Improving THT-AOI Image Classification through Federated Learning: A Study on Model Performance and Training Stability under Various Data Distributions

Rachinger B, Thielen N, Meier S, Franke J, Risch F (2025)

Publication Type: Conference contribution, Conference Contribution

Publication year: 2025

Publisher: IPC International

City/Town: Bannockburn, Illinois

Pages Range: 762-774

Conference Proceedings Title: IPC APEX EXPO 2025 Technical Conference Proceedings

Event location: Anaheim, Kalifornien

Abstract

Automated optical inspection (AOI) systems are commonly applied for in-line quality control of through hole technology (THT) solder joints. However, these systems usually rely on conventional image processing, often leading to high pseudo error rates and extensive workloads during subsequent manual inspection. Deep learning approaches have shown promise in reducing pseudo errors by classifying solder joint images flagged as defective by AOI systems into two classes: real defects and pseudo errors. Yet, their performance is limited by the availability of sufficient local training data, and while cross-company data sharing could address this issue, it remains constrained due to privacy concerns.

Federated learning (FL) offers a promising solution by enabling collaborative model training across companies without the need to share raw data, thus preserving data privacy. This study evaluates the application of FL for THT-AOI image classification, specifically addressing common FL challenges such as divergence of local model updates and performance decay under non-independent and identically distributed (non-IID) client data. Therefore, local, centralized and federated learning are compared based on real and artificially separated data under IID and non-IID conditions.

Our results demonstrate that convolutional neural network (CNN)-based FL can be effectively applied for THT-AOI image classification, achieving training stability and model performance comparable to centralized learning. By implementing FL, we reduce AOI pseudo errors by an average of 23.6% on an aggregated test set containing heterogeneous data from multiple electronics manufacturers, significantly lowering the need for manual inspection. This represents a 12.1% relative improvement in pseudo error reduction over local learning, with both approaches maintaining a controlled error slip rate of 1.0%.

Authors with CRIS profile

How to cite

APA:

Rachinger, B., Thielen, N., Meier, S., Franke, J., & Risch, F. (2025). Improving THT-AOI Image Classification through Federated Learning: A Study on Model Performance and Training Stability under Various Data Distributions. In IPC APEX EXPO 2025 Technical Conference Proceedings (pp. 762-774). Anaheim, Kalifornien, US: Bannockburn, Illinois: IPC International.

MLA:

Rachinger, Ben, et al. "Improving THT-AOI Image Classification through Federated Learning: A Study on Model Performance and Training Stability under Various Data Distributions." Proceedings of the IPC APEX EXPO 2025, Anaheim, Kalifornien Bannockburn, Illinois: IPC International, 2025. 762-774.

BibTeX: Download