THANK YOU FOR SUBSCRIBING
Machine learning: A benefit for productions
By Marco Huber, Head of Center for Cyber Cognitive Intelligence, Fraunhofer IPA
The topic of learning machines is present in numerous research agendas for a little while. Beginnings of the technology even date back to the 1950s, when among other things the Turing test was developed. Only a few years ago, the technical prerequisites for bringing ML into industrial applications became available: The digitization and growing networks of production machines in the context of industry 4.0 combined with powerful sensors and high computing capacity enabled processing and evaluation of large amounts of data.
Methodical procedures for complex structures
ML methods are based on structures or procedures with which software can gain “knowledge” from data and derive actions from it. This learning process usually requires a large amount of so-called training data or examples. To save resources and to speed up learning, simulation environments are used, which allows faster implementation of the real application.
ML is particularly suitable when the relationships between cause and effect are difficult or impossible to describe analytically or are even unknown. Then automatically analyzing the data can help to recognize patterns. ML also helps when optimizations by using physical models would be too time-consuming.
Optimizing productions on various levels
Integration possibilities of ML in productions are multifaceted and offer companies numerous advantages throughout entire production processes. Starting with production planning, the construction of plants can run more automated. Commissioning within a few days– which is certainly a vision–would only be possible with the help of artificial intelligence.
Once productions are running, the quality of products gains increased importance. Today, production managers get information about the quality of the products and thus, about the entire production process only when they see the production results. Knowing this information in advance and being able to change parameters directly when necessary would be more effective.
We are developing the software for bin-picking; while new objects were so far entered manually using CAD data, the software shall teach itself the handling of unknown or complex objects using ML methods
Another topic for ML methods is predictive maintenance. To date, maintenance is usually preventive–which means too early–or reactive–which means too late. However, with the right data analysis, companies can predict when a particular part might fail and replace it at the right time thereby reducing the downtime of production.
Besides, completely new business models based on the available data will be feasible. A well-known example is selling a service instead of a machine. The customer rents the machine and pays for the usage. However, this only works if the manufacturer can guarantee a certain service level, for example, 99 per cent availability. Data analysis and ML help suppliers to predict that the machine will work as required.
Latest research activities
The most dominant research topic in ML is deep learning, i.e., deep neural networks. These methods can be used to solve complex problems that were previously thought to be reserved for humans like the game of Go. Reinforcement learning is important for planning actions. Transferred to production, one can imagine a robot that no longer needs to be programmed, but continuously learns and improves from its actions. This is similar to the “trial and error” learning known from children.
Another topic is meta-learning, i.e., learning how to learn. This principle of human learning aims at the most efficient use of data. An example of this is the transfer of what has already been learned to similar tasks.
Last but not the least, the interpretability or comprehensibility of the results in ML is a further field of research. Most ML methods are a black box: One enters data into the algorithm and obtains results, but it is not comprehensible how the result is achieved.
In the near future, further benefits for productions can be expected from the basic and application-oriented research in ML. Politics supports the research as well: For instance, the German Federal Government will invest three billion Euros by 2025 to make the slogan “AI made in Germany” an international trademark. Similar initiatives are going on in other countries.
In addition to productions, industrial robotics is also a major recipient of ML’s increased research activities. ML is more and more used for the intuitive instruction of robots. Programming times of industrial robots are being shortened by the use of automatic trajectory planners and program generators. AI can also improve robot performance. Accuracy increase, speed optimization, or service life improvement can be achieved by typical methods such as genetic algorithms or reinforcement learning.
Additionally, image processing is a key technology for identifying and locating objects or for capturing environments. AI and ML algorithms are now widely used to optimize recognition performance. The Fraunhofer Institute for Manufacturing Engineering and Automation IPA is further developing the software for bin-picking. While new objects were so far entered manually using CAD data, the software shall teach itself the handling of unknown or complex objects using ML methods. In addition, even noisy or incomplete sensor data should lead to reliable gripping hypotheses. Gripping strategies for parts, for example, with a risk of jamming, should be learned as well.
These are all advances that expand the application area of industrial robots in production and thus contribute to further increase the degree of automation in production.