Koopman-theoretic Approach for Identification of Exogenous Anomalies in Nonstationary Time-series Data

Alex Mallen, Christoph A. Keller, J. Nathan KutzarXiv 2022 · OpenReview

Abstract

In many scenarios, it is necessary to monitor a complex system via a time-series of observations and determine when anomalous exogenous events have occurred so that relevant actions can be taken. Determining whether current observations are abnormal is challenging. It requires learning an extrapolative probabilistic model of the dynamics from historical data, and using a limited number of current observations to make a classification. We leverage recent advances in long-term probabilistic forecasting, namely Deep Probabilistic Koopman (DPK), to build a general method for classifying anomalies in multi-dimensional time-series data. We also show how to utilize models with domain knowledge of the dynamics to reduce type I and type II error. We demonstrate our proposed method on the important real-world task of global atmospheric pollution monitoring, integrating it with NASA's Global Earth System Model. The system successfully detects localized anomalies in air quality due to events such as COVID-19 lockdowns and wildfires.

Introduction

Detecting anomalies in real-world time-series data is a fundamental challenge in data science, particularly when systems are nonstationary. In domains like atmospheric chemistry, variables such as nitrogen dioxide ($NO_2$) levels do not remain constant; they are subject to complex diurnal, weekly, and seasonal cycles. Traditional anomaly detection methods often fail to distinguish between these expected fluctuations and genuine exogenous shocks—events caused by external factors rather than internal system failures.

To address this, we present a framework that utilizes Deep Probabilistic Koopman (DPK) to model "business as usual" (BAU) dynamics. By establishing a robust counterfactual—what the system should look like in the absence of an anomaly—we can accurately identify deviations caused by events such as the COVID-19 lockdowns or extreme wildfires.

Methodology

Deep Probabilistic Koopman (DPK)

The core of our approach is the DPK model, which bridges Koopman theory and deep learning. Koopman theory suggests that non-linear dynamical systems can be represented in a high-dimensional space where their evolution is linear.

  1. Linear Representation: DPK maps observations into this latent space, expressing the time-series as a combination of periodic sinusoids (capturing daily, weekly, and annual cycles) and a neural network ($g_\Theta$) that parameterizes a probability distribution.
  2. Probabilistic Forecasting: Unlike deterministic models that provide a single point prediction, DPK generates a probability distribution. This allows for the calculation of a z-score (standardized deviation) for new observations, providing a statistical basis for anomaly detection.
  3. Multidimensional Scaling: To monitor regional or national trends, we employ Principal Component Analysis (PCA) to reduce dimensionality, followed by the calculation of the Mahalanobis distance ($Z_{R,t}$). This metric identifies when a collective group of sensors deviates significantly from the expected norm.

Domain Knowledge Integration

A significant innovation in this work is the integration of NASA’s GEOS-CF (Global Earth System Model). While DPK learns from historical patterns, GEOS-CF provides weather-aware predictions. By comparing the two, the system can "explain away" deviations caused by known meteorological factors. For instance, if $NO_2$ levels drop due to a predicted storm, GEOS-CF accounts for this, preventing a false anomaly flag (Type I error). Conversely, if levels remain low during a period where weather suggests they should be high (e.g., during a lockdown), the system correctly identifies the anomaly (Type II error reduction).

Experimental Validation

Datasets

We validated the framework using ground-based $NO_2$ observations from the OpenAQ platform and the European Environment Agency (EEA). The models were trained on data from 2018–2019 and evaluated against the highly anomalous year of 2020.

Key Results

  • COVID-19 Impact: The system successfully detected the onset of lockdowns in major global hubs, including Milan, Madrid, and various cities in China. In China, the model captured the return to baseline activity by May 2020, while European cities showed sustained anomalies.
  • Wildfire Detection: In Seattle, the system identified a massive $NO_2$ spike caused by wildfires. By integrating GEOS-CF, the model could distinguish this exogenous event from standard urban pollution patterns.
  • Statistical Reliability: During "normal" pre-lockdown periods, the generated p-values were uniformly distributed. This calibration confirms that the model does not produce excessive false positives under standard operating conditions.

Discussion and Future Work

While the DPK framework is highly effective, it faces challenges such as "overconfidence" in long-term forecasting, where the model underestimates variance. We addressed this through hyperparameter tuning of the sensitivity ($\alpha$ level). Additionally, the model currently requires approximately two years of historical data to accurately capture annual cycles.

Future research will focus on Automated Subset Selection. Rather than manually defining regions (e.g., "Spain" or "Wuhan"), we aim to develop methods that allow the system to self-discover the geographic boundaries of anomalous events, enabling more granular and automated global monitoring.

Summary updated

Abstract

In many scenarios, it is necessary to monitor a complex system via a time-series of observations and determine when anomalous exogenous events have occurred so that relevant actions can be taken. Determining whether current observations are abnormal is challenging. It requires learning an extrapolative probabilistic model of the dynamics from historical data, and using a limited number of current observations to make a classification. We leverage recent advances in long-term probabilistic forecasting, namely {\em Deep Probabilistic Koopman}, to build a general method for classifying anomalies in multi-dimensional time-series data. We also show how to utilize models with domain knowledge of the dynamics to reduce type I and type II error. We demonstrate our proposed method on the important real-world task of global atmospheric pollution monitoring, integrating it with NASA's Global Earth System Model. The system successfully detects localized anomalies in air quality due to events such as COVID-19 lockdowns and wildfires.
BibTeX
@misc{mallen2022koopman,
  author = {Alex Mallen and Christoph A. Keller and J. Nathan Kutz},
  title = {{Koopman-theoretic Approach for Identification of Exogenous Anomalies in Nonstationary Time-series Data}},
  year = {2022},
  eprint = {2209.08618},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2209.08618},
}
Figure