Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Image-based Novel Fault Detection with Deep Learning Classifiers using Hierarchical Labels

Nurettin Sergin, Jiayu Huang, Tzyy-Shuh Chang, Hao Yan

TL;DR

This work proposes a methodology on how, when available, labels regarding the fault taxonomy can be used to increase unknown fault detection performance without sacrificing model performance, and demonstrates increased detection performance on novel fault detection in inspection images from the hot steel rolling process.

Abstract

One important characteristic of modern fault classification systems is the ability to flag the system when faced with previously unseen fault types. This work considers the unknown fault detection capabilities of deep neural network-based fault classifiers. Specifically, we propose a methodology on how, when available, labels regarding the fault taxonomy can be used to increase unknown fault detection performance without sacrificing model performance. To achieve this, we propose to utilize soft label techniques to improve the state-of-the-art deep novel fault detection techniques during the training process and novel hierarchically consistent detection statistics for online novel fault detection. Finally, we demonstrated increased detection performance on novel fault detection in inspection images from the hot steel rolling process, with results well replicated across multiple scenarios and baseline detection methods.

Image-based Novel Fault Detection with Deep Learning Classifiers using Hierarchical Labels

TL;DR

This work proposes a methodology on how, when available, labels regarding the fault taxonomy can be used to increase unknown fault detection performance without sacrificing model performance, and demonstrates increased detection performance on novel fault detection in inspection images from the hot steel rolling process.

Abstract

One important characteristic of modern fault classification systems is the ability to flag the system when faced with previously unseen fault types. This work considers the unknown fault detection capabilities of deep neural network-based fault classifiers. Specifically, we propose a methodology on how, when available, labels regarding the fault taxonomy can be used to increase unknown fault detection performance without sacrificing model performance. To achieve this, we propose to utilize soft label techniques to improve the state-of-the-art deep novel fault detection techniques during the training process and novel hierarchically consistent detection statistics for online novel fault detection. Finally, we demonstrated increased detection performance on novel fault detection in inspection images from the hot steel rolling process, with results well replicated across multiple scenarios and baseline detection methods.
Paper Structure (37 sections, 4 theorems, 20 equations, 14 figures, 3 tables)

This paper contains 37 sections, 4 theorems, 20 equations, 14 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Out-of-distribution Detection
  4. Classifier-based models
  5. Generative models
  6. Hierarchical Classification
  7. Proposed Methodology
  8. Formulation of baseline OOD detection scoring methods used in this study
  9. Maximum Softmax Probability
  10. Out-of-DIstribution detector for Neural networks (ODIN)
  11. Deep Mahalanobis Detection
  12. Hierarchical Regularization in Hierarchical Classification
  13. Proposed Methods
  14. Proposed Hierarchically Consistent Score Function
  15. Insights to improve upon the state-of-the-art OOD Methods
  16. ...and 22 more sections

Key Result

Proposition 1

If we define $s(f_{k})=-\sum_{k=1}^{K}l_{k}^{\text{soft}}(\hat{y})\log f_{k}(\bm{x},\bm{\theta}^{*})$ as a function of $f_{k}$, where $f_{k}$ is the output of the Softmax layer, which satisfies $\sum_{k}f_{k}=1$. The hierarchically consistent score $s(f_{k})$ will be minimized if and only if $f_{k}=

Figures (14)

  • Figure 1: Graphical illustration of the hierarchy of defects in the hot steel rolling dataset.
  • Figure 2: Methodology Flowchart
  • Figure 3: Graphical illustration of the soft labeling logic. A hypothetical two-level fault taxonomy is given. The distances show how the least common ancestor-based distances manifest themselves for this structure. Using these distances and \ref{['eq:soft-label']}, and taking $\beta=5$, we obtain the soft labels in the second row under the leaf labels, given the real label is $L_{11}$. For comparison, one-hot labeling is also shown for the same case.
  • Figure 4: ResNet18 Architecture HeZRS16.
  • Figure 5: Example flowchart for critical value
  • ...and 9 more figures

Theorems & Definitions (8)

  • Proposition 1
  • Proposition 2
  • Proposition 3
  • Remark 1
  • Remark 2
  • Proposition 4
  • proof
  • proof