Artificial Intelligence: Targeting the Future with Standards

What do we mean by artificial intelligence?

There is no universally accepted definition of "artificial intelligence". The range of technologies and perspectives with regard to AI is too broad for this.

Under the heading of "artificial intelligence", however, the DKE primarily considers technologies that can identify patterns in extremely large, differing data sets, for example in image and speech recognition, and that can optimize their own performance within predefined ground rules on the basis of data from training and ongoing operation as, for example, in robotics or autonomous driving. AI systems continue to be deterministic, but internally so complex that humans are only able to understand their behavior to a very limited extent.

How does artificial intelligence actually "work"?

The mode of operation has constantly evolved since the term "artificial intelligence" was coined in the 1950s.

The AI systems that dominate today are those neural networks that bear some resemblance to the way the human brain works. Numerous nodes (neurons) are interconnected, with the importance of the individual connections being gradually determined by training a specific task. It is only through this learning process that a neural network becomes functional and is able to be used as an algorithm to solve tasks. But how does the learning process actually work? The keyword here is: machine learning.

Where will artificial intelligence be used in the future?

AI is being used in a wide variety of areas, and the trend is rising – from Industry 4.0 and the energy industry to autonomous driving and healthcare. In the following, we describe some examples:

Industry 4.0

For tomorrow's industry, i.e. "Industry 4.0", artificial intelligence offers a huge opportunity to revolutionize the entire automation industry. Industry 4.0 focuses on the networking of all plants and robots. Applications based on AI can help to optimize process flows using a vast amount of data. However, this will only work if big data ultimately becomes smart data.

By using AI it is possible to make machines capable of learning – so they will be able to optimize production processes on their own. With the growing use of AI the number of automated guided transport vehicles that drive around the production site independently and take on tasks on their own will also increase. Another example based on collected data that is analyzed using AI applications is predictive maintenance. This involves predictive plant or machine maintenance. Potential problems are identified before they even occur. The result: fewer breakdowns and lower costs.

Energy Management

Renewable energies ?already account for a significant share of electricity consumption in Germany. However, power generation and distribution alone are only the first step. Artificial intelligence can help to generate, store and distribute electricity efficiently. AI-based applications can be used, for example, to optimize energy management in buildings by taking weather forecasts, consumption data and energy prices into account.

The expansion of renewable energies goes hand in hand with technological advancements. Microgrids and smart grids take advantage of a decentralized energy supply. Algorithms determine forecasts for feed-in, storage and consumption. Smart meters provide consumers with support when it comes to to evaluation and optimization in their own homes. Another advantage is that energy suppliers can offer more individualized tariff models based on consumption data.

Artificial intelligence is also a crucial technology in the convergence of the different energy sectors – in other words?integrated energies . At the same time, predictive infrastructure maintenance is being pursued. As in the case of Industry 4.0 the aim is to use pattern recognition in order to identify problems before they even arise. Again: a precautionary check is cheaper than a grid failure. Energy networks that are controlled with the support of artificial intelligence are also able to respond to incidents much faster and thus contribute to network stability.

Where do standards come into play for artificial intelligence?

AI today is reminiscent of the situation when it came to electricity at the end of the 19th century: a lot of innovative but incompatible products, accidents caused by insufficient safety standards, and a combination of hype and skepticism in science, industry, and society. Back then the enormous positive potential of electricity was leveraged, among other things, by standardizing transmission networks, plug systems, safety requirements, switchgear, etc. In similar fashion AI today can make a lot of headway through standardization.

Here are some examples of aspects of AI for which standardization would be helpful and is already being prepared in part:

• Training methods: The most common AI systems at the moment are based on deep neural networks. They are trained with existing data in order to learn how to draw the right conclusions from future data. Neither training data sets nor training methods have been standardized thus far so that it is hardly possible, for example, to merge AI systems and training data across company boundaries.
• Description of self-learning systems: The characteristic feature of many AI systems is that they are able to learn independently through ever new input data while changing their behavior from day to day. But how can several such systems be reliably combined into a larger whole (e.g., a production plant or an operational robot) if their properties are not stable? A practical description of self-learning systems and their properties (e.g. safety, reliability, etc.) is an important task for standardization for just this reason, among others.
• AI methods and scientific and technical terminology: In order to be able to precisely discuss such a complex and interdisciplinary field like AI in expert circles, it must be ensured that there is a common understanding of the most important AI methods and technical terms. For this reason, an international standardization committee with the participation of VDE DKE is already developing corresponding definitions.

Standardization of test methods for AI systems

Effective test methods have been developed and standardized for conventional products over many years in order to test all of the relevant operating and overload scenarios. Since it is not possible to test all possible cases with a finite amount of effort, testers use their engineering knowledge in order to understand how the product works and to select "critical" scenarios accordingly.

This approach has hardly been possible for common AI systems up to now. Deep neural networks, in particular, are a "black box" whose internal "wiring" created during training is beyond analysis. Therefore, it is difficult to make the right selection of test scenarios. New test procedures must therefore be developed and standardized so that AI systems can also be certified and provided with seals of approval.

Another challenge is that AI systems are often self-learning. An AI system may have passed a test today, but may behave differently tomorrow, thus rendering the test result invalid. Therefore, new test procedures must be developed and standardized that describe, for example, continuous testing of the system and/or version control.

Adaptation of previous standards

The body of national, European and international standards for products of all kinds that has been developed in countless expert committees over decades is in principle technology-neutral, i.e. it describes the behavior of products and not their technical implementation. Nevertheless, the emergence of AI as a new technology results in a need for clarification or refinement of numerous existing standards.

One example is the IEC 61508 standard for the functional safety of all systems with electrical components and similar standards derived from it in areas such as industry and mobility. IEC 61508 assumes predictable system behavior, but this is precisely what can no longer be assumed when AI is in use. In this case the tried and tested family of standards must be further developed while at the same time, for example, formulating requirements for the division of responsibilities and/or the interaction between conventional components and those supported by AI.

A conventionally operating "safety kernel" which provides AI with the freedom to play to its strengths in optimizing system behavior, but which at the same time sets demonstrably fixed limits, is being discussed particularly for safety-critical applications such as robotics, for example.

European and international efforts

CEN and CENELEC has established a focus group for artificial intelligence at the European level.?The focus group provides support to CEN and CENELEC in investigating the need for European standardization for AI within CEN-CENELEC and reports to the meetings of the Technical Boards of CEN and CENELEC. It is chaired by Dr. Sebastian Hallensleben and Patrick Bezombes, Vice Director of the Joint Centre for Concepts, Doctrine and Experimentation, from France.

At the international level the Market Strategy Board of IEC (International Electrotechnical Commission) plays a significant role in the context of AI standardization.?Also, the?SEG 10 ?Assessment Group Standardization for Autonomous and AI-based Applications has been established in order to advance the topic of artificial intelligence at the international level. It is chaired by Dr. Sebastian Hallensleben and Tangli Liu from China, Chief Strategic Planning Officer and Director of Smart Cities Department at the China Electronics Standardization Institute.