Building Manufacturing Deep Learning Models with Minimal and Imbalanced Training Data Using Domain Adaptation and Data Augmentation (2306.00202v1)

Abstract: Deep learning (DL) techniques are highly effective for defect detection from images. Training DL classification models, however, requires vast amounts of labeled data which is often expensive to collect. In many cases, not only the available training data is limited but may also imbalanced. In this paper, we propose a novel domain adaptation (DA) approach to address the problem of labeled training data scarcity for a target learning task by transferring knowledge gained from an existing source dataset used for a similar learning task. Our approach works for scenarios where the source dataset and the dataset available for the target learning task have same or different feature spaces. We combine our DA approach with an autoencoder-based data augmentation approach to address the problem of imbalanced target datasets. We evaluate our combined approach using image data for wafer defect prediction. The experiments show its superior performance against other algorithms when the number of labeled samples in the target dataset is significantly small and the target dataset is imbalanced.

• The paper introduces a novel DL pipeline that combines domain adaptation with autoencoder-based data augmentation to address limited and imbalanced manufacturing data.
• It employs an adversarial domain adaptation model to transfer knowledge from a data-rich source domain to a sparse target domain, enhancing defect detection accuracy.
• Experimental results on wafer manufacturing datasets confirm that the proposed approach outperforms traditional models under challenging data conditions.

Insights into Domain Adaptation and Data Augmentation in Manufacturing Deep Learning Models

The paper presents a sophisticated approach to address the limitations associated with training deep learning (DL) models in the manufacturing domain, particularly under conditions of minimal and imbalanced training data. By integrating domain adaptation (DA) techniques with data augmentation strategies, the authors strive to enhance defect detection models that rely heavily on substantial labeled datasets.

Problem Context and Methodology

Manufacturing processes, such as wafer defect detection, traditionally rely on manual inspections, which are both time-intensive and costly. The advent of DL models offers a promising alternative; however, these models require extensive labeled datasets which are often difficult to obtain. Moreover, these datasets are frequently not only limited in size but also exhibit class imbalances, presenting significant challenges to developing effective models.

To address these challenges, the authors introduce a novel DA approach, which adapts knowledge gained from a source dataset to a target dataset with potentially different feature spaces. This method is coupled with an autoencoder-based data augmentation technique aimed at balancing the target dataset. Specifically, the proposed framework initially utilizes DA to transfer knowledge from a source domain with an abundance of data to a target domain with sparse data. Subsequently, it employs an autoencoder to generate synthetic data to bolster minority classes within the target dataset. This combination serves to mitigate the sim-to-real problem commonly encountered with synthetic datasets.

Experimental Evaluation

The efficacy of the proposed approach is experimentally validated using datasets from wafer manufacturing. The experiments demonstrate improved performance on wafer defect prediction tasks, showing superiority over existing algorithms, particularly when faced with small and imbalanced labeled samples in the target dataset. The experimental results underscore the method's potential for generating highly accurate DL models while relying on limited target data.

Key Contributions

The paper delineates several significant contributions:

1. DL Pipeline for Data Scarcity and Imbalance: The researchers propose a DL pipeline designed to overcome the limitations posed by limited and imbalanced datasets by leveraging DA and data augmentation through autoencoders.
2. Adversarial Domain Adaptation Model: An innovative adversarial DA model is developed, enabling effective knowledge transfer even when source and target datasets possess different dimensional feature spaces.
3. Empirical Evaluation: The efficacy of the pipeline is comprehensively evaluated against benchmark datasets, highlighting its performance in harsh conditions where traditional models often falter.

Implications and Future Directions

The implications of this research are twofold. Practically, it provides a viable solution for enhancing DL models within the manufacturing sector, a domain where labeled data scarcity and imbalance are prevalent. Theoretically, it expands the understanding of how DA and data augmentation can be synergistically applied to improve model generalization.

Moving forward, the approach suggests promising directions for further research. One potential avenue involves exploring multi-source domain adaptation, which could allow aggregation of knowledge from multiple related source domains. Additionally, the challenges associated with unlabeled or partially labeled datasets in the target domain could be further addressed, potentially by incorporating semi-supervised or unsupervised learning techniques.

Overall, the work offers an insightful and effective framework for leveraging DL in domains with data constraints, supporting the advancement of AI applications in manufacturing and beyond.