While there are many attractive use cases for AI in manufacturing, this game-changing technology is still at the start of its journey to adoption.
The sophisticated connected industrial machinery that epitomizes Industry 4.0, combined with the desire to further optimize manufacturing processes, makes industrial manufacturing a prime candidate for adoption of artificial intelligence. ABI Research’s figures have the total installed base of AI-enabled industrial machines growing at a CAGR of 64.8% in the next five years, to reach 15.4 million units in 2024.
Today, however, the installed base is still relatively small. While artificial intelligence technology is revolutionizing many other sectors and empowering businesses around the world to get value from their data, industrial manufacturing has not yet fully embraced this game-changing technology. What unique factors in this segment are holding back the adoption of AI?
AI has potential applications in all the different phases of industrial manufacturing. It may be used in generative design for product development, or for production forecasting for inventory management. AI can also be used in machine vision applications on the production line, performing tasks such as defect inspection or production optimization, and can be used in predictive maintenance systems for machinery. Today, some of these applications are finding their way into factories, while others are still waiting to break through.
“Given that unplanned downtime has historically been the main adversary in the manufacturing world’s rogue’s gallery, predictive maintenance has clearly been a key application utilizing AI over the past few years,” said Chetan Khona, director of industrial, vision, healthcare, and sciences at Xilinx.
Chetan Khona (Image: Xilinx)
“We have seen the sophistication and the scope of the approaches to AI increase as awareness and access to newer methods become available.”
For example, complex digital-twin systems can enable all kinds of what-if AI-based analyses, including predicting failures. (Digital twins are complex digital models of large real-world systems that typically rely on extensive sensor networks to gather information from the real world to feed into the model).
Avoiding the cloud
A key factor that has rapidly accelerated the adoption of AI in other business sectors, including finance, is access to readily available computing power in the cloud. Unfortunately, utilizing the cloud is not an option for factories. Like most practical solutions to real-world technical problems, Khona noted, AI is subject to the inconvenience of physics, as well as to people’s perceptions and fears.
“Physics dictates local application deployment, because the control rate of most industrial systems is 10 milliseconds or below — often below 1 millisecond,” he said. “Even if your network operated at the speed of light — the theoretical best-case scenario — you are not going to make it to the cloud and back in the required timescales to impact foundational operations with AI algorithms.”
Xilinx also continues to hear that its customers’ customers have security and privacy concerns.
“[Manufacturers] don’t want their data, at least not all of it in its rawest form, leaving the factory floor,” Khona said. “There is legitimacy in their concerns because the reality of the situation is, even with the best embedded and system-wide cybersecurity available today, cyber-threats are constantly evolving. When you consider the expected lifecycle of factory assets, your security gets weaker over time because the threats are getting stronger.”
In practice, this means AI computation in the manufacturing environment must therefore be performed in the edge device. Implementing AI in edge devices at the different system levels comes with various challenges, said Khona.
At the lowest level of the industrial hierarchy, in embedded systems such as industrial drives and motor control systems, AI has the biggest potential to affect efficiency and reliability because of the sheer number of such systems.
“However, embedded systems development takes time — often four to six years from inception to production — so we haven’t seen what is to come hit the market yet, despite the interest,” said Khona. “Also, AI is only welcome in embedded systems if it can be applied without significantly compromising power, lifecycle, operational performance, and price.”
At the system level, AI can manage all the distinct pieces of machinery and orchestrate their synchronization. Security and latency are both big challenges when combining data from system-wide physical, operational, and human assets at this level.
Khona also highlighted the comparative lack of development of algorithms for analyzing factory data.
“Machine data that feeds into AI algorithms is often time-series data, and seemingly much of the world’s most popular AI work is with image data, which is not considered time-series,” he said. “While image data makes up the majority of data used in AI today due to the sheer number of surveillance cameras and the density of image data, machine data will come close to matching it in a few years. Continuing development of RNNs [recurrent neural networks] and LSTM [long short-term memory]-based models will need to happen to make AI more pervasive in industrial applications.”
In the meantime, some applications can be processed using common models based on convolutional neural networks (CNNs). This is achieved by plotting the time-series data and applying models on that image, which is the approach Xilinx takes with its SPYN-AI motor control kit.
KUKA robots in operation in the aerospace industry (Image: KUKA)
The large number of use cases, all at different levels of maturity, makes AI adoption in manufacturing an extremely complex affair.
“Some AI use cases are much more mature, such as predictive maintenance and machine vision, as they can rely on rule-based AI in the case of predictive maintenance, or benefit from the advancements in other sectors in the case of machine vision. Other, more advanced use cases still require many trials and much R&D,” said Lian Jye Su, principal analyst at ABI Research.
Lian Jye Su
(Image: ABI Research)
Su said that leading industrial machine and robot vendors have in fact been implementing rule-based AI for some time. While these systems generate and collect large amounts of data, they are kept proprietary and are governed by stringent protocols to ensure the highest levels of accuracy and precision. This means industrial manufacturing has been slower to implement data-driven AI solutions than sectors such as finance and enterprise software. “The industrial manufacturing sector has missed out on the boom of data-driven AI that has transformed many other industries,” he said.
One particularly big challenge manufacturing companies face is building and training in-house data science teams for AI implementation. “I believe any highly skilled AI talent will more likely choose to work for major cloud AI vendors than for a manufacturer,” said Su. “Already, there has been an AI talent war going on in the industry, and any manufacturer who is trying to get into the war will only be on the losing end.”
The shortage of AI and data science talent is effectively shaping the industry as manufacturing companies instead rely on partnerships with cloud service providers and with some of the growing number of pure-play AI startups to develop their AI capabilities (see sidebar). System integrators, chipset and industrial server manufacturers, and connectivity service providers complete the picture.
U.S. versus China?
Industrial robotics will benefit from AI when the industry adopts it. (Image: Xilinx)
Manufacturing is key to the economies of both the U.S. and China. In both markets, labor is getting more expensive and profit margins are shrinking. AI-enabled industrial robots have become competent enough to replace human workers and can help each side build competitive advantages through automation.
ABI Research figures place the installed base of AI-enabled end devices in both markets at a similar level, but perhaps unsurprisingly, there are striking differences in the way AI is entering the industrial sector in the world’s two most powerful countries.
In the U.S., the ecosystem of companies applying AI to manufacturing is growing quickly. “The U.S. adopts a diverse approach when it comes to AI investment strategy,” Su said. “Top AI startups in the United States come from various sectors, including self-driving cars, industrial manufacturing, robotics process automation, data analytics, and cybersecurity. Naturally, most top industrial AI companies come from the U.S.”
The large U.S. cloud service providers, AWS and Microsoft, have also established partnerships with industrial AI development platform vendors, AI software vendors, and system integrators, Su said. Those partnerships will develop solutions for manufacturing use cases such as operational efficiency.
Xilinx offers its Alveo accelerator cards to enable the data collection, aggregation, processing, and modeling that goes into AI-guided decision-making in a secured, low-latency, on-premises setup. (Image: Xilinx)
On the other hand, China has chosen to focus on certain strategic areas of AI technology. “This has translated into the creation of hugely successful machine vision AI startups with large market reach and deep technology, including SenseTime, Yitu Technology, and Megvii, but [China] has been slow in pushing for digital transformation in its manufacturing industry,” Su said. “A lot of industrial and manufacturing startups in China are still focused on the connectivity layer, instead of pure-play industrial AI solutions.”
And while China’s cloud service providers are no less visionary than America’s (Alibaba has developed its own suite of AI-based solutions for industrial manufacturing, for example), technology on its own is not enough to ensure rapid adoption.
“These [Chinese] cloud AI companies do not often have the right connections and go-to-market channels to reach many provincial and municipal manufacturers where the majority of small- and medium-size manufacturers reside,” Su said.
Meanwhile, Europe is a recognized center for advanced manufacturing technologies, including many of the well-known commercial robotics companies, including ABB, KUKA, Universal Robot, and MiR. Su said that ABI Research has been monitoring innovative manufacturing AI use cases from companies such as BP, Volkswagen, Airbus, Novartis, and Sanofi.
“The manufacturing sector in Europe is much smaller as compared to the U.S. and China, but they probably fall in between the U.S. and China in terms of AI adoption rate in industrial manufacturing,” said Su. “The rate of AI adoption is slower than in the U.S. due to the lack of scalability.
“Unlike the U.S.’s single market structure, Europe is made up of many nations with diverse geopolitical, language, and cultural differences between them, creating much bigger localization and distribution challenges for AI vendors.”
>> This article was originally published on our sister site, EE Times Europe.