The fully autonomous factory has always been a provocative vision, much used in speculative fiction. It’s a place that’s nearly unmanned and run entirely by artificial intelligence (AI) systems directing robotic production lines. But this is unlikely to be the way AI will be employed in manufacturing within the practical planning horizon.
The realistic conception of AI in manufacturing looks more like a collection of applications for compact, discrete systems that manage specific manufacturing processes. They will operate more or less autonomously and respond to external events in increasingly intelligent and even humanlike ways—events ranging from a tool wearing out, a system outage, or a fire or natural disaster.
AI in manufacturing is the intelligence of machines to perform humanlike tasks—responding to events internally and externally, even anticipating events—autonomously. The machines can detect a tool wearing out or something unexpected—maybe even something expected to happen—and they can react and work around the problem.
Historians track human progress from the Stone Age through the Bronze Age, Iron Age, and so on, gauging evolutionary development based on human mastery of the natural environment, materials, tools, and technologies. Humankind is currently in the Information Age, also known as the Silicon Age. In this electronics-based era, humans are collectively enhanced by computers, leverage unprecedented power over the natural world, and have a synergistic capacity to accomplish things inconceivable a few generations ago.
As computer technology progresses to be more capable of doing things humans have traditionally done for themselves, AI has been a natural development. People have choices about how machine learning and AI are applied. One thing AI does well is helping creative people do more. It doesn’t necessarily replace people; the ideal applications help people do what they’re uniquely good at—in manufacturing, that could be making a component in the factory or designing a product or part.
Increasingly, it’s about the collaboration of humans and robots. Despite the pervasive popular impression of industrial robots as autonomous and “smart,” most of them require a great deal of supervision. But they are getting smarter through AI innovation, which is making collaboration between humans and robots safer and more efficient.
Today, most of the AI in the manufacturing industry is used for measurement, nondestructive testing (NDT), and other processes. AI is assisting in the design of products, but fabrication is still in the early stages of AI adoption. Machine tools remain relatively dumb. Automated shop tooling is in the news, but many of the world’s factories continue to rely on older equipment, often with only a mechanical or limited digital interface.
Newer fabrication systems have screens—human-computer interfaces and electronic sensors to provide feedback on raw material supply, system status, power consumption, and many other factors. People can visualize what they’re doing, either on a computer screen or on the machine. The way forward is becoming clear, as is the range of scenarios for how AI is used in manufacturing.
The nearer-term scenarios include monitoring the machining process in real time and monitoring status inputs like tool wear. Such applications fall under the heading of “predictive maintenance.” It’s an obvious opportunity for AI: Algorithms that consume continuous streams of data from sensors find meaningful patterns and apply analytics to predict problems and alert maintenance teams to resolve them before they happen. Sensors inside the machine can monitor that something’s happening. It could be an acoustic sensor listening for the belts or gears starting to wear out, or it could be a sensor monitoring the wear of the tool. That information would be linked to an analytic model that could predict how much life is left in that tool.
On the shop floor, additive manufacturing is becoming an important modality and has prompted adding many new types of sensors to the system, monitoring new conditions affecting materials and fabrication technology only widely adopted in the past 10 years.
AI is making possible much more precise manufacturing process design, as well as problem diagnosis and resolution when defects crop up in the fabrication process, by using a digital twin. A digital twin is an exact virtual replica of the physical part, the machine tool, or the part being made. It’s much more than a CAD model. It’s an exact digital representation of the part and how it will behave if, for example, a defect occurs. (All parts have defects; that’s why they fail.) AI is necessary for the application of a digital twin in manufacturing process design and maintenance.
Large enterprises have a lot to gain from AI adoption, as well as the financial strength to fund these innovations. But some of the most imaginative applications have been funded by small- to medium-size enterprises (SMEs), such as contract designers or manufacturers supplying technology-intensive industries like aerospace.
Many SMEs are trying to leapfrog larger competitors by rapidly adopting new machinery or new technology. Offering these services is differentiating in the fabrication space, but in some cases, they are implementing new tools and processes without the necessary knowledge or experience. This could be true from a design point of view or a manufacturing point of view; it’s challenging to break into additive manufacturing because of this. In this scenario, SMEs could have greater incentives for AI adoption than large enterprises: Using smart systems that can provide feedback and assist setup and operationalizing could help a small upstart establish a disruptive foothold in the market.
Essentially, end-to-end engineering expertise can be built into a manufacturing process. That is, the tooling with onboard AI can be delivered with the knowledge to direct its installation, adoption, sensors, and analytics for detecting operational and maintenance issues. (Those analytics are likely to include so-called “unsupervised models,” trained to look for patterns of feedback from the sensors not associated with known problems by looking for odd or “wrong” aspects to be investigated.)
A real-world example of this concept was DRAMA (Digital Reconfigurable Additive Manufacturing facilities for Aerospace), a £14.3 million ($19.4 million) collaborative research project started in November 2017. Autodesk is among a consortium of companies working with the Manufacturing Technology Centre (MTC) to prototype a “digital learning factory.” The entire additive-manufacturing process chain is being digitally twinned; the facility will be reconfigurable to meet the requirements of different users and to allow testing of different hardware and software options. Developers are building an additive manufacturing “knowledge base” to aid in technology and process adoption.
In DRAMA, Autodesk plays a key role in design, simulation, and optimization, fully taking into account the downstream processes that occur in manufacturing. Understanding the effect of the manufacturing process on each part is critical information that humans can automate and then bring into the design process through generative design to allow the digital design to perform closer to the physical part.
This scenario suggests an opportunity to effectively package an end-to-end work process to sell to a manufacturer. It could include everything from software to the physical machinery in the factory, the digital twin of the machinery, the ordering system that exchanges data with the factory’s supply-chain systems, and the analytics to monitor manufacturing methods and collect data as inputs move through the system. Essentially, creating “factory in a box” systems.
Such a system would allow a manufacturer to look at the part that made today, compare it to the part made yesterday, see that product quality assurance is being done, and analyze the NDT that’s been done for each process on the line. The feedback would help the manufacturer understand exactly what parameters were used to make those parts and then, from the sensor data, see where there are defects.
The utopian vision of that process would be loading materials in at one end and getting parts out the other. People would be needed only to maintain the systems where much of the work could be done by robots eventually. But in the current conception, people still design and make decisions, oversee manufacturing, and work in a number of line functions. The system helps them understand the actual impacts of their decisions.
Much of the power of AI comes from the ability of machine learning, neural networks, deep learning, and other self-organizing systems to learn from their own experience, without human intervention. These systems can rapidly discover significant patterns in volumes of data that would be beyond the capacity of human analysts. In manufacturing today, though, human experts are still largely directing AI application development, encoding their expertise from previous systems they’ve engineered. Human experts bring their ideas of what has happened, what has gone wrong, what has gone well.
Eventually, autonomous AI will build on this body of expert knowledge so a new employee in, say, additive manufacturing benefits from operational feedback as the AI analyzes onboard sensor data for preventive maintenance and to refine the process. That’s an intermediate step toward innovations like self-correcting machines—as tools wear out, the system adapts itself to maintain performance while recommending replacement of the worn components.
AI applications aren’t limited to the fabrication process itself. Think of this from a factory-planning standpoint. Facility layout is driven by many factors, from operator safety to the efficiency of process flow. It may require that the facility is reconfigurable to accommodate a succession of short-run projects or frequently changing processes.
Frequent changes can lead to unforeseen space and material conflicts, which can then create efficiency or safety issues. But such conflicts can be tracked and measured using sensors, and there is a role for AI in the optimization of factory layouts.
When adopting new technologies where there’s a lot of uncertainty, like additive manufacturing, an important step is using NDT after the part’s been made. Nondestructive testing can be very expensive, especially if it incorporates capital equipment CT scanners (used to analyze the structural integrity of manufactured parts). Sensors in the machines can link to models that are built up from a large data set learned from the manufacturing process for specific parts. Once sensor data is available, it’s possible to build a machine-learning model using the sensor data—for example, to correlate with a defect observed in the CT scan. The sensor data can flag parts that the analytic model suggests are likely to be defective without requiring the part to be CT-scanned. Only those parts would be scanned instead of routinely scanning all parts as they come off the line.
The operation can also monitor how people are using the equipment. Manufacturing engineers make assumptions when the equipment is designed about how the machinery will be operated. With human analysis, there may be an extra step happening or a step being skipped. Sensors can accurately capture that information for AI analysis.
AI also has a role in adapting manufacturing processes and tooling to various environmental conditions where they might be applied. Take, for example, humidity. Developers of additive-manufacturing technology have found that some machines don’t work as designed in certain countries. Humidity sensors in the factories have been used to monitor conditions, sometimes discovering counterintuitive things. In one case, humidity created issues in what was supposed to be a moisture-controlled environment: It turned out that somebody was leaving the door open when he or she went outside to smoke.
Effectively using sensor data requires the development of effective AI models. Those models have to be trained to understand what they’re seeing in the data—what can cause those problems, how to detect the causes, and what to do. Today, machine-learning models can use sensor data to predict when a problem is going to occur and alert a human troubleshooter. Ultimately, AI systems will be able to predict issues and react to them in real time. AI models will soon be tasked with creating proactive ways to head off problems and to improve manufacturing processes.
AI has an important role in generative design, a process in which a design engineer enters a set of requirements for a project and then design software creates multiple iterations. Recently, Autodesk has collected large volumes of materials data for additive manufacturing and is using that data to drive a generative-design model. This prototype has an “understanding” of how the material properties change according to how the manufacturing process affects individual features and geometry.
Generative design is an adaptable optimization technique. A lot of traditional optimization techniques look at more general approaches to part optimization. Generative-design algorithms can be much more specific, focusing on an individual feature, applying an understanding of the mechanical properties of that feature based on materials testing and collaboration with universities. Although designs are idealized, manufacturing processes take place in the real world, so conditions might not be constant. An effective generative-design algorithm incorporates this level of understanding.
Generative design can create an optimal design and specifications in software, then distribute that design to multiple facilities with compatible tooling. This means smaller, geographically dispersed facilities can manufacture a larger range of parts. These facilities could be proximal to where they’re needed; a facility might make parts for aerospace one day and the next day make parts for other essential products, saving on distribution and shipping costs. This is becoming an important concept in the automotive industry, for example.
AI can be also used to optimize manufacturing processes and to make those processes more flexible and reconfigurable. Current demand can determine factory floor layout and generate a process, which can also be done for future demand. Those models can then be used to compare and contrast them. That analysis then determines whether is it better to have fewer large additive machines or lots of smaller machines, which might cost less and be diverted to other projects when demand slows. “What-if” analysis is a common application for AI.
Models will be used to optimize both shop floor layout and process sequencing. For example, applying thermal treatment on an additive part can be done straight from the 3D printer. It could be that the material comes in pre-tempered or it needs to be retempered, requiring another heat cycle. Engineers could run various what-if scenarios to determine what kind of equipment the facility should have—it may make more sense to subcontract parts of the process to another company nearby.
These AI applications could change the business case that determines whether a factory focuses on one captive process or takes on multiple products or projects. The latter would make the factory more resilient. In the example of aerospace, an industry that’s experiencing a downturn, it may be that its manufacturing operations could adapt by making medical parts, as well.
Another key area of focus for AI in manufacturing is predictive maintenance. This allows engineers to equip factory machines with pretrained AI models that incorporate the cumulative knowledge of that tooling. Based on data from the machinery, the models can learn new patterns of cause and effect discovered on-site to prevent problems.
Design, process improvement, reducing the wear on machines, and optimizing energy consumption are all areas AI will be applied in manufacturing. That evolution has already begun.
The machines are getting smarter and more integrated, with each other and with the supply chain and other business automation. The ideal situation would be materials in, parts out, with sensors monitoring every link in the chain. People maintain control of the process but don’t necessarily work in the environment. This frees up vital manufacturing resources and personnel to focus on innovation—creating new ways of designing and manufacturing components—rather than repetitive work, which can be automated.
As with any fundamental shift, there has been resistance to AI adoption. The knowledge and skills required for AI can be expensive and scarce; many manufacturers don’t have those in-house capabilities. They see themselves as effective in specialized competencies, so to justify the investment to make something new or improve a process, they need exhaustive proof and may be risk-averse to upscaling a factory.
This can make the concept of “factory in a box” more attractive to companies. More enterprises, especially SMEs, can confidently adopt an end-to-end packaged process where the software works seamlessly with the tooling, using sensors and analytics to improve. Adding the digital twin capability, where engineers can try out a new manufacturing process as a simulation, also makes the decision less risky.
There’s also a role for AI in quality inspection, a process that generates a lot of data so is naturally suited to machine learning. Consider additive manufacturing: One build generates as much as a terabyte of data on how the machine produced the part, the on-site conditions, and any issues discovered during the build. That volume of data is beyond human scope for analysis, but AI systems can do it now. What works for additive tools can easily work for subtractive manufacturing, casting, injection molding, and a broad range of other manufacturing processes.
When complementary technologies such as virtual reality (VR) and augmented reality (AR) are added, AI solutions will reduce design time and optimize assembly-line processes. Line workers have already been equipped with VR/AR systems that let them visualize the assembly process, providing visual guidance to improve the speed and precision of their work. The operator might have AR glasses that project diagrams explaining how to assemble the parts. The system can monitor work and offer prompts to the worker: You’ve turned this spanner enough, you’ve not turned it enough, or you’ve not pulled the trigger.
Larger companies and SMEs have different focus areas for AI adoption. SMEs tend to make a lot of parts whereas bigger companies often assemble a lot of parts sourced from elsewhere. There are exceptions; automotive companies do a lot of spot-welding of the chassis but buy and assemble other parts such as bearings and plastic components.
Regarding the parts themselves, an emerging trend is the use of smart components: parts with embedded sensors that monitor their own condition, stress, torque, and so on. This idea is especially provocative for auto manufacturing, as these factors depend more on how the car is driven rather than how many miles it goes; if driven over a lot of potholes every day, more maintenance will probably be required.
A smart component can notify a manufacturer that it has reached the end of its life or is due for inspection. Rather than monitoring these data points externally, the part itself will check in occasionally with AI systems to report normal status until conditions go sideways, when the part will start demanding attention. This approach cuts down on the volume of data traffic within the system, which at scale can become a significant drag on analytic processing performance.
The greatest, most immediate opportunity for AI to add value is in additive manufacturing. Additive processes are primary targets because their products are more expensive and smaller in volume. In the future, as humans grow AI and mature it, it will likely become important across the entire manufacturing value chain.
This article has been updated. It was originally published in January 2021.
Andy Harris is an engineering manager based in the London. He is a part of the Autodesk Industry Futures team and leads the R&D effort for this group. Harris has a background in aerospace, automotive, and materials science with 15 years of experience in this area. He has a master’s degree in aerospace engineering and a doctorate in materials science from the University of Surrey. At Autodesk, Harris works directly with industrial partners and universities to provide innovative solutions.
Emerging Tech
Emerging Tech
D&M
