Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unsupervised anomaly detection in MeV ultrafast electron diffraction

Mariana A. Fazio, Salvador Sosa Güitron, Marcus Babzien, Mikhail Fedurin, Junjie Li, Mark Palmer, Sandra S. Biedron, Manel Martinez-Ramon

TL;DR

This paper presents an unsupervised anomaly-detection framework for MeV ultrafast electron diffraction (MUED) images using a convolutional autoencoder (CAE) to reconstruct diffraction tiles and a residual-based detector that models reconstruction errors with a two-distribution likelihood. By estimating the posterior probability $p(N|e)$ of an image being normal and automatically deriving a decision threshold, the method provides both detection and uncertainty guidance without requiring labeled training data. The approach demonstrates high detection performance with low false-positive rates on a substantial MUED dataset and remains robust when the anomaly fraction is small. Practically, this enables automated data quality control in MUED experiments, reducing manual inspection and expediting ultrafast materials research, with computation carried out on a high-performance computing platform.

Abstract

This study focus in the construction of an unsupervised anomaly detection methodology to detect faulty images in MUED. We believe that unsupervised techniques are the best choice for our purposes because the data used to train the detector does not need to be manually labeled, and instead, the machine is intended to detect by itself the anomalies in the dataset, which liberates the user of tedious, time-consuming initial image examination. The structure must, additionally, provide the user with some measure of uncertainty in the detection, so the user can take decisions based on this measure.

Unsupervised anomaly detection in MeV ultrafast electron diffraction

TL;DR

This paper presents an unsupervised anomaly-detection framework for MeV ultrafast electron diffraction (MUED) images using a convolutional autoencoder (CAE) to reconstruct diffraction tiles and a residual-based detector that models reconstruction errors with a two-distribution likelihood. By estimating the posterior probability of an image being normal and automatically deriving a decision threshold, the method provides both detection and uncertainty guidance without requiring labeled training data. The approach demonstrates high detection performance with low false-positive rates on a substantial MUED dataset and remains robust when the anomaly fraction is small. Practically, this enables automated data quality control in MUED experiments, reducing manual inspection and expediting ultrafast materials research, with computation carried out on a high-performance computing platform.

Abstract

This study focus in the construction of an unsupervised anomaly detection methodology to detect faulty images in MUED. We believe that unsupervised techniques are the best choice for our purposes because the data used to train the detector does not need to be manually labeled, and instead, the machine is intended to detect by itself the anomalies in the dataset, which liberates the user of tedious, time-consuming initial image examination. The structure must, additionally, provide the user with some measure of uncertainty in the detection, so the user can take decisions based on this measure.

Paper Structure

This paper contains 13 sections, 9 equations, 9 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Experimental Methods
  3. Computational Methods
  4. Data pre-processing procedure
  5. Convolutional autoencoder for reconstruction
  6. Anomaly detection
  7. Computational resources
  8. Experiments
  9. Training and test of the autoencoder
  10. Statistical analysis of the reconstruction error
  11. Estimation of the detection threshold
  12. Experiment with reduced number of anomalies
  13. Conclusions

Figures (9)

  • Figure 1: Schematic representation of the experimental setup for the MUED instrument located in the Accelerator Test Facility at Brookhaven National Laboratory.
  • Figure 2: Single shot diffraction patterns obtained for Ta$_2$NiSe$_5$: (a) typical diffraction pattern and (b) - (d) anomalous patterns.
  • Figure 3: Structure of the used convolutional encoder. The input is a tile of an image, with dimensions $80\times 80$ for the present application, and the output is the reconstructed image contained in the tile.
  • Figure 4: Structure of the detector.
  • Figure 5: Comparison between 10 original and reconstructed test tiles of an image and their corresponding reconstruction error. The graphs show the logarithm of the amplitudes and the squared error.
  • ...and 4 more figures