[Paper Review] Application of federated learning in manufacturing

Vinit Hegiste, Tatjana Legler+1 more|arXiv (Cornell University)|Aug 9, 2022

Privacy-Preserving Technologies in Data24 references4 citations

TL;DR

This paper proposes federated learning (FL) for collaborative quality inspection in manufacturing, where multiple factories train a shared deep learning model without sharing raw data. By fine-tuning a VGG-19 model with ImageNet weights and aggregating only model weights across clients, the global model achieved high accuracy in classifying defective USB components with 5 communication rounds, outperforming training from random weights and proving effective for privacy-preserving industrial AI.

ABSTRACT

A vast amount of data is created every minute, both in the private sector and industry. Whereas it is often easy to get hold of data in the private entertainment sector, in the industrial production environment it is much more difficult due to laws, preservation of intellectual property, and other factors. However, most machine learning methods require a data source that is sufficient in terms of quantity and quality. A suitable way to bring both requirements together is federated learning where learning progress is aggregated, but everyone remains the owner of their data. Federate learning was first proposed by Google researchers in 2016 and is used for example in the improvement of Google's keyboard Gboard. In contrast to billions of android users, comparable machinery is only used by few companies. This paper examines which other constraints prevail in production and which federated learning approaches can be considered as a result.

Motivation & Objective

• To address data privacy and intellectual property constraints in industrial machine learning by enabling collaborative model training without sharing raw production data.
• To evaluate the feasibility and performance of federated learning in a real-world manufacturing quality inspection scenario involving limited, non-uniform defect data.
• To compare different model initialization strategies—random vs. pre-trained weights—for federated training in a low-data, industrial setting.
• To demonstrate that a globally aggregated model can generalize well across diverse client data distributions in a manufacturing context.
• To validate the effectiveness of federated learning using a real-world dataset of defective USB components across three clients.

Proposed method

• Federated learning was implemented using a central server that aggregates model weights from three client factories, each with its own local dataset of USB component images.
• A VGG-19 architecture was used, with feature extraction via pre-trained ImageNet convolutional layers and fine-tuning of the final classifier head on local defect data.
• Model weights were updated locally on each client’s device and sent to the central server after each communication round (CR), using FedAvg as the aggregation algorithm.
• The global model was reconstructed by averaging the client model weights at each CR, ensuring data privacy since raw images were never transmitted.
• Clients trained their models using local data and evaluated performance on external test sets, including third-party USBs not in any client’s training data.
• The study compared training from random initial weights versus pre-trained ImageNet weights, measuring convergence speed and generalization performance.

Experimental results

Research questions

• RQ1 Can federated learning effectively train a high-accuracy quality inspection model in manufacturing when defect data is scarce and non-uniformly distributed across factories?
• RQ2 How does the use of pre-trained ImageNet weights impact convergence speed and model performance in federated learning for industrial defect detection?
• RQ3 To what extent can a globally aggregated federated model generalize to unseen defect types or third-party products not included in any client’s training data?
• RQ4 How do communication rounds affect the accuracy and robustness of the global model in a federated setup with limited data per client?
• RQ5 Does federated learning outperform centralized training in terms of privacy while maintaining comparable or better generalization in industrial quality inspection?

Key findings

• A fine-tuned VGG-19 model initialized with ImageNet weights achieved high accuracy in defect classification using only 5 communication rounds, significantly outperforming models trained from random weights.
• The global federated model generalized well to third-party USB components not in any client’s training data, correctly classifying hidden defects with high confidence.
• Training from random initial weights required at least 20 communication rounds to show any meaningful accuracy improvement, and even then performance was poor on external data.
• The global model trained via federated learning detected all defects in live classification with slightly higher confidence than a centrally trained VGG model, demonstrating superior generalization.
• Using pre-trained weights reduced the number of required communication rounds and improved convergence, proving critical for practical deployment in industrial settings with limited data.
• The federated approach successfully preserved data privacy, as no raw images were shared, and all training occurred locally on client devices.