The use of artificial intelligence (AI) is generally expected to further increase productivity as well as the quality of products while simultaneously reducing the consumption of resources. This creates significant growth potential for AI-based applications.
Significant amounts of data are now becoming available at many (sub-)production steps, which are essential for the application of machine learning methods. Due to the rapid increase in data volume and the greatly increased complexity of production processes, on the other hand, the application of “conventional” analysis and optimization methods is often no longer useful or even possible. And this is precisely where the key properties of AI-based systems come into play: namely, they can learn automatically from the data material already available and also adapt to changing environmental conditions.
Overview of AI methods
Today, the term “artificial intelligence” is predominantly used in the context of machine learning, but also includes a variety of other applications such as expert systems or digital assistants (with which the automated creation of a formalized knowledge representation enables subsequent logical reasoning).
The decisive factor here is the ability to deal successfully with new or unknown situations, to process and “understand” new information and thus to generate new knowledge from the available knowledge, to solve new tasks. It is precisely this ability that distinguishes AI systems from classic rule-based IT systems, which have to be adapted manually every time the task changes.
Currently, machine learning methods are divided into the following three subgroups:
- Supervised learning (Supervised Learning),
- Unsupervised learning,
- Reinforcement learning.
In supervised learning, the AI algorithm is trained with a large set of exemplar input and output data (reference data). A mathematical model is created from the relationship between the input and output data, which also allows prediction of output values given a previously unknown combination of input data. This is called a regression task. Alternatively, the supervised learning algorithm can also realize the assignment of features obtained from the input values to object classes (classification task or pattern recognition). In this case, it is assumed that the relationship between features and object classes is already known. However, if these are not known, the relationship between the data must be established subsequently by so-called “labeling”, which of course can be quite costly.
In unsupervised learning, an exemplary set of data is also required, but without any already given “labels” or predefined object classes. It is learned with the help of suitable algorithms, how the “normal” data looks like. In case of anomalies, these are recognized and separated accordingly. The interpretation of the obtained results lies here however with humans. He has to connect the automatically recognized object classes with the real production context. Thus, despite the saving of the labeling process, unsupervised learning is no less costly due to the necessary interpretation of the results.
Reinforcement learning is based on reward principles, which leads to the adaptation of the solution in terms of optimization. For example, parts of a solution can be evaluated with scores, where a higher score corresponds to a better solution. The greatest challenge here is the identification of suitable reward strategies and the admissibility of the try-and-error principle under concrete production conditions.
Promising application areas
Promising AI application areas for manufacturing SMEs are very diverse and it is simply not possible to create a conclusive list. Therefore, only a rough overview is given within the scope of this article.
When using AI-based predictive models, you can determine exactly what condition your machines are in. This allows you to make timely decisions about which of your assets need maintenance. By adjusting maintenance scheduling, you can avoid breakdowns in your machinery. The basis for these AI algorithms is historical and current machine data.
Process optimization and process control
This is another application field of artificial intelligence in production. Here, manufacturing or process engineering processes are optimized. The focus of optimization here is not only on the machines themselves, but also on the production steps or entire process chains. For example, based on machine and process data, AI models can be used to determine important influencing factors and make optimal process settings. The AI algorithm can thereby evaluate the production process in real time and make the appropriate adjustments as needed. This allows you to minimize both the waste and the energy requirements of your processes.
Here, AI-based processes, such as image recognition or image classification, help you automate the time-consuming and error-prone manual quality controls. After a learning phase, AI can detect the anomalies on images captured by cameras and provide appropriate guidance to the operator. Alternatively, the AI can sort out the defective parts fully automatically.
Product and process development
AI solutions, such as using artificial neural networks to predict the mechanical properties of products, can help you optimize your design and engineering processes. One of the many uses for AI is also the use of machine learning to evaluate test or simulation data to uncover complex relationships in product development via pattern recognition. The use of AI for planning and optimization tasks is also useful to save manual effort and accelerate the development process.
Assistance systems are now an integral and potential component of AI applications. The task of assistance systems is to support humans in a wide range of activities. Although assistance systems are not intended to replace humans, they do provide them with important information and warn them of incorrect interventions. Machine learning and, above all, deep learning open up numerous new possibilities and areas of application for assistance systems – even in increasingly complex application scenarios.
Artificial intelligence can be used to establish causality relationships and optimize production processes in a self-learning manner. This results in self-optimizing machines and systems that can adapt to changing production conditions and aim to optimize machine utilization. The cross-process-step evaluation of numerous data from machine groups, plants and production lines, but also from operating data, enables cost-optimized production through AI.
AI systems are already supporting companies with action planning and optimization algorithms in warehousing. In the process, critical stock levels are identified as well as suggestions for material quantities to be kept in stock are issued on the basis of forecasting models. AI is also increasingly being used in the area of goods transport. Examples include route planning for driverless transport systems and dynamic route adaptation for autonomous vehicles, which allows them to recognize and avoid obstacles, for example.