Comparing few-shot learning methods in various application scenarios for quality prediction in ultrasonic wire bonding

Few-shot learning refers to the problem of learning underlying patterns in data from just a few training samples, which contrasts with traditional deep learning methods that usually rely on large datasets. The collection of large datasets is typically costly and time-consuming and often requires significant computational resources. In ultrasonic wire bonding production processes, it is of fundamental importance that a high quality of the joints is ensured, while the application scenario often differs depending on the process, e.g. different materials, process sequences or machine settings. In this paper, we compare various few-shot learning approaches for predicting bond qualities in ultrasonic wire bonding. This is done by quantitatively predicting the shear force of the bonding joint across different application scenarios. The prediction is based on key process variables in the form of time-series data, such as deformation, ultrasonic power, frequency and current. These time-series vary in length depending on the bonding process. The few-shot learning approaches are compared in three application scenarios: Changing the bonding program (application scenario 1), changing the transducer (application scenario 2), and changing the bonding machine (application scenario 3). For example in application scenario 1, with just 15 retraining samples - less than 2.3% of the original training data - the shear force as a quality criterion for a 380 μm aluminum wire can be predicted with high accuracy. Using an autoencoder yields a mean absolute error of 90 cN, while meta-learning improves this to just 81 cN (less than approx. 5% of the average shear force).