Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Autoencoder-based framework for anomaly detection in stellar spectra: application to the MaNGA Stellar Library

Akihiro Suzuki

Abstract

A machine-learning-based method is developed to identify objects with unusual stellar spectra. The method employs an autoencoder, a neural network trained to compress spectral data into a low-dimensional representation and subsequently reconstruct it. Spectra that deviate significantly from the dominant patterns in the training dataset are identified using the reconstruction error as an anomaly score. The models are applied to selected datasets from the MaNGA Stellar Library, an empirical library of stellar spectra. Several spectra are flagged as anomalous: an object with likely instrumental and/or reduction issues, two carbon stars, and an oxygen-rich thermally pulsating asymptotic giant branch star. The sources of the large reconstruction errors are examined, and the effectiveness and limitations of autoencoder-based approaches for detecting anomalous stellar spectra are discussed.

Autoencoder-based framework for anomaly detection in stellar spectra: application to the MaNGA Stellar Library

Abstract

A machine-learning-based method is developed to identify objects with unusual stellar spectra. The method employs an autoencoder, a neural network trained to compress spectral data into a low-dimensional representation and subsequently reconstruct it. Spectra that deviate significantly from the dominant patterns in the training dataset are identified using the reconstruction error as an anomaly score. The models are applied to selected datasets from the MaNGA Stellar Library, an empirical library of stellar spectra. Several spectra are flagged as anomalous: an object with likely instrumental and/or reduction issues, two carbon stars, and an oxygen-rich thermally pulsating asymptotic giant branch star. The sources of the large reconstruction errors are examined, and the effectiveness and limitations of autoencoder-based approaches for detecting anomalous stellar spectra are discussed.
Paper Structure (16 sections, 2 equations, 13 figures, 2 tables)

This paper contains 16 sections, 2 equations, 13 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Training data and preprocessing
  3. Data selection
  4. Preprocessing
  5. Autoencoder architecture, training, and application
  6. Network architecture
  7. Implementation and training
  8. Application to the large sample
  9. Unsuccessfully reconstructed cases
  10. Object with possible artifacts
  11. Carbon star
  12. Object with MaNGA ID 3-115120061
  13. Object with MaNGA ID 7-17219806
  14. Properties of two carbon stars
  15. Oxygen-rich thermally pulsating AGB star
  16. ...and 1 more sections

Figures (13)

  • Figure 1: Color-magnitude diagram (absolute $G$-band magnitude vs BP-RP color) for the samples used in this study. The samples from small and large datasets are plotted as small diamonds and circles. Three objects flagged as outliers by the autoencoder models (see Section \ref{['sec:unsuccess']}) are also plotted as a big circle, square, and star as indicated in the lower panel. In the upper panel, the distribution of the BP-RP color for the large dataset is also presented as a histogram. Alt text: Scatter plot showing the distribution of sample data and a hitogram on top of it.
  • Figure 2: Schematic representation of the autoencoder model in this work. Alt text: Schematic illustration showing the model.
  • Figure 3: Training (upper) and validation (lower) losses as a function of the training epochs. In each panel, all the ten autoencoder models are presented. Alt text: Tow plots with ten lines each.
  • Figure 4: Reconstruction error distributions for all the ten models. In each panel, the distribution is plotted as a histogram. The 1st, 10th, 50th, 90th, and 99th percentiles of the distributions are indicated by vertical dashes lines (from left to right). Alt text: ten histograms showing distribution of errors in ten different models.
  • Figure 5: Resoncstructed stellar spectra for a randomly selected object from the large sample. Alt text: Ten plots showing the model and observed spectra wit thin and shick lines and residuapl plotted on top of each panel.
  • ...and 8 more figures