Currently there are a wide variety of manual and automatic processes for checking tool wear. These systems are constantly stepping up the game with incremental improvements, but they are about to get warp speed.
The kind of incremental improvements we see are undoubtedly useful. You can buy indexable cutting tool inserts that are coated to make erosion more obvious. New methods employing optical, radioactive, or electrical sensors have been tried and implemented on CNC machines. These have improved accuracy while reducing downtime.
But how many tools still break mid-job? How many hours does tool monitoring cost your shop despite investments, planning, and regular changes? There is still plenty of room for development… at least until next gen cutting tool monitoring technologies roll out.
That’s because a bunch of experimentation and development is finally coming to a “head.” Literally, machining is about to get an Artificial Brain.
Let’s talk about the most recent tools that will bring highly accurate automated tool monitoring to your shop in the early ‘20s if not next year.
Cutting Tool Monitoring Data Types Are Overwhelming, the Last Decade Has Shown
There are already quite a few methods of measuring tool wear, but each has its own limitations. You might see sensors that analyze the actual geometric parameters of the cutting tool, or after-the-fact measurement of surface roughness, or tool tip heat, but these cannot be reliably measured during cutting. Many are difficult or impossible to apply — at least in a way that produces significant value for machine shops.
Lately, more indirect methods have implemented modern sensors and sophisticated analytics technologies to process external signals, using data such as cutting force, vibration characteristics, acoustic emission, temperature, and surface characteristics. These are easier to apply on real shop floors, not just artificial environments, which makes indirect measurements more promising by far.
Over the last couple of decades, the number of data types has grown. It’s still growing actually, as the following results show.
Patterns in Chip Color, Acoustics and Spark Proliferation Can Predict Tool Wear
One exciting paper published this year looked into two uncommonly measured factors: chip color and spark proliferation. The signs of wear would be too subtle for humans to reliably record or use, but the differences are no problem for sensors to measure. Tests did find a significant relationship between these factors.
These kind of experiments have forwarded the kind of data we can consider when developing a cutting tool monitoring system, but they also complicate the field. With so many options for measurement and analysis, we haven’t developed the right measurement and analysis. This is the perfect problem-type for Machine Learning because we are using a large number of variables and inputs to predict a measurable outcome.
Neural Networks Sort through the Data to Deliver 92.59% Accuracy
But indirect measurements haven’t led to consistently accurate predictions on the shop floor. Experiments like the one above provide useful insights, but they are typically constrained by fairly rigid conditions. Experimenters will use a handful of projects to gauge results, so their data will only be applicable to a very small set of real-world scenarios.
Meanwhile each experiment adds to the grand equation of tool wear: 2,000 data points when all you need is one good predictor of tool wear. The equations, datasets, and models have piled up, which makes cutting tool monitoring is finally ready for AI.
Published in this year’s Complexity; Hoboken, an experiment with artificial neural networks and an assortment of sensors broke new ground in cutting tool monitoring predictiveness.
What is an Artificial Neural Network (ANN)?
Artificial Neural Network (ANN) simulates the human brain to process information in the service of a clearly defined task. It’s totally different than other algorithms because traditionally, algorithms are developed by humans. ANN develops its own algorithms based on learning samples. The more samples you expose the program to, the better it gets. ANN learns how to reach your target objective in “its own way.”
Google DeepMind is one of the best known examples of ANN. If you haven’t seen DeepMind in action, check out this Starcraft II game between AI and a pro gamer to see what it can do.
Researchers have attempted to use ANN for cutting tool monitoring in the past, but there have been a few obstacles:
- Adaptability – existing monitoring models can be accurate, but when they are, they are narrowly defined. They might work on one specific kind of project but not very many others.
- Cost of sample data – ANN thrives in large data sets where it can adjust to different conditions and develop a nuanced model that plots inputs to the right output. But acquiring large sample sizes of various project types is challenging and costly.
- Finding the right sensor set-up – Even with ANN doing the heavy analytical lifting, deciding what type of data to feed it is itself a complicated undertaking.
The latest and greatest model was developed by Maohua Du and a team of researchers. Their ANN was tested across a wide variety of environments and posted a 92.59% accuracy in predicting tool wear. That isn’t just in one set-up either. The system works across different machining operations such as milling, drilling, boring, forming, and shaping, each combined with different cutting tools and workpiece materials.
To be sure more development will occur between now and when CNC machines will boast ANN-driven cutting tool monitoring systems, but with working models already in existence, you can bet on seeing them soon.