Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Autoencoder-based time series anomaly detection for ATLAS Liquid Argon calorimeter data quality monitoring

Vilius Čepaitis

TL;DR

Problem: Detect rare and unseen anomalies in ATLAS Liquid Argon calorimeter data during data quality monitoring. Approach: an unsupervised LSTM autoencoder trained on good-quality time-series data reconstructs input; anomalies are flagged when the reconstruction loss exceeds a threshold, $L_{th}=1.3$, with loss defined as $L = (1/N) * \sum_{i=1}^{N} (X_i - g(f(X_i)))^2$. Data: 16-dimensional time series derived from topo-cluster $Q$-factor and timing $\tau$, aggregated across four EM partitions and normalized. Findings: validation on noise-burst data shows the AE separates noisy events from good data and can detect additional anomalous windows beyond a reference detector, with evidence of generalization to cosmic and heavy-ion data. Significance: demonstrates a practical, automated early-warning capability for LAr calorimeter DQM and potential applicability to transient calorimeter issues.

Abstract

The ATLAS experiment at the LHC employs comprehensive data quality monitoring procedures to ensure high-quality physics data. This contribution presents a long short-term memory autoencoder-based algorithm for detecting anomalies in ATLAS Liquid Argon calorimeter data, represented as multidimensional time series of statistical moments of energy cluster properties. Trained on good-quality data, the model identifies anomalous intervals. Validation is performed using a known short-term issue of noise bursts, and the potential for broader application to transient calorimeter issues is discussed.

Autoencoder-based time series anomaly detection for ATLAS Liquid Argon calorimeter data quality monitoring

TL;DR

Problem: Detect rare and unseen anomalies in ATLAS Liquid Argon calorimeter data during data quality monitoring. Approach: an unsupervised LSTM autoencoder trained on good-quality time-series data reconstructs input; anomalies are flagged when the reconstruction loss exceeds a threshold, , with loss defined as . Data: 16-dimensional time series derived from topo-cluster -factor and timing , aggregated across four EM partitions and normalized. Findings: validation on noise-burst data shows the AE separates noisy events from good data and can detect additional anomalous windows beyond a reference detector, with evidence of generalization to cosmic and heavy-ion data. Significance: demonstrates a practical, automated early-warning capability for LAr calorimeter DQM and potential applicability to transient calorimeter issues.

Abstract

The ATLAS experiment at the LHC employs comprehensive data quality monitoring procedures to ensure high-quality physics data. This contribution presents a long short-term memory autoencoder-based algorithm for detecting anomalies in ATLAS Liquid Argon calorimeter data, represented as multidimensional time series of statistical moments of energy cluster properties. Trained on good-quality data, the model identifies anomalous intervals. Validation is performed using a known short-term issue of noise bursts, and the potential for broader application to transient calorimeter issues is discussed.

Paper Structure

This paper contains 7 sections, 1 equation, 4 figures.

Table of Contents

  1. Introduction
  2. Methodology
  3. Liquid argon calorimeter inputs
  4. Anomaly detection algorithm
  5. Results
  6. Conclusion
  7. Funding information

Figures (4)

  • Figure 1: A sketch of the data processing pipeline
  • Figure 2: The LSTM autoencoder architecture.
  • Figure 3: Illustrations of LAr noise bursts and their reconstruction loss distributions.
  • Figure 4: Original (blue) and AE-reconstructed (cyan) time series for calorimeter partitions (top 16 rows). The bottom three rows show the reconstruction loss, $L$, events with $L > 1.3$, and noisy event labels from a reference algorithm. An anomalous window near event index 900 is flagged by the AE but not by the reference algorithm. Figure taken from Ref. ATL-DAPR-PUB-2024-002.