this may provide helpful data. I do wonder, however, how applicable AE would be to manufacturing situations, since the sensor seems to have rather significant placement restrictions.

Finally, it may be beneficial to develop a fast and simple sensor system for the sole purpose of detecting sudden breakage. This may be more relevant to smaller drill sizes, but then we need to be relevant to as many drill sizes as possible.