Creating reliable and robust quality control systems that identify process errors while having a low number of false rejects is a considerable challenge in the automated manufacturing industry. Especially in the pharmaceutical industry, where a product's quality has to be ensured at all costs, a large amount of false rejects is acceptable to guarantee the integrity of all released products. As standard quality control systems mainly perform a binary classification, most of them do not provide insights about the reason behind rejections. As a result, the underlying reason for the rejects, such as degradation in equipment or wrong settings in process parameters, often goes unnoticed. Yet, these systems are based on conservative approaches that incorporate the uncertainties related to the measurement system and process variation such as batch-to-batch variations and assembly tolerances. In this contribution, a new data-driven quality control system is suggested. The system is based on well-established machine learning methods that differentiate multiple types of errors in the assembly processes of medical products. Trained on process data, the system's functionality is demonstrated in a pre-study and two real industrial use cases. Moreover, application-specific differences are discussed. It is shown that for the two use cases and a limited number of batches the system not only detects 100% of all defective products but also limits the number of false rejects to an acceptable amount. In all of the application examples, the system has the potential to be executed as a soft real-time system that allows integration into industrial processes. Moreover, it is shown that the algorithm can present the extracted knowledge in various forms understandable for humans, allowing for more informed decision making.

在自动化制造行业中，创建可靠，强大的质量控制系统以识别过程错误，同时减少少量的次品，这是一个巨大的挑战。尤其是在制药业中，必须不惜一切代价确保产品的质量，为保证所有已发布产品的完整性，可以接受大量的次品。由于标准质量控制系统主要执行二进制分类，因此大多数系统都无法提供有关拒收原因的见解。结果，通常不会引起人们对次品的根本原因的关注，例如设备性能下降或工艺参数设置错误。然而，这些系统基于保守的方法，结合了与测量系统和过程变化（如批次之间的变化和装配公差）相关的不确定性。在此贡献中，提出了一种新的数据驱动的质量控制系统。该系统基于完善的机器学习方法，可区分医疗产品组装过程中的多种类型的错误。经过过程数据培训后，该系统的功能在预研究和两个实际工业用例中得到了证明。此外，还讨论了特定于应用程序的差异。结果表明，对于两个用例和有限数量的批次，系统不仅检测出所有有缺陷产品的100％，而且还将错误拒收的数量限制为可接受的数量。在所有应用示例中，该系统具有作为软实时系统执行的潜力，可以集成到工业流程中。而且，表明该算法可以以人类可以理解的各种形式呈现所提取的知识，从而允许进行更明智的决策。