Are you ready for next-gen industrial AI?

Are you ready for next-gen industrial AI?

Artificial intelligence (AI) has tremendous potential in industrial applications. Now, the challenge is about turning that potential into business value at scale. This will involve creating novel processes and overcoming new and interesting challenges, many of which are more human than technical. 

Would you trust the algorithm?

A bad decision in an industrial context can have serious consequences. So when a senior engineer disagrees with a recommendation from an advanced AI system, how will organizations decide? To what extent will organizations trust AI to control critical systems? To understand the attitudes, benefits and barriers to implementing AI, Siemens partnered with Longitude Research to survey over 500 senior executives from energy, heavy industry, infrastructure and transportation. Their responses offer a unique insight into the future of industrial AI.

Download the report “Next-Gen Industrial AI”

The research explores five factors driving next-gen industrial AI

  • Artificial understanding:
    AI’s ability to integrate data and use contextual data will lead to powerful use cases.
  • AI everywhere:
    The Internet of Things, edge computing and embedded intelligence will drive greater predictive power, more autonomous machines and richer simulations.
  • Smart data exchanges:
    Next-gen AI breakthroughs will depend on organizations sharing data. Who will share with whom?
  • Human-AI collaboration:
    As AI evolves from tools to colleagues, what happens when humans and AI disagree?
  • Managing AI safety concerns:
    Our machines, vehicles, and equipment are becoming autonomous. How can we trust AI and ensure that it makes safe decisions?

Expert interview: “An extremely fast way for industrial organizations to make use of AI”

Gerhard Kreß is responsible for data services at Siemens Mobility, including the global MindSphere Application Centers for Rail which co-create data-driven solutions with customers. In the interview he explains how the combination of new generation of algorithms and an increasingly automated data management could support some of the most impressive use cases in the next generation of Industrial AI.

How will the next phase of AI development differ from how things have been done to date?

Until recent years, most AI applications were developed using a supervised learning approach. In other words, you would tell the system what is correct and incorrect, what the main parameters are, when to learn. For this you need a training dataset with hundreds of labels and label conditions. In an industrial setting where applications require deep domain knowhow and there is never enough labelled data, this is a cumbersome approach. Today, with deep learning approaches, you don’t need to pre-label things, because the networks essentially teach themselves. This technology, which is widely used for image recognition, is now being applied in other fields like quality control in manufacturing, medical imaging and autonomous driving. It is an extremely fast way for industrial organizations to make use of AI. While technically still difficult to do, powerful toolsets that make this technology more accessible are becoming increasingly available. As a result, applications can now be built in a couple of days without the need for Ph.D. scientists or specialist programmers.

What are the implications of this kind of approach for the future capabilities of AI?

I think we are at a tipping point, where AI applications are reaching or surpassing the performance of humans, in specialized areas. So far, this has been mainly in narrow areas. For example, for quality control in manufacturing, where AI can identify product defects far better than humans.   But what has changed in the past year or two, is the emergence of more generalized algorithms, like DeepMind’s AlphaZero, which taught itself how to master the games of chess, shogi, and go, beating a world-champion program in each case. This shows that a single algorithm can acquire new knowledge and abilities.   This is the beginning of a move towards systems that are more topic independent, or general purpose. Over the next few years we will see more impressive examples that can handle multiple tasks better than human beings. 

Will many areas of analytics and decision-making become more automated? Are there specific areas that will still require experienced or well-trained human minds?

Historically you would spend about 70% of your time preparing datasets for machine learning applications. That is certainly becoming more automated. Also, the heavy lifting involved in deriving insights from those datasets is starting to become automated. We have the first models in place, not yet deep learning based, but good enough to support partial automation of findings. In two to four years, I believe off-the-shelf packages will be readily available to automate insight creation, without the need to train it on your use case or domain. While we might be able to automate data insights, determining the best course of action remains tricky. This will still require strong industrial domain knowhow and a wider understanding of customer processes, industry outlook and risk levels. Let’s consider the risk factor. In the B2C world, if an algorithm makes a movie recommendation that does not match your interests, no big deal. But in an industrial setting, an inaccurate recommendation could have catastrophic consequences. These high impact decisions won’t be done by algorithms any time soon. But the insight generation, in five years this could be 100% automated.

What about assembling and organizing those large datasets in the first place – this can be a major barrier for organizations – how are things changing in this area? 

Over the past 10 to 15 years there has been a trend of pulling all data together in one data warehouse and building a dashboard on top to get an overview of everything. This can take years to achieve as you need to normalize the data while you assemble it. It has never really worked because it takes too long, costs too much and does not cope well with change.   What you increasingly see is an approach where you keep the data where and as it is – in different silos and formats. You don't integrate it in the traditional sense but add a layer on top that glues it together and makes it accessible. This is attractive, because it saves a huge amount of time and effort, and data owners do not need to give away control of their databases or make big changes to their data management model.   This is the approach we use for our customers in the rail sector. It takes two to three days to develop an application instead of two to three months and requires up to 10 times as much CPU power. But with CPU power becoming cheaper, this approach makes more sense as we transition to an Industrial Internet of Things, where I don’t know today, what I will ask of my data tomorrow.

Benefits and barriers of AI adoption in industrial organizations

Over the next three years, as industrial organizations increase the adoption of artificial intelligence (AI), many of the long-standing barriers to progress will fall in significance. According to the respondents, data integration and quality issues will fall the most, from 73% saying this is a major or moderate barrier today, to 31% by 2022. At the same time the advance of next generation AI applications will see many more organizations benefit. In the visualization below you can explore how respondents rated several specific barriers and benefits, both for today and for three years’ time.
Podcast with Michael May

“Al can explore solutions that no expert has thought about”

Podcast with Michael May, AI expert at Siemens

How can companies who are not Amazon and Facebook make money from AI? Michael May, Head of Company Core Technology Data Analytics and AI at Siemens, explains how industrial organizations can benefit from AI and how soft criteria, such as human behavior, can be a crucial condition for success.

Listen to podcast

Additional Spotlights

Related Content

Knowledge Graphs

Knowledge Graphs

Artificial intelligence has reached the point where it may be set to trigger the next wave of technological disruption. Graphic representations of relationships fill the gap where deep learning comes to its limits. A world of new applications may be the result.
AI on the shopfloor

AI on the shopfloor

A whitepaper, envisioning the future of intelligent automation and its impact on manufacturing.

AI in industry

Transporters find their way through factory halls on their own, plants optimize their power consumption during live operation, and machines perform quality-control checks while manufacturing is still in progress.

Digital transformation

Leading by example: With the Digital Enterprise, Siemens is consistently pursuing the digital transformation of industry – both at the customer’s site and in-house.

Intelligent disinfection robots

Developed by Siemens and Aucma in just one week, from design to sample production, intelligent disinfection robots join the battle against the corona virus and other viruses in hospitals.

Cloud-based pump solution

Setting new standards in the food and beverage industry: A digital, cloud-based solution enables a food processing company to monitor motors, pumps, and systems, and use plant data for process optimization.

Increasing competitiveness

A manufacturer of clamping technology and gripping systems is using digital twins to help plant manufacturers and users increase their competitiveness. Digital twins enable the virtual simulation and testing of all process steps.