Machine Learning for Machine Tools

With computers beating both chess grand masters and world-class Go players, plus ChatGPT writing everything from college term papers to love poems, artificial intelligence (AI) is arguably the hottest topic in tech today.

But what about in manufacturing? Theoretically, with enough data and the right algorithm, AI and machine learning could assess machine condition, adjust programs to improve throughput or quality, and perhaps more. So, what’s the reality?

Design Magic

Let’s start with generative design. Computer algorithms can automatically design a part to meet a diverse set of constraints and requirements.

“Your engineering intent defines where the part needs to sustain specific load cases. It could be forces, pressures, temperatures,” explains Kirill Volchek, chief technology officer for Oqton Inc., a software company owned by 3D Systems, Rock Hill, S.C. “The part could be working under vibration. You name it. You also define an imaginary volume to be populated with this part. Our algorithm then suggests designs that function best under the load cases.”

Another beneficial feature of this tool is the ability to quickly explore optional approaches, Volchek says, such as “minimizing the amount of material required, or looking into the potential areas of the part that could be loaded more, or loaded less.”

Volchek describes this as an iterative, goal-seeking process, but one that remains based on physics. Conversely, he explains, machine learning is coming to the fore now as computers make design decisions that go beyond the physics.

For example, Volchek says, computers can be trained to consider the requirements of specific manufacturing domains. “Volume generation for additive manufacturing should look different from a casting or CNC machining,” he adds. This broader “intelligence” suggests even more exciting possibilities.

The Unmanned Factory?

AI is beginning to reduce the need for human labor in manufacturing much like self-driving cars automate personal travel. “Let’s imagine that you have a manufacturing bureau that needs to produce thousands of unique parts each day,” Volchek notes. “You cannot afford to have an engineer look at and be involved in the manufacturing of each individual part,” he says, noting that it’s a challenge that some customers meet on a daily basis.

Oqton was established five years ago and used “a huge amount of data” to train algorithms that can turn certain manufacturing areas “into completely self-driving, or unmanned, processing,” according to Volchek. This includes the 3D printing of dental implants, which, as Volchek puts it, are each unique, though similar.

Oqton’s AI is able to examine the geometry of an incoming order, classify the part, and determine how to handle each feature. “You build a flat surface differently from an edge,” Volchek says. “Some areas that go into a human body must not be post-processed. And so on.”

The AI automatically designs any sacrificial support structures needed to build each part, as well as orients each part on the build tray, nesting parts together to maximize throughput.

“It’s brilliant. Our designers can drag parts into a folder, and it just prepares and builds the jobs for us,” asserts John Dumas, owner and orthodontist at Motor City Lab Works in Birmingham, Mich. “The parts are built, assembled and the support structures are there, ready to go,” he adds.

Not only does the AI have close to 100% success rate with the builds, Volchek says it can interface with a company’s ERP system to intelligently schedule jobs and route them to the proper machine.

While dental implants are one of the best examples of fielded automation, Oqton also has a cooperation agreement with Houston-based Baker Hughes Co. to automate the 3D printing of oil and gas parts. But, as Volchek points out, 3D printing accounts for only about 0.1% of all manufacturing. So, what about the other 99.9%?

Heidenhain’s Integrated Process Monitoring learns the speeds, feeds and torques experienced by each axis throughout a cut in producing a good part, and then tracks and displays any deviations for analysis and adjustments thereafter. (Provided by Heidenhain)

Process Optimization in Subtractive Machining

Oqton was able to pioneer new AI tools because they were already “part of the MES (manufacturing execution software) space,” with agreements that provided access to anonymously learn from vast amounts of geometric data, according to Volchek. Given the digital nature of 3D printing, that data was golden.

“Additive manufacturing is not just producing the outer shape of a part,” Volchek says. “It’s producing the internal microstructure. It produces a material’s metallurgical properties, if you’re talking about metal parts.”

That’s not the case with most manufacturing. Volchek estimates that about 90% of the parts being produced now don’t have corresponding 3D drawings. Furthermore, the majority of 2D drawings are still in raster format, essentially digital blueprints that can’t be scaled and which are devoid of searchable process information. In order for AI to “break boundaries” in subtractive machining and other traditional disciplines, Volchek says developers need a large number of annotated 3D models. Most manufacturers don’t have them, and generally don’t want to share the information they do have, he maintains.

Gisbert Ledvon, vice president of marketing at Heidenhain Corp., Schaumburg, Illinois, echoes this sentiment. In contrast to other CNC manufacturers that use cloud-based machine monitoring and data-sharing services, Ledvon says, “We still believe that it’s better for our end users to maintain their data within their own network.”

Heidenhain’s architecture facilitates secure communication within and between plants. But the company doesn’t want to host the data, as it’s vulnerable to theft or loss.

There’s still a lot that can be done with a company’s own data. But full-blown AI-controlled production probably isn’t feasible without much greater data sharing.

Heidenhain, however, offers an intriguing process optimization tool that uses AI on a job-by-job basis. Called Integrated Process Monitoring (IPC), the option was introduced at IMTS 2022 for machines with the TNC7 control. IPC learns the speeds, feeds, and torques experienced by each axis throughout a cut, then tracks and displays any deviations for analysis and adjustments.

First, the operator activates the option and cuts a model part. Assuming he’s satisfied with the outcome (and, of course, can inspect it fully on a CMM to be sure), he then establishes tolerances for each machine motion. For example, the speeds and feeds for X, Y, and Z motion might remain within ±5% of the ideal, while the spindle torque cannot vary more than 20%.

IPC tracks every machine motion on all subsequent parts and graphically displays the resulting part in 3D, coloring each surface to highlight any changes. If every surface is colored green, you can be sure you have another good part, without having to measure it with a probe or camera (assuming a reasonable tolerance is set). Any surface displayed in yellow indicates that something has changed and should be investigated. Red indicates a major deviation, indicating the part likely will not be acceptable.

“We take the intelligence of all the Heidenhain devices making up the machine motion to determine what went wrong and where there are problems on the final part,” Ledvon summarizes.

The entire process is graphed in real time and recorded, so users can see how well each machine motion remains in the tolerance band. The operator can adjust each band if it’s too restrictive or too loose. And, if the color coding indicates the need for a quality check, Ledvon says, operators can focus on the red and yellow areas. This can cut inspection time by 80%.

A threshold for the machine’s motions also can be set. For example, tool breakage would ordinarily only be detected with a laser or another measuring device. But with IPC, “If your spindle load went to zero when your program expected the tool to be deep in a cut, the machine would know your process is 100% out of tolerance and automatically shut down,” Ledvon says. “The operator would then see what had happened and bring in another tool and continue.”

Boosting OEE by Predicting Problems

One area being indisputably revolutionized by machine learning across the manufacturing spectrum is predictive maintenance. And improvements in this area typically lead directly to improved productivity.

CpK Interior Products Inc., a manufacturer of automotive interiors, offers an example. Starting with a proof-of-concept project without any preconceived notions about what it would learn or accomplish, Gregory Farrar, head of r&d and plant manager for CpK’s Belleville, Ont., plant, says the company provided a year’s worth of machine data from its compounder to an AI software vendor.

“We provided 20 different pieces of data. Analog tags, such as motor amperage, motor speed, water temperature, and water flow,” elaborates Chris Murray, production manager.

These are values the machine control captured every few milliseconds. Although the vendor knew nothing of CpK’s history, its algorithm was able to detect several data irregularities that CpK was able to tie to real world problems.

“One was an amperage anomaly in our cooling mixer, which was a precursor to a raw material issue we had with our base resin,” Murray recalls. “That was a quality indicator. The second was a temperature anomaly in the water supply going into our cooling mixer. This indicated a heat exchanger failing about three months in advance.”

The point here is not just that the algorithm found important problems in advance, but that it found them in a mass of data that no human could effectively process. As Murray puts it, “There are mechanisms to identify those types of anomalies, but not in a timely manner. Not fast enough to intervene and prevent an issue from happening. That’s a lot harder to do when you’re relying on somebody to put two and two together and raise their hand and say, ‘Hey, we potentially have a problem here.’

“We took the human element out of that aspect of the process,” he continues. “When using AI to detect anomalies, you still rely on a human being to take that information and make a decision. But calculating and determining abnormalities is actually the difficult part.”

CpK’s proof of concept resulted in a 10% improvement in OEE from the start. The company now is working with another vendor, Magic Systems, whose machine learning software promises to combine multiple discrete data inputs and determine the biggest contributor to any anomaly, as well as quantifying three or four secondary contributors. That’s virtually impossible for a human to do, according to Murray. “We can’t compute all that data quickly enough to understand the relationships between them and how they’re contributing to the problem.”

CpK also uses Magic’s system to visualize its downtime, coding the root causes and compiling the data to identify where they are losing the most time.

“By using that data, we can focus our continual improvement projects,” Murray explains. “A big part of our downtime here is setups, changeovers and shutdowns. We were able to reduce our changeover time by something like 50%.”

FANUC America Corp., Rochester Hills, Mich., has an interesting approach that uses just two data points from a machine’s motors to identify mechanical problems not related to the motors themselves. Called AI Servo Monitor, the system tracks the speed and torque command of each motor at a rate of one millisecond.

“After it has a sufficient amount of data to determine what is normal, it establishes what we call an ‘anomaly score’ for each axis,” explains Robert Taylor, program manager for machine tool data collection and connectivity. “Each day it collects more data and provides another set of numeric scores and it graphs them in comparison to the first set, the model.”

The graph makes it easy to see any trend that indicates things are changing in the machine. But Al Servo Monitor doesn’t just record and graph motor data. Taylor says the “AI” is in the application of a fast Fourier transform algorithm that analyzes data and makes predictions about possible mechanical failures. It can’t predict exactly what’s wrong, but it can point to problems in a specific axis that could be caused by a worn rail or ballscrew, chip build-up, poor lube and the like, providing operators a targeted nudge to investigate in time to prevent a failure.

Taking just a few days to establish a baseline, AI Servo Monitor works best in a production environment. It takes longer to define “normal” in a job shop that’s constantly changing parts and machining parameters. Likewise, it helps if the machine is in good condition. The system is limited to machines equipped with the FANUC control, drives and motors. But it’s available for a reasonable one-time fee, according to Taylor, whereas most AI solutions require costly monthly subscriptions.

Not AI, but Smart Investments

AI Servo Monitor uses motor data and AI to predict mechanical problems. FANUC and Heidenhain (and other control manufacturers) also offer first-rate tools for monitoring the health of the motors themselves, as well as the other motion control elements, even if these tools don’t use AI. FANUC monitors motor health through temperature and the insulation resistance to ground via its MT-LINKi data collection software, Taylor says.

“It’s PC software that resides on an on-premise server that is connected via a network to the equipment and is collecting data from the equipment,” he explains. “The necessary connection to it is standard on all FANUC controls.

The software will also capture and display data from non-FANUC equipped machines. And, while this doesn’t analyze the data, it can display machine status for a whole factory, making it easier to manage.

“What MT-LINKi can do for us is, if it shows (operators) are getting several alarms on this certain shift, it allows us to go down and see what opportunities there are for additional training,” says Landon Garrison, manufacturing engineer at Stober Drives Inc., Maysville, KY.

The system helped Stober increase utilization and productivity, adds Industrial Engineer Nathan Landreth, who describes MT-LINKi as a “game changer.”

Similarly, Heidenhain’s State Monitor software runs offline and networks throughout the shop using UPC UA, MT Connect or Heidenhain’s own DNC protocol. In addition to the inherent feedback mechanisms within motors, linear scales and encoders, “the machine tool builder can set up to seven different sensors,” according to Ledvon. “You could have a sensor for coolant or torque, for example.”

Then again, Taylor cautions, “When you start collecting huge amounts of data, it makes it very hard to go through and analyze. Some customers might not know exactly what to look for. So they say, ‘I want everything.’ That just muddies the waters. You really need to be focused on what you’re looking for. And that’s where FANUC excels, because they know what to look for.”

It’s a sentiment shared by Ledvon and Heidenhain. Look at what’s available from the machine tool and control builders, and pick your AI battles carefully.

Written by:  Ed Sinkora, Contributing Editor, SME Media, for SME.