Manufacturing Shifts To AI Of Things

AI is being infused into the Internet of Things, setting the stage for significant improvements in manufacturing productivity, improved uptime, and reduced costs — regardless of market segment.

The traditional approach to improving manufacturing equipment reliability and efficiency is regular scheduled maintenance. While that is an improvement over just fixing or replacing equipment when it breaks, it’s far from optimal. Even with periodic maintenance, equipment can suddenly break down, idling workers, delaying shipments, and disappointing customers.

This is where AI fits in. Industry 4.0 — called smart manufacturing in most regions outside of Europe — has been on the horizon for nearly a decade. Progress has been spotty, in part because the technologies needed to make it work have developed unevenly, and in part because adoption has been limited, sometimes even within the same organization. But those technologies now are beginning to reach a level of maturity where they can be relied upon to improve productivity and avoid unanticipated equipment breakdowns.

Much of this can be correlated with the build-out of the edge, improvements in the cloud, and evolutionary advances in algorithm training and inferencing. Collectively, they are providing the foundation for what increasingly is being called the Artificial Intelligence of Things, or AIoT.

Beneath that umbrella concept, some confusing terms are coming into better focus. AI refers to machine intelligence, while machine learning — a subset of AI — is being defined as the process of acquiring knowledge essential to AI. So machine learning is used to train a system to make the right decision or selection, and AI is the system that makes those decisions.

At first, machines can be expected to make mistakes. But over a period of time and many trials and errors, they can “learn” to make the correct choices, or to infer the right answers. For example, an AI system can be given 1 million pictures to learn what a dog looks like. At first, it will learn that a dog has two ears, two eyes, a nose, and a mouth, but so do other animals. After many samples, the AI system can learn the subtle differences.

In a manufacturing operation, AIoT enables devices and software to network with other AIoT devices, including smart sensors, smart gateways, and local servers. Add in AI data analytics, and some decisions can be automated, which is often referred to as actionable intelligence.

AIoT reaches beyond the benefits of IoT by getting more accurate, real-time information for stronger problem solving. Additionally, better data analytics produce scalable, lower-cost solutions. The result is better a customer experience, enhanced product development, and the potential for higher revenue.

In the past, all these decisions were made in the cloud. Sensors collected data, sent it back to the cloud, and waited for the cloud to process that data and make sense of it. That approach has proved to be inefficient and expensive, while bogging down networks with data traffic. Imagine, for example, millions of IoT device endpoints doing this simultaneously.

With edge computing, AIoT decision-making happens at the network endpoint, or at least much closer. That reduces latency, bandwidth requirements, and results in much greater efficiency in manufacturing operations.

“Introducing edge computing to store data locally reduces transmission costs,” said Ron Lowman, strategic marketing manager for IP at Synopsys. “However, efficiency techniques are required to remove data waste, and the predominant method today is adding AI capabilities on low-power edge computing server CPUs, with connectivity to AI acceleration SoCs in the form of GPUs and ASICs or an array of these chips.”

AIoT is gaining momentum in manufacturing, industrial and chemical processing, energy management, surveillance, lighting, traffic management, and smart parking. It also has found a role in noise and water quality monitoring, remote medical devices, and home health care. Fleet management, smart waste management, leakage detection, smart retailing, and supply chain management also rely on AIoT, as do robotics, smart home, and an increasing list of other applications.

Ultimately, the goal is to optimize the production process, product quality, and cost effectiveness.

Smarter chips
The underlying technologies that make this all possible are various types of smart chips, which can be trained to recognize what is considered a normal operation and establish a performance profile for the manufacturing equipment. When the behavior of the equipment is outside the predetermined profile, the AIoT will sound a warning.

“A smart chip is more than an SoC,” said Paul Daigle, senior product marketing manager, industrial IoT, at Silicon Labs. “A combination of onboard sensors, analog-to-digital converters (ADCs), microprocessor units (MPUs), and radio — along with automatically improving algorithms — makes next-generation chips smarter. These embedded devices with machine learning capability can be initially trained in a known environment. Depending on the applications, it may take as little as 1 to 2 minutes, or as much as 30 minutes, for the device to be trained to make correct decisions. Trained devices have all kinds of potential use cases. A trained device can be attached to the outside of a pump to monitor its wear and tear and provide early warning when the bearings need to be replaced. An occupancy sensor with embedded machine learning (ML) and a microphone can enhance safety and security by identifying the sound of broken glass. Or a smart chip with a low-resolution infrared sensor can count people without a camera to help building managers resolve energy performance.”

Fig. 1: Hierarchy of intelligence. Source: Silicon Labs

What information is collected?
Various types of manufacturing require a wide variety of production equipment. While the equipment for processing semiconductors, pharmaceuticals, or even food is very different, they all share some common parameters, such as temperature, vibration, noise, energy consumption, humidity, and in some cases magnetic fields. Intelligent sensors, which are part of AIoT, can be used to capture some or all of these parameters in real time. Using the captured data, AIoT analytics then can determine if the equipment is operating within its acceptable profile.

“There are various sensor types available including accelerometers, temperature sensors, voltage sensors, microphones, cameras, IR, RF, and others that can be used to assess normal operation and detect changes that could signal the need for preventive maintenance,” said Dana McCarty, vice president of sales and marketing for Flex Logix’s inference products. “Multi-sensor machine learning models will combine information in ways not obvious to humans to perceive the state of equipment repair.”

Others point to similar trends. “AI can be very useful in manufacturing,” said Dennis Laudick, vice president of marketing for machine learning at Arm. “In the case of preventive maintenance of motor control, for example, AI will first learn what is normal, based on the manufacturer’s specifications. Then it will detect the parameters outside normal. AI can tell the difference between normal versus abnormal noises. It can tell the difference between the sound coming from motor operation within specification or a defective motor. With sensor fusion, together with temperature, vibration, and noise, AI can provide a very accurate prediction on the condition of a running motor.”

Improving accuracy of results
Accuracy is critical, and in a manufacturing facility it often requires some technology innovations.

“It’s essential to have good sensors, because false sensing gives you poor data,” said Shawn Slusser, senior vice president of sales and marketing at Infineon. “Now, when you combine multiple sensors together, you can do sensor fusion, so even if one sensor tells you one thing, you have multiple sensors that can validate what it’s telling you. It’s kind of a voting system for different kinds of sensors. That’s becoming more and more affordable and easy with the compute and the different sensors.”

That includes everything from electronic sensors to MEMS devices, and putting them together in a single package with other sensors. MEMS devices, such as pressure sensors, gyroscopes, and accelerometers, are common in many applications. In fact, accelerometers first appeared in high volume during the 1990s for measuring tire pressure and for automobile airbag deployment. They now are used in cellphones, automobiles, aircraft, drones, microphones, and smart cities to measure temperature, vibration, sound, gases, and chemicals.

While MEMS always has involved complex packaging and integration with other chips, it takes on a whole new level of complexity — and potential usefulness — when these sensors are combined with other sensors in a sensor-fusion type of application.

SEMI has been promoting smart manufacturing for the better part of a decade, and the German government has been working on its own version, called Industry 4.0, since it was first introduced at the Hannover Fair in 2011. The initial goal for both was to fuse together computerization with manufacturing. Since then, SEMI has created a number of standards for advanced analytics and process control, data management, and surface-mount technology assembly lines automation. It has further proposed standards to support the “Intelligent Future” of smart data and artificial intelligence, including intelligent homes, workplaces, transportation, health, infrastructure, grid, and smart manufacturing.

“Our goal is to ‘sensorize’ the world,” said Tim Brosnihan, executive director of SEMI’s MEMS & Sensors Industry Group. That will likely include some combination of MEMS and other sensors going forward, particularly as the edge build-out continues.

MEMS will continue to be an integral part of this shift. “To our knowledge, all sensors used for preventive maintenance in manufacturing today are MEMS-based, be it mechanical or electrical,” said Sree Harsha Angara, product marketing manager for internet, compute, and wireless at Infineon. “Size, power, and cost are key requirements for preventive maintenance applications, and MEMS technology plays a key role in hitting them.”

Connecting systems
Unless you are building a smart factory from the ground up, achieving preventive maintenance requires digitalization and a great deal of planning and new investment. Within most factories, old equipment without intelligence is still being used, sometimes retrofitted with sensors, while at other times digitally isolated from newer equipment.

Alongside of that, different types of connectivity are being used today, both wired and wireless. In the early days of factory automation, serial interfaces such as RS-232 and RS-422 were popular. Today, different types of connections must co-exist. They include wired Ethernet, WiFi, Bluetooth, and various types of LPWAN. Connecting AIoT, sensors, and IoT devices to the servers and/or the cloud can be challenging, and a pathway to connect the old and the new is necessary.

“There are some nuances here, but broadly the trend is WiFi/Bluetooth for the nodes on the machine, followed by an aggregator that translates from WiFi/Bluetooth to cellular connectivity for cloud connections,” said Infineon’s Angara. “Connecting directly to the enterprise, WiFi is not always allowed for security reasons, so even if the node has WiFi, there can be a WiFi-to-cellular aggregator. Wired Ethernet is also seen, but it is less popular for the retrofit model as running wires is added infrastructure cost.”

Additional challenges include mixing smart sensors, dumb sensors, and devices with and without edge computing capabilities. Careful planning and deployment are called for. The technical community as a whole is leaning toward using edge computing to replace cloud-centric computation.

So far, however, there is little quantitative information comparing the performance based on cloud computing, edge computing, and a hybrid model of mixing cloud and edge. While this isn’t slowing down the migration to AIoT, it does raise questions for which there are no answers yet.

“With edge computing it is essential to do some necessary computations in a centralized place,” said Andy Heinig, head of the department of efficient electronics in Fraunhofer IIS’ Engineering of Adaptive Systems Division. “Only the centralized computational unit has all the information from different sensor values needed for sensor data fusion. The centralized places also can be organized hierarchically. Some centralized computational units can be near the edge devices, with others in the cloud. Currently no software support takes into consideration edge and centralized node performance differences in order to co-optimize edge and central node AI algorithm partitioning and efficiently process network traffic.”

The future
Preventive maintenance is an important part of smart manufacturing, but this is just the beginning. AIoT can be deployed in many different areas in a factory to further increase productivity. For example, it can be used for incoming inspection. Traditionally, the quality control department performs sample inspection. Instead of inspecting 100% of the components used to build a device, only a sample — say 10% — will be audited. With the installation of a 3D HD camera, AIoT can inspect 100% of the components and screen out defective parts at an early stage. Additionally, a robotic arm can pick out defective components or those of different colors and/or shapes, further reducing reject rates.

AIoT also can be used to improve worker safety, resulting in lower worker compensation payments. For example, a warehouse can be equipped with AIoT cameras to ensure only authorized workers wearing appropriate safety equipment can enter the warehouse.

Smart manufacturing no doubt will increase productivity and profitability. But careful planning on how everything should work, including machine learning requirements and implementation, is important to future success. Realistic ROI expectations of investment should be included in the planning.

“Investment in sensor/processing/communications technology should be at the MCU level,” advised Flex Logix’s McCarty. “You also need to be realistic about the need for potential lengthy training and algorithm development and field testing, but two to three years should be sufficient to develop reliable technology to support predictive maintenance for many applications.”