Experts say artificial intelligence (AI) can add tremendous value, but automakers should, in particular, look at production and logistics. In these areas, they say, efficiency boosts and cost savings can be realized in a relatively short period of time.
A new McKinsey study calls AI a future value-adding machine for the auto industry. By 2025, automakers around the world are expected to reap $215 billion in value from it. Even in the short-term, savings of $61 billion seem to be possible in production alone. Although these figures are for the global auto industry, they are attracting the attention of manufacturing managers in Germany, a country known for high wages. They have been asking themselves for a long time how they can integrate the technology into operations that have already been analyzed and optimized again and again.
McKinsey partner Matthias Kaesser, co-author of the study, believes the forecast underlines the potential of AI in the automotive sector. Depending on the suitability of specific processes, he said, cost savings of 10 to 15 percent are possible – assuming that the companies engage in the right kind of transformation. Processes with AI systems analyzing and evaluating image and audio data are ideal. One example: A suitably programmed AI could automatically analyze the sound of the cylinders and valves during the first engine test at the end of the assembly line. At an early stage, it would determine whether the sample was a “lemon” with a high risk of problems. That would be predictive maintenance in the broadest sense.
Martin Ruskowski, head of the Innovative Factory Systems research area at the renowned German Research Center for Artificial Intelligence (DFKI) in Kaiserslautern, also believes AI can make a huge contribution to manufacturing efficiency. “The most obvious application of AI in production —which has been the subject of in-depth practical work — is anything targeting production monitoring and predictive maintenance,” he said. The opportunities involve production machines in a narrow sense. Controlling them is yet another option. In each case, deviations from the norm can be identified before equipment exceeds critical limits. Ruskowski sees the use of AI to support workers as another promising field. For example, AI systems could use their capabilities to analyze sequences of movements, take over inputting or verify the results of certain steps. With the help of AI, the system can determine quite easily whether a particular screw is tightened enough.
Of course, the AI system would have to look over the worker’s shoulder with a camera during work. This is one reason why Ruskowski urgently recommends the early integration of works councils and employees when systems are introduced. He assigns the greatest economic potential of AI use to the monitoring of production facilities. If the AI system sends out notifications about looming problems, it can help to avoid stoppages and limit the time spent on repairs with smart methods of failure identification and location.
Ruskowski quickly allayed some fears about the new technology. AI isn’t and doesn’t need to be sorcery, he said. The applications require no especially complex or futuristic technologies at all. Neuronal networks? Machine learning? Deep Learning? They are all important subfields of AI, but simple, affordable
approaches often do the trick in manufacturing. “If you are identifying deviations from the norm, relatively simple statistical models, decision trees or descriptor models usually suffice,” Ruskowski said. A large portion of DFKI’s research deals with apparently simple approaches that theoretically make these areas of AI low-hanging fruit. The applications enable clear economic improvements at a manageable cost.
Still, when it comes to AI, people are mainly talking about its potential, not about what has actually been achieved. Before users can reap the benefits, problems have to be solved and obstacles overcome. Issues that have to be addressed include the above-mentioned integration of employees affected by AI on the organizational level, and the availability of the right data on the technical level. In manufacturing environments, machines are only rarely networked according to an overarching plan. Users are mostly dependent on individually programmed interfaces to feed the key learning data from the machines into the AI systems.
That drives up costs. “Breaking-in an AI system can quickly cost a few hundred thousand euros,” Ruskowski said. “You have to find the right model. You need data experts who understand the data, and you have to program the connection. That is a huge obstacle especially for smaller companies.”
As in Industry 4.0, one goal is to create standardized data access for AI. The communication protocol OPC UA, which is often used in mechanical engineering, is not sufficient. Standards are needed to allow the transmitted data to be interpreted and placed in a meta context, with the goal of making a learning process possible. This process is, in turn, a precondition for successful, productive operations. What’s key is to have a framework that allows the classification of the learned data. Said Ruskowski: “The systems have to know how the data should be interpreted so they can do this.”
By Christoph Hammerschmidt