Artificial Intelligence in the Factory Big Data in Machine Learning
As a core technology of Industry 4.0, self-learning systems are likely to find their way into factories, especially if they are introduced step-by-step and prove that machine learning can make money.
Machine learning (machine learning, ML) as a sub-area of artificial intelligence (AI) is especially relevant in industrial production. ML enables systems to understand their environment, plan actions, react to obstacles and communicate with people. Machines learn to recognize independently recurring patterns and objects on the basis of operating data and intelligent algorithms. The acquired knowledge can then be applied to unknown and unsorted data. This allows you to identify sources of error, plan and optimize processes and to make forecasts.
The fact that machine learning is currently experiencing a hype, even though the concept actually dates back to the 1980s, is due to the new possibilities of data processing. Only with big data applications, high computing power and huge cloud storage was the right infrastructure created. Up until now, Internet giants have primarily used this infrastructure for their own purposes. But industry is following suit. “With regard to robotics, we are keeping a very close eye on what big players such as Google or Amazon are developing with their IT skills and infrastructures and what they are researching in the area of production technology," confirms Prof. Jörg Krüger, Head of the Automation Technology Business Unit at the the Fraunhofer Institute for Production Systems and Design Technology (IPK), Berlin. But the examples from the IT companies simply cannot be transferred one-on-one to industrial applications.
Many, large companies from the control and automation segment are already infected with the "ML virus". But according to industry experts, the use of machine learning in industrial applications is still in its infancy. A few impressive exceptions cannot disguise this fact, for example, when IBM introduced its Watson system in the Cognitive Factory, or when Festo used fascinating exhibits such as the "elephant's trunk", an intelligent bionic handling assistant, to answer the question of how people in tomorrow's factory can interact with machines simply, efficiently and, above all, safely. The technology is available. It is exciting and stimulates the imagination, but transferring it into real products that promise sales and profit is likely to take years to come.
SMEs and Start-Ups Are on the Move
The fundamental question is whether machine learning is only suitable for global players and their idea of a comprehensive concept of a digital factory, or whether, in addition to a top-down development by financially strong large companies with their competent research and development departments, a breakthrough from flexible, innovative small and medium-sized enterprises would also be conceivable.
“Artificial intelligence is an important topic for the future," says Dr. Wilfried Schäfer, Managing Director of VDW (Association of German Machine Tool Builders) and organizer of EMO Hannover 2017, the world's leading trade fair for metalworking. “Therefore, small and medium-sized producers should also deal with machine learning if they want to be able to derive opportunities for their own development in good time".
For Dr. Cord Winkelmann, Managing Director of Sensosurf, a Bremen-based company, many things are already moving. “Large companies tend to develop their own solutions, often very complex and comprehensive, sometimes spectacular and effectively marketed ones," he notes. “Among them there is a highly active group that wants to get informed, exchange information, network and make a difference. That's where digitization is a matter for the boss."
Innovative start-ups can play their part in the development process. Sensosurf starts with the slogan "Sensor integration meets machine learning". Established in 2016 as a spin-off of the Chair of Microsensors, Actuators and Systems (IMSAS) at the University of Bremen, the company transfers microsystems technology into the harsh conditions of mechanical engineering. Sensosurf integrates sensors directly into standard machine components. These include flange and pillow block bearings, linear guides and threaded rods. “We are dealing with areas from which there has been little or no information available," says Winkelmann. Machine learning is used for data evaluation to exploit machine and processing information.
Large amounts of data are a prerequisite for machine learning, because only with them ML can be realized at all. Winkelmann says that rapid market penetration depends on generating information that pays off from the start. Great things always start with small steps," he explains. This includes data evaluation at the machine, networking of the machines with each other, determination of the characteristic features of a process. “When you first see the data that is acquired, evaluated and visualized, you quickly get used to the new findings and the possibilities they offer," says Winkelmann. Machine manufacturers are convinced by the fact that machines can learn how to protect themselves from operating errors. The data obtained can also be used to fend off unjustified warranty claims.
Fraunhofer expert Prof. Krüger also confirms that "it is important to offer companies ways of introducing machine learning in small digestible steps". At present, he sees the focus of ML for machine tool manufacturers primarily in condition monitoring. In this context, it is essentially a matter of interpreting data automatically by means of pattern recognition technology. The knowledge that is required to determine the condition of a process or a machine can be trained by means of machine learning methods.
Potentials in Energy Management
In addition to predictive maintenance, condition monitoring and quality management, self-learning systems can also advance energy management. At EMO Hannover 2017, the Munich-based company Gerotor presented its high-performance HPS energy storage system for the first time. It is designed to reduce energy and connection costs using intelligent algorithms. The idea for the product has its origins in Formula 1, more precisely from the KERS (Kinetic Energy Recovery System) used there. The system was originally ordered for environmental reasons, as it recovers the energy generated during braking maneuvers by transferring it to the drive axle via a rotating flywheel system.
The Gerotor founders saw great potential in "using this efficient and wear-free technology, not only for cars that drive around in circles", as Gerotor board member Michael Hein put it. In the search for an application that also involves frequent braking and acceleration, the company identified machine tools and tool spindles. In many cases they are slowed down and accelerated every second. The advantages of digitizing and networking the energy storage system became obvious: "Whoever is in control of the energy circuit, also has control of the information center.”
Connected directly to the system without requiring its own power supply, the new energy storage unit increases the efficiency of the entire system by means of energy recovery, peak smoothing and digitalization. The system measures all currents and cycles, captures data and information, improves its own algorithms and draws conclusions. While conventional control strategies can achieve energy savings of 10 to 25 % at the most, according to Hein, intelligent strategies are likely to double the savings for users. For Hein, energy management offers a particularly simple and efficient introduction to the ML. “Energy systems must be 100 % forward-looking," he stresses. "We need intelligent control strategies and an infrastructure that readjusts itself."
Return on Investment Is Decisive
However, Hein admits that the concept of machine learning practically does not occur in customer discussions. Rather, the decisive factor is ROI (return on investment): "We sell exclusively on the basis of the argument that we save more than we cost". In fact, this may be one of the reasons why many companies are getting quiet when the question of their ML strategies arises. Machine learning is a means to an end, not a sales argument.
However, there is no blueprint for the introduction of individual strategies. It is advisable to draw on expert knowledge, be it through the various Fraunhofer Institutes or external service providers. As Jörg Krüger explains, each company must first of all clarify what kind of intelligence it requires from a machine, system or robot, such as the determination of the machine's condition, autonomy, and automatic adaptation to changes such as tool wear and tear or component properties. The ability to independently reschedule and self-organize production processes, to understand human commands and gestures for simplified programming are also among the skills that a machine could learn on its own. But Krüger also points out that one question must be answered: "Who checks whether something has been learned correctly before the machine applies the new knowledge automatically?
Questions on IT security and data protection also have to be answered, as well as who is responsible for decisions made by an intelligent system. Could this be another example of "discomfort" in dealing with cognitive systems and possible loss of control? Cord Winkelmann does not think so. A much more serious obstacle to machine learning, as well as to digital transformation in general, is the frequently inadequate provision of high-speed Internet in many places, especially for businesses in rural areas.
This article was first published by Maschinen Markt.