Artificial intelligence (AI) processes are gaining a foothold on the factory floor. As is also the case with any other tool, a decision has to be made as to whether or not and when it makes sense to use AI – says Professor Marco Huber from the Fraunhofer Institute for Production Technology and Automation IPA. In an exclusive interview with Schall Trade Fairs, he explains that there are many good reasons why humans and AI systems can and should work hand in hand. And he describes a practical approach for small and mid-sized companies who are getting started with AI.
Professor Huber, people say that artificial intelligence (AI) is gaining a foothold in industrial manufacturing – there’s also discussion about machine learning and deep learning. For the purpose of clarification, please give us your definitions and explain any delimitations.
Unfortunately, there’s no unequivocal definition of the term AI. For me, artificial intelligence is about solving problems that require humans to act intelligently. In addition to machine learning which is currently the most active branch in the field of AI, planning and search algorithms, as well as logic and robotics, play a role as well. Machine learning takes advantage of patterns in data in order to make data-based decisions. The term “learning” makes reference to the recognition of these patterns and their transformation into a statistical model, which can then be applied to new data. There are many different types of models within this context. Artificial neuronal networks are currently very popular, whose mode of operation is modelled on human nerve cells in the brain (neurons). If these networks consist of a very large number of layers, that is if they’re deep, we speak of “deep learning”.
AI has been a research topic for several decades, but it has only really gathered momentum in recent years due to the availability of massive computing power, big data and intelligent algorithms. AI offers many opportunities, but the transition from research results to actual practice is often difficult for companies due to a lack of knowledge or resources. Another challenge is the “black box character” of many of the methods based on machine learning. This means that even experts are not always able to understand exactly why a given result has been obtained. This may lead to legal problems, as well as to a lack of trust and acceptance on the part of the users. This is why we’re intensively researching methods for explaining AI processes (xAI, explainable AI).
In which areas of production are AI processes already being used? Where does AI make good sense on the factory floor – and where is it entirely inappropriate?
Quality inspection by means of camera systems is one of the classic applications in the field of artificial intelligence. Camera-supported systems supplement humans as quality inspectors, or take on this job entirely by themselves if an automated solution offers sufficient added value. For example, where damage to a vehicle’s paint needs to be detected or weld seams have to be inspected. To an ever-greater extent, image data is evaluated using algorithms that have their origins in the field of artificial intelligence.
Another field of applications involves predictive maintenance. The idea is to know in advance when the service life of a particular production machine will come to an end, so that the machine can fully exploited before it fails. At the point in time of failure, the repair crew and the required spare parts are already on hand. Special algorithms from the field of AI can also be used to this end. During the course of a scientific study at the Fraunhofer IPA, we were able to prove that the AI approach works better in the field of maintenance than previously used heuristics and mathematical optimisation methods. Production output was increased because there were fewer unplanned stoppages. And production costs were lower because there were no so-called fatal failures, and instead only planned failures.
In essence, AI should be regarded as a tool. As is also the case with any other tool, a decision has to be made as to whether or not and when it makes sense to use AI. Machine learning is suitable when the cause-effect relationships are unknown or can’t be described, but sufficient data are otherwise available. For example, if a mathematical model is available, as is the case with many control problems, AI is usually unnecessary or merely a supplement to the existing procedure.
The fact that machine learning relies on data is of course another aspect within the context of whether or not machine learning makes sense in manufacturing facilities. Unfortunately, these data are not available everywhere. For example, this may be the case in the previously mentioned example of quality inspection with cameras in production. The availability of sample data describing possible defects would be helpful, so that they can be reliably identified. But very few defects should actually occur in a well-run production facility. Consequently, we’re working on generating the data artificially in order to train an algorithm that can detect defects of this sort. Or an algorithm is chosen that doesn’t actually detect any defects, but is capable of determining whether or not the machine is still running in its normal state – because there’s lots of data on the normal state in a smoothly running production facility.
Are there any disadvantages and risks? One objection could be that “eventually the system takes over all of the processes, the human being is left out of the loop and loses control!” Are objections of this sort justified?
We are of course aware of these doubts and I think it’s very important to take them seriously and to address the concerns of the users of AI applications. In the short to mid-term, it will certainly come to pass that AI-based technologies will support people, perhaps performing specific subtasks autonomously on their own as well, but humans will still have a pivotal function in the production environment.
Steven Pinker’s famous quote still applies today in the field of AI: “The hard problems are easy and the easy problems are hard.” This means that problems which are easy for humans, such as creativity, manipulation or perception, are very difficult for AI. Other problems such as lifting heavy objects or calculations with large numbers are difficult for humans, but easy for artificial systems. And thus there are many good reasons why humans and AI systems can and should work hand in hand. For example, AI can take on tasks which are extremely monotonous or computationally too intensive for human beings, thus permitting human workers to focus on tasks of a higher order which benefit from their human strengths. In order to achieve this, a people-oriented approach must be taken when developing AI systems, which means that AI systems should be geared towards the needs, the values and the well-being of the persons involved in order to gain trust and acceptance, and to always allow people to remain in control as well. In this way, AI systems become workable for everyone: they take notice of people, understand them, imitate them and actively support them.
What benefits can a manufacturing company reap from the use of AI?
It’s not easy to offer a general answer to this question because the requirements and goals of the individual company have to be known in order to be able to precisely evaluate the advantages provided through the use of AI. In principle, AI, or more precisely the use of machine learning, facilitates a higher degree of autonomy (in the sense of “automating automation”) and in many cases also improves accuracy or solutions competence in the execution of a specific task. In concrete terms this might mean that a company saves money when it plans maintenance on the basis of AI because, as mentioned above, downtime and failures are reduced. If AI is used to optimise production planning and control, knowledge that was previously confined to certain individuals can then be incorporated into the AI system, making it available more quickly and for more extensive use.
AI can thus help to optimise processes and make them more efficient. Where robotics is concerned, AI reduces programming effort because each individual process step no longer has to be programmed manually – instead, the robot system can adjust itself to deviations in the process or new workpieces and adapt its parameters accordingly. This is especially decisive for increasingly personalised production, where previously necessary programming effort can quickly render the use of robots uneconomical. Last but not least, AI can simplify challenging tasks for robots such as bin picking, for example by enabling powerful image recognition, even for reliable detection of reflective or mirrored objects, or by making it possible for the robot system to recognise jammed components and figure out how to disentangle them.
How can small and mid-sized enterprises get started with the practical use of AI?
We’ve already completed hundreds of AI projects with SMEs. Based on this extensive experience we know that there’s usually lots of interest, which sometimes takes on overwhelming proportions. Sometimes, however, there’s a lack of time and/or money, and especially a lack of expertise. Not least of all, there’s often simply a lack of data or meaningful data processing which would make the implementation of machine learning applications possible in the first place, because this branch of AI is based on relatively large amounts of data which should be available or must be generated in real life, or by means of simulation.
My experience indicates that there are four factors which determine whether or not an AI solution will pay off quickly for any given company:
- Start small, think big: look for low-hanging fruits first. This yields quick results and initial findings, and thus provides support for argumentation within the company. Armed with this experience, greater challenges can then be tackled.
- Get an early start: identify meaningful applications for AI as early as possible. Take advantage of short development cycles in order to make rapid progress, and to be able to respond to difficulties in an agile manner. Prototypes are obtained quickly in this way.
- Focus on the benefits: in the final analysis, the success of an AI project depends on its added value. Let the respectively involved department take the lead – not necessarily the IT department.
- Get everyone involved: don’t forget your employees. Convincing them of the benefits of an AI project will prevent unnecessary barriers.
Especially in the event of initial difficulties or any lack of clarity about the advantages and benefits of AI for a company, I recommend seeking expert advice. For example, the AI Progress Centre of the Fraunhofer Institutes IPA and IAO offers support in various formats. They tackle entry-level questions, offer short feasibility studies and facilitate the setting up of demonstrators.
How do you and your institute demonstrate the use of AI in industrial production processes at live trade fairs, for example at the next Motek in Stuttgart in October 2022?
We have a number of demonstrators and exhibits at our institute that we will definitely be presenting at trade fairs this year, and also within the scope of in-house events. Let’s hope that more on-site presence will be possible again as summer approaches ...
Our AI picking exhibit shows how machine learning and simulations significantly improve applications in terms of autonomy and performance. This is demonstrated by our scientists based on the example of a robot picking objects from undefined positions within a bin. An AI-based object position assessment yields reliable and accurate object positions within a few milliseconds. New objects can be taught in quickly and easily with the help of CAD models. The software can also detect the previously mentioned jams and disentangle them, and it can reliably deal with packaging materials as well. The robot had already been extensively trained in simulation, after which the acquired knowledge was transferred to the actual application. The software also generates and evaluates gripping moves based on this knowledge.
NeuroCAD is a software solution for assembly automation. With the help of machine learning methods, it analyses component properties on the basis of which it generates an assessment as to how well a component is suited for assembly automation. Users can upload their STEP files to a website free of charge and find out within seconds how easy or difficult it is to separate any specific component. The tool also evaluates the component’s gripping surfaces and alignability. Furthermore, the neural network indicates how probable it is that the result is correct.
Another demonstrator has arisen from my research group, which addresses numerous sectors and applications. We use it to demonstrate how previously non-transparent AI processes can become comprehensible and explainable. As mentioned already, transparency or the so-called “white box character” of AI applications is becoming increasingly important due to legal requirements and in order to generate trust amongst the users. And because this issue is becoming increasingly relevant, we offer the right solution for all user needs: from the identification of use cases to the creation of explainable AI models and their use in customer applications, right on up to the generation of explanations for previously taught-in AI models and procedures for creating transparency.
Curriculum Vitae – Univ.-Prof. Dr.-Ing. habil. Marco Huber: After studying computer science and successfully completing his doctorate at the University of Karlsruhe (TU), Professor Huber headed the “Variable Image Acquisition and Processing” research group at Fraunhofer IOSB in Karlsruhe from 2009 to 2011. He then worked as a senior researcher at AGT International in Darmstadt, Germany, until 2015. Professor Huber was responsible for product development and data science services for the Katana division at USU Software AG in Karlsruhe from April 2015 until September 2018. At the same time, after successfully completing post-doctoral studies, he taught as a private lecturer in the field of computer science at the Karlsruhe Institute of Technology (KIT). He has held the professorship for cognitive production systems at the University of Stuttgart since October 2018. At the same time, he heads up the departments for Image and Signal Processing, as well as Cyber Cognitive Intelligence (CCI), at the Fraunhofer Institute for Production Technology and Automation IPA in Stuttgart. His research is focused on machine learning, explainable artificial intelligence (xAI), sensor data analysis, image processing and robotics in production environments.