Edge AI Is Starting To Transform Industrial IoT

A slew of wireless and increasingly multi-modal sensors is being targeted at the Industrial Internet of Things (IIoT), setting the stage for significant improvements in efficiency, higher yield, and reduced downtime.

Wired IIoT devices, such as smart energy meters and breakers, industrial network gateways, and environmental sensors already are well established in factory settings. They have long been designed with redundancy, and are manufactured with hardened materials to withstand harsh conditions, such as exposure to high and low temperatures, moisture, gases, oil, power generation, construction, transportation, and even radiation. Those sensors read the environment and provide alerts if something is out of the ordinary.

Wireless and multimodal sensors take those capabilities much further. A key driver toward AI-enabled IIoT is the concept of Industrial 4.0, which involves digitalization, connectivity, and analytics throughout a manufacturing process.

“Traditionally, we’ve seen customers like TI and NXP in this space, and if you look at their 10-K statements, about 40% of the revenue comes from industrial,” said Pallavi Sharma, director of product management at Imagination Technologies. “But a lot of that is driven by analog. We’re seeing the use case changing now, starting with predictive maintenance, or building an AI into anomaly detection and the like. On the edge, there’s a requirement to make that more intelligent and incorporate more AI, so we’re seeing the transition into Industrial 4.0.”

When it comes to processing AI, there is a considerable difference between factory automation systems and edge AI embedded in industrial IoT. “In factory automation, things are changing in that there are a lot of virtual private clouds available,” said Sathishkumar Balasubramanian, head of products at Siemens EDA. “You can hook up your edge devices and do some processing on that, but they take most of the heavy loads back into the cloud so that someone else does the workload and sends it back. Automation is getting more centralized, more on the cloud, and this happened before AI.”

With edge AI IIoT, engineers are building intelligence directly into hardware devices. “Let’s say I’m operating in the factory in a furnace or a boiler,” said Balasubramanian. “You’re figuring out something is wrong or something needs to change. We could easily automate it. If it looks like the calcium deposit is going up, or the temperature is going up, you go change it. That kind of intelligence needs to be built in. That’s why we call it intelligence at the edge and edge-based intelligence. Most of the things they used to do in a data center now need to be done on the edge because they need to be real-time, and they need to operate when everything else breaks.”

Complicating matters are combinations like mixed-signal design plus machine learning, which means multiple layers of processing need to happen right at the deep sensor edge. “We talk about intelligent sensing,” said Satish Ganesan, chief strategy officer at Synaptics. “We talk about running some algorithms right there at the lowest power profile possible, when everything else is sleeping and the power is very low. Then you come to the next level, the native AI processor. At 100-milliwatt power, you can run 270 million parameters. For a washer, dryer, or an HVAC system in an enterprise or an industrial setting, or a collaborative robot in a manufacturing line, you want a workload-specific model that helps when you have issues with debugging or diagnostics because you’re doing processing at that particular point inside.”

Fig. 1: AI and LLMs are moving from the cloud into IoT, with industrial in the middle of the spectrum. Source: Synaptics

IIoT sits somewhere in the middle of the requirement spectrum between consumer and automotive when it comes to robustness, reliability, and security. “The design cycles are like automotive,” said Hezi Saar, executive director of product management for mobile, automotive, and consumer IP at Synopsys. “But adoption of technology has been slower due to the spread of applications, including industrial automation, building automation, energy, lighting, medical, security, entertainment, and mil/aero.”

Fig. 2: Architectures integrating AI capabilities into IoT devices (AIoT). Source: Synopsys

Edge AI makes the requirements for large-scale IoT easier by enabling devices to become intelligent, autonomous decision-makers. “Processing data locally on the device, or the edge, directly enhances the operational efficiency and safety of IoT applications, especially in challenging, remote, or mission-critical environments,” said Saar. “However, it introduces challenges related to hardware constraints and system maintenance, which can make it harder if not managed correctly. While the benefits will outweigh the hardships, a lot of industrial automation is quite fragmented, and large-scale deployment of AI technologies may take time based on the level of electrification.”

The biggest challenge is the agility and the speed with which the industrial sector is changing. “People want more automation, and that’s driving change on multiple levels,” said Samuel Imgrueth, CEO of Toradex, which specializes in industrial embedded compute. “If you look back 10 years ago, an industrial product was shipped and forgotten. They designed it, they qualified it. It’s a very long process — three to five-year development, verification, validation, then you start shipping your product, and you never touch or update the software. Now, however, there are government regulations like the EU Cyber Resilience Act that mandate device manufacturers to update their devices within two weeks if there is a known vulnerability. There’s been a lot of change in this historically very persistent, slow-moving market, which is now required to be much more agile and dynamic.”

Robotic IIoT and simulation
The line is blurring between Industrial IoT and the Internet of Robotic Things (IoRT), which includes automated guided vehicles (AGVs), autonomous mobile robots (AMRs), wheeled robots, collaborative robots (co-bots), and eventually, humanoid robots. This is where cognitive edge AI can shine.

“The convergence of multi-modal sensor fusion and edge AI is moving robotics and automation beyond fixed, pre-programmed tasks toward systems that are truly cognitive and highly adaptable,” said Synopsys’ Saar. “The next big thing is the shift from automation to autonomy, such as cognitive co-bot systems, bringing true human-robot collaboration. The current generation of co-bots is limited by speed and proximity sensors. Multi-modal edge AI enables a deeper understanding of the human partner, with intelligent task handover and contextual safety.”

Future robots will fuse vision (reading body language and posture), sound (detecting a sudden verbal warning or abnormal machine sound), and force/torque data in real-time. “The robot can interpret a human worker’s intent — such as reaching for a specific part — and automatically adjust its pace, trajectory, or even pre-stage the next tool, rather than just stopping when a human enters a safety zone,” said Saar. “This drastically improves workflow efficiency and psychological comfort for the human worker.”

Together, industrial sensors and edge AI can see, hear, detect, and process environmental information. Robotics will also need touch sensors.

“Whether it’s automotive, industrial, or consumer, you have elements of wanting to sense what’s happening in the real world with mixed-signal capability,” said Synaptics’ Ganesan. “In a robotic hand, a humanoid, or anything else like a co-bot in industrial automation, what do you want to do? You have to understand what kind of material you’re picking up. Is it soft? Is it hard? What’s the dexterity required? That’s how a human hand works. The mixed signal design is something that we focus on, and we add elements of machine learning. It’s not like edge AI running teraflops. It’s megaflops and gigaflops, doing machine learning with very small memory, doing very small processing to run some algorithms at the edge.”

One crucial requirement for letting more robotics loose in the factory is efficient simulation. “A radar simulation tool, for example, can run on a GPU and model one of these little wheeled robots on the factory floor, that runs around and picks stuff off the shelves,” said Matt Commens, director of product management at Ansys, part of Synopsys. “It can generate synthetic data to help train AIs — to train that robot to understand what it is seeing when it sees reflections, for instance.”

By the time the radar signals get down to the CPU or the AI on board, they encounter sensor fusion. “At the simulation level, it’s pretty much software,” said Commens. “They do cross-train both camera and radar. The AI should be able to start inferring from limited data sets what they could be coming up against. We’re not doing synthetic aperture radar with this technology. We’re just bouncing signals and getting signals back. There’s some phase shift in there. It’s turned into IQ data that looks like gobbledygook. But if you push enough of it through an AI, it can start to infer some patterns.”

Simulation also can help create an effective test plan. “The design engineer must first imagine and anticipate the absolute worst conditions under which the device will operate,” according to Keysight. The design engineer must then test individual components within a device, as well as the finished product, to fully explore those conditions.”

Fig. 3: The IIoT requires low-energy, low-maintenance, and high-performance devices that can run in challenging environments in real-time. Source: Keysight

Digital twins are also key. “If I’m able to build a factory, I need to be able to mimic how the factory’s electronic circuits or the control unit behaves,” said Siemens’ Balasubramanian. “When we are talking about a digital twin, we are really focused on digitizing the entire factory before you even build a factory, or before you even build a car.”

Edge language models for IoT
In addition to edge AI/ML capabilities, IIoT increasingly will be connected to a domain-specific language model. “We can create a foundation model for a car manufacturing factory that might not be huge, but it still needs processing power to run,” Balasubramanian noted. “If you go to a terminal of the IIoT device in a factory and ask how a particular system is doing, it should be able to run the diagnosis right there at the edge and give you the results. Those things need some power. That’s where some of the shift is happening on the edge, as well.”

Specifically, some edge devices will be connected to a number of small language models — what Infineon calls edge language models (ELMs) — to meet different needs, thereby avoiding going back to a larger model on the cloud if it doesn’t know the answer. “What’s running on the MCU is not just a language model,” said Steve Tateosian, senior vice president of Infineon‘s IoT, compute, and wireless business unit. “There’s a set of models on there. There’s a weight word detection model running in a low power domain, so we’re talking about always-on audio in a battery-operated device that will last for weeks, not a few hours.”

Previously, there were specialty models on the edge, with one for voice and another for audio, for example. Those specialty approaches for ML on the edge are being replaced by consolidated models that support a lot of different inputs or contexts. “That’s requiring a change on the hardware side, in terms of doing things a different way,” Tateosian said. “That software stack then requires a different set of attention to it in terms of how we deliver this performance or those results on the edge.”

And because of the way ELMs are created and the focus of the model, the answers are more accurate than those generated by an LLM. “You have less concern that you’re going to get some hallucination, or something that is just wrong coming out of the model,” he said. Over time, the models will continue to be more responsive. “I can recognize the user voice inputs. I can have vision inputs in the future, or environmental audio sources. You can think of all of the different types of data that can come into these systems, and then it can analyze that data and make a recommendation for you.”

Fig. 4: A robotic head with edge AI processing. Source: Semiconductor Engineering/Infineon

Processors used in IoT devices
IoT devices typically use a combination of MCUs and NPUs to process edge AI and LLMs, but GPUs increasingly have a place as AI becomes more prominent. OEMs also might put in accelerators that do very specific things, such as vision accelerators to process bug detection on the factory automation assembly line.

“They’re still going to need very powerful computer vision accelerators, but when it comes to other things that are running AI algorithms, or enhancing whatever that function or feature is, they cannot think of everything today,” said Imagination’s Sharma. “There might be newer things in two to three years. Having powerful GPUs built in here protects them from that. They can use that same SoC, and they can run a lot of these general-purpose algorithms.”

The rise of wireless IIoT
While autonomous mobile robots are driving the trend toward more wireless devices, fixed IIoT devices may be a combination of wired for power and wireless for software updates and data transmission.

“If you have to upgrade the software in those remote sensors, and it’s the same software upgrade, having to go to each one of them and do it is a very time-consuming process,” said Ananda Roy senior product manager for low-power edge AI at Synaptics. “You can just send the upgrade to the hub, and then because it’s connected wirelessly to all these devices, you can do the one-time upgrade of all those devices around you.”

Devices can process data locally, then send their findings to a central processing hub. “This is where you get all your data from different sensors into one device so that you can read off of it,” said Roy. “With wired devices, you have to bring in wires from all over, and that’s very expensive. You have to drill holes in the walls because you cannot have wires missing. The devices across your factory floor could still be wired, but this hub definitely can be wireless, and then you export data from the hub to the cloud very easily using a wireless connection.”

One of the key challenges of using wireless technology in industrial applications is latency. “Typically, if you are transmitting data wirelessly and you have dropped packets because of a noisy environment, it introduces latency,” said Roy. “Most industrial applications are very latency-constrained, but that’s what can be solved with Wi-Fi 7, which brings down latency to sub-10 milliseconds. Our robotics partner, Grinn, built a robotic arm you can control through a tablet entirely over Wi-Fi. Robotic applications have to be low latency. You cannot press a button and then wait for the robot to respond a few seconds later.”

Grinn also implemented a Synaptics touch controller into the robotic hand to enable real-life detection of pressure on different areas of the fingers. “We can create different types of sensors,” said Robert Otręba, CEO of Grinn. “We can implement even more sensing points, or fewer, depending on the complexity of the end application. In this way, we have not only a controlling feature, but also constant feedback. So the same robotic hand could work with hard or soft components without breaking them.”

Further, mobile robotics systems need to be able to communicate with each other, and this likely will be done via ultra-wideband (UWB) wireless, which uses short-range radio waves, sending 2ns pulses to provide precise location and distance measurements. The technology is already established for asset tracking big tools in a warehouse or hospital, or for tracking small, expensive equipment that gets lost easily. It also enables robot-to-robot and human localization tracking so they can work more closely together.

“It does the ranging between the anchors,” said Schweta Bagchi, senior manager application marketing at Infineon’s IoT and Sensor Solution Group. “For example, if I’m wearing a wristband and walking, the robot will stop if I come closer, even if I am not paying attention. It knows exactly there’s a human nearby.”

Built on 22nm, Infineon’s UWB accuracy is at the centimeter level. “In a warehouse, if you have an automated guided vehicle, you can have a lot of technology like lidar, but everyone will have a blind spot,” said Bagchi. “Ultra-wideband is something that can communicate between the two devices, so two robots can talk to each other. Safety is one aspect of it. The second is, when they talk to each other, they also can understand, ‘I have X amount of work. I can take on more. Shall I come to you?’”

This robot-to-robot communication might have ML behind it, with simple data communication based on certain logic, or it could be an LLM. “When we are working with human lives, we expect that they might speak more natural conversation,” said Bagchi.

Fig. 5: A mobile robot easily deployed on a factory floor to lift and transport goods. Source: Infineon

Wireless technologies are no longer just a bit pipe to get data from point A to point B. “You can do a lot more things like Wi-Fi sensing, channel sounding using Bluetooth, which is a ranging technology, so that additional ranging like proximity detectors or PIR sensors, or radar is needed,” said Roy. “You can use the Wi-Fi that’s already in the device to know more about your environment.”

Designers can also get creative about generating more power without being plugged in or relying on batteries. “Once you get to very, very low power, you can start using environmental stray energy to power the chip,” said Marc Swinnen, director of product marketing at Ansys. “If you’re stuck up against some girder, and the machines are rolling and moving, there are constant vibrations. You can tap into that to top up your battery or keep it topped up, just out of vibrations, or even the pressure or temperature differences.”

Conclusion
The industrial IoT is changing rapidly due to recent developments in edge AI and ML, multimodal sensors, and robotics to form a cognitive, mobile, and wireless network of devices across the modern factory floor.

“Industrial sensors used to be pretty rudimentary, such as, ‘I’m reaching a critical temperature. I need to shut down,” said Imagination’s Sharma. “It was a tabular form of seeing something and making a decision. Now it’s a lot more nuanced. It can look at the thermal sensor, it can take data from the sound sensor, and it can correlate that to make some decisions. The AI models are being trained for these specific industrial applications on the edge. And we’re just scratching the surface.”

Related Reading
Complex Mix Of Processors At The Edge
Diversity of compute elements proliferates for inference, but the mix varies by application.
Smarter Packaging: How AI Is Reshaping Assembly And Materials Control
From predictive maintenance to excursion monitoring, AI is redefining yield management in multi-die assembly.
Packaging With Fewer People And Better Results
Manufacturers are turning to automation, AI, and robotics to boost repeatability, cut costs, and support heterogeneous integration.
Increasing Roles For Robotics In Fabs
AI and robotics are taking on bigger, more complex, and increasingly autonomous tasks, but integration with existing equipment and processes remains a formidable challenge.
Predicting And Preventing Process Drift
AI/ML are increasingly vital for good yield and reliability, but there are still plenty of pitfalls to avoid.