Towards Accurate and Robust Classification in Continuously Transitioning Industrial Sprays with Mixup

Hongjiang Li, Huanyi Shui, Alemayehu Admasu, Praveen Narayanan, Devesh UpadhyayarXiv 2022 · DOI · OpenReview

Abstract

Automating the classification of fuel injector spray morphology is critical for optimizing internal combustion engine performance and reducing emissions. However, deep learning models in this domain face significant challenges due to limited experimental data and the continuous nature of spray transitions, which lead to ambiguous class boundaries. This paper investigates the application of Mixup—a data augmentation technique that creates synthetic samples through linear interpolation—to bridge the gap between discrete labels and physical reality. By leveraging transfer learning with a ResNet50 backbone and Mixup, researchers achieved a classification accuracy of over 98% on both internal and large-scale independent test sets. The findings suggest that Mixup acts as a form of physical modeling, enabling neural networks to better generalize across the morphological continuum of industrial sprays.

Introduction

In internal combustion engine design, the spatial distribution and morphology of fuel sprays are primary determinants of combustion efficiency and pollutant formation. Engineers typically categorize these sprays into distinct regimes: "no collapse," where individual plumes are clearly separated; "transitional"; and "collapse," where plumes merge into a single central jet.

The morphological continuum of fuel injector sprays, illustrating the transition from distinct plumes (no collapse) to a merged jet (collapse).

Automating this classification using Deep Convolutional Neural Networks (DCNNs) is desirable for high-throughput analysis but is hindered by two "real-world" constraints. First, scientific data is scarce; high-fidelity spray experiments are costly, resulting in datasets far smaller than standard computer vision benchmarks. Second, spray morphology does not exist in discrete states but rather as a continuum. This inherent fluidity makes labeling difficult even for experts, leading to "label corruption" where boundary cases are assigned arbitrary discrete labels. Standard models often overfit these ambiguous samples, attempting to memorize noise rather than learning the underlying physical transition.

Methodology

The study employed transfer learning, utilizing the ResNet50 architecture as a feature extractor. The model was fine-tuned on a proprietary dataset of 878 Mie-Scattering (laser-based) spray images from Ford’s internal laboratories. The data was categorized into four classes: Pre/Post-injection, No collapse, Transitional, and Collapse.

To address data scarcity and the continuum problem, the authors implemented Mixup. Unlike traditional augmentations such as rotation or flipping, Mixup generates new training samples by blending pairs of images ($x_1, x_2$) and their corresponding labels ($y_1, y_2$) using a mixing weight $\lambda$:

  • $\tilde{x} = \lambda x_1 + (1 - \lambda)x_2$
  • $\tilde{y} = \lambda y_1 + (1 - \lambda)y_2$
Visual examples of Mixup augmentation at different alpha values, showing how synthetic images are generated by blending existing samples.

This approach provides a specific physical interpretation in the context of spray dynamics. By producing "soft" labels—such as a sample that is 60% "transitional" and 40% "collapse"—the model is encouraged to learn the transition between regimes. This prevents the network from becoming overconfident on boundary samples and improves its ability to generalize to unseen experimental conditions.

Results and Discussion

The integration of Mixup yielded significant performance gains. While the baseline ResNet50 model achieved a respectable 94% accuracy, the addition of Mixup (with an optimal $\alpha$ parameter of 0.2) increased test accuracy to 98%.

The model's robustness was further validated on an independent "real-world" test set comprising 7,200 images. On this larger dataset, the model maintained an accuracy of 98.7%. Notably, the system achieved 100% recall for the "no collapse" regime, which is the most critical state for engineers seeking to optimize spray patterns for lower emissions.

A key insight from the study is that while traditional augmentations (e.g., shifting, rotation) provide some benefit, excessive stacking of these methods can degrade performance when combined with Mixup. This suggests that for scientific imagery, augmentations must respect the underlying physics; for instance, vertical flips are physically nonsensical for a downward-firing injector and may introduce misleading features.

Conclusion

This work demonstrates that mathematical data augmentation techniques like Mixup can effectively mirror the physical transitions of industrial processes. By treating class boundaries as a continuum rather than hard divisions, the researchers developed a model that is both highly accurate and robust to the noise inherent in experimental labeling. Future research will focus on integrating high-fidelity Computational Fluid Dynamics (CFD) data to further ground the model in physical truth and exploring advanced interpolation methods to refine the classification of complex transitional states.

Summary updated

Abstract

Image classification with deep neural networks has seen a surge of technological breakthroughs with promising applications in areas such as face recognition, medical imaging, and autonomous driving. In engineering problems, however, such as high-speed imaging of engine fuel injector sprays or body paint sprays, deep neural networks face a fundamental challenge related to the availability of adequate and diverse data. Typically, only thousands or sometimes even hundreds of samples are available for training. In addition, the transition between different spray classes is a continuum and requires a high level of domain expertise to label the images accurately. In this work, we used Mixup as an approach to systematically deal with the data scarcity and ambiguous class boundaries found in industrial spray applications. We show that data augmentation can mitigate the over-fitting problem of large neural networks on small data sets, to a certain level, but cannot fundamentally resolve the issue. We discuss how a convex linear interpolation of different classes naturally aligns with the continuous transition between different classes in our application. Our experiments demonstrate Mixup as a simple yet effective method to train an accurate and robust deep neural network classifier with only a few hundred samples.
BibTeX
@misc{li2022towards,
  author = {Hongjiang Li and Huanyi Shui and Alemayehu Admasu and Praveen Narayanan and Devesh Upadhyay},
  title = {{Towards Accurate and Robust Classification in Continuously Transitioning Industrial Sprays with Mixup}},
  year = {2022},
  eprint = {2207.09609},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2207.09609},
}
Figure