ML is becoming increasingly important in industrial contexts, and especially in quality assurance. This often entails a lot of cost and effort, particularly when strict fault tolerance specifications, which can be in the range of less than 0.1 percent, have to be observed. In such cases, every wrong prediction of the analysis system takes many extra man-hours to remedy.
Suitable ML methods could significantly reduce this added work. In industrial and production environments, however, ML-based process optimizations and decision-making supports are difficult to implement at present. The reason for this is that most well-established processes, which already have low error rates, tend to produce one-sided data material and too little of it.
To address this problem, the “Data Efficient Automated Learning – DEAL” research group at the Center for Applied Research on Supply Chain Services is carrying out research into the possible applications of machine learning in the industrial and manufacturing sector. The group is further developing deep learning methods that facilitate the training of high-quality models even with unbalanced data sets.
A feasibility study on quality assurance for a customer from the automotive sector confirmed that this approach works. The spot weld inspection system on-site, which used conventional image processing methods for its analyses, identified far more errors than actually existed. The DEAL group employed standard deep learning approaches to improve the decision-making. In developing the algorithm, it was vitally important to have the right evaluation strategy. This made it possible to achieve the required scores on key performance indicators as regards quality and computation time.