By Peter Seeberg and Kai Peters
Policymakers in Europe have recognized that assuming a leading position in the global artificial intelligence (AI) race will not come for free. In light of this, Germany has pledged to invest 3 billion euros into AI by 2025, while the European Commission is talking about 20 billion euros of public and private investment in the research & innovation of AI by 2020. However, questions still remain on the best way to use these funds to leverage private investment and where research can trigger necessary short-term success of AI.
Betting on catching up or on defending the pole position?
The sums being mentioned by national governments and European Commission may initially sound impressive, however, when considering the urgency in light of fierce global competition, they actually appear to be rather less significant. A shining example of this is the Chinese city of Tianjin. A single city alone is planning to invest USD 16 billion for the research and development of AI and smart manufacturing. From a European perspective, it is therefore vital that the available funds are invested where quick and substantial benefits can be achieved. With this in mind, the fields, applications and levers which promise rapid diffusion and effective leverage effects – such as industrial production and engineering - and which encourage interactions between AI experts and other disciplines should be prioritized.
In order to succeed, Europe can use its advantage of still having excellent and worldwide competitive industries. Although global competitors in the B2C sector may enjoy a competitive advantage ahead of Europe, industrial and mechanical engineering can still play a pioneering role in the use of AI in both the B2B sector and in industrial applications. Caution must be taken here, as this logic can be applied to both sides of the coin. If the integration of AI in industrial and engineering applications is unsuccessful, the leading role of European engineering companies will certainly be lost to competitors from other technology world regions.
The role of mechanical engineering as the linchpin of industrial innovation
As providers of industrial solutions, companies in the mechanical engineering and production equipment sector are the key players in driving technology forward. Embedded in production technologies such as robotics, automation or sensor technology, AI is enabling new solutions and efficiencies in industrial value chains. The investments in AI should therefore not only be made in the ivory towers of the IT industry labs, but also in the boundaries between AI and industrial applications, where the use of AI is highly rewarding, both with regards to scalability and real-world benefits.
However, it is also clear that the use of AI in real-world industrial processes, distinguished by safety, reliability, quality and precision requirements, creates different challenges to those in media or e-commerce. Making low probability predictions will not suffice in addressing the safety, reliability, quality and precision requirements in an industrial setting. As a result, great efforts must be made to increase the reliability of these results, particularly in areas where AI is not yet meeting the requirements of the industry.
- Being intelligent and precise without big data
One of the biggest challenges for industrial AI applications is that there is often a lack of big data. Relevant incidents (such as failures or machine defects) occur less often than, for example, online sales decisions. The challenge here is to enable AI to make very precise predictions with only a limited supply of data, which sounds difficult but is feasible in reality.
The key to this challenge lies in the context knowledge and high-quality data. If a machine only suffers a breakdown once a month, or even more sporadically, it can be difficult to recognize potential patterns leading up to the failure. However, AI models can be trained using the data from a specific period and then concentrate on the fully functional machine. The breakdown, and more interestingly any situations leading up to the breakdown, will show as anomalies to the pattern of a fully functioning machine model. Context knowledge and transfer learning from relevant and similar situations can help analyze these patterns. Research for AI in mechanical engineering should therefore above all be geared towards specific applications in business and industry, on the basis of context-dependent acquisition, selection and assurance of data quality.
- Ensuring explainability and transparency in industrial processes
If an autonomous system, such as a self-driving car, fails, an explanation of this occurrence (and any possible causalities) may be required during legal procedures. Even if the technology is in perfect working order, accepting AI requires a certain understanding of why things happen. Against this background, two decisive topics for industrial AI are the abilities to explain and repeat results. If the reasons behind algorithms finding correlations are inexplicable, these correlations can neither be confirmed nor rejected as causal relations.
Consequently, the ability of an AI model to recognize and predict the occurrence of a specific event needs to be increased to be as close to 100% as possible. If the success rates of predictability models are closer to 50% than 100%, then these models only add very little value. The robustness of an AI algorithm, which is based on a combination of the abilities to explain and repeat results, is a research area in which mechanical engineering can make a large contribution by providing results from specific industrial production use-cases.
- Accelerating data cleaning through interoperability
Approximately 80% of the time invested in a typical project in the context of AI is spent on data preparation and cleaning. Data is extracted from often heterogeneous data sources, outliers are deleted, faulty entries eliminated, timestamps aligned, metadata added, and the clean data then transformed into the required format. The OPC UA format has become the de facto standard architecture for industrial interoperability.
If data is available in an OPC UA information model, the data cleaning phase can be almost entirely skipped, resulting in time savings of up to 80%. In addition, the analysis results are fed back into the OPC UA information model; as such, they are available for immediate use by third parties, making the entire AI process far more rapid. Companies and research projects shall build upon the opportunities for maintaining a global leading position by exporting machines with embedded AI on top of OPC UA around the world.
- Earnest consideration of ethical questions, but check against reality
Fostering the acceptance and success of AI technologies will rely on an ethical approach to shaping the human-machine cooperation. "Human-centered AI", as promoted by the European Union, is the approach to resolving issues raised by unintentionally or deliberately harmful applications. However, the discussion on ethics is not equally relevant for every application scenario, with It playing a reduced role in industrial use, or other areas where AI predominantly focuses on the interactions of machines or processes and not with humans.
The debate on AI ethics must therefore not result in hastily drawn and undifferentiated red lines and unnecessarily restrictions of AI-based applications. Research projects, pilot projects and sandbox settings can certainly contribute to assessing where ethical or legal questions may come into play, provide guidance and check whether the current legal framework is still fit for purpose.
- Taking AI solutions mainstream - transferring technologies to companies
Current AI success stories may well be impressive in terms of their outcome, but they often require huge resources and long development times; two factors which might not pay off when it comes to resolving problems in industrial settings. The current status quo means that only professional data scientists can extract knowledge from data by using complex tools - a dilemma for many companies, as the ability to employ data and AI experts is not a given.
In order to counter the shortage of specialists and time, expertise must be made available in the most productive form possible. The research areas of Autonomous Analytics and Guided Analytics aim to minimize the required interactions and the need to involve a data scientist. Applications in the near-future will be built upon the current applications used by data scientists, thus enabling domain experts to extract value from "their" data. Furthermore, this will help facilitate the use of AI by domain experts in companies and speed up development times.
All things considered, AI can only become a European success story if the technology is made available to a wide range and large number of mid-sized industrial companies. It is therefore important to ensure the efficient transfer of technologies and low-threshold access to new technologies, projects, results and networks.