acceptodds
Under review as a conference paper at ICLR 2027

Bridging the Sim-To-Real Gap with a Robust Deep Learning Framework for Extracting Rare Transient Anomalies in Complex Stochastic Time-Series

Abstract

Detecting rare, morphologically diverse transient anomalies in data-scarce, highly stochastic time series remains a fundamental challenge in representation learning. Real-world temporal data often suffers from extreme class imbalance, variable sequence lengths, and significant domain shifts when relying on simulated training data. To address these challenges, we introduce a novel framework that leverages a Conditional Variational Autoencoder (CVAE) paired with a Power Spectral Density (PSD)-matched colored-noise padding strategy. This approach effectively bridges the simulation-to-real domain gap by synthesizing robust training representations that preserve underlying spectral characteristics. Furthermore, by treating temporal sequences as 1-dimensional images, our architecture employs a Convolutional Neural Network (CNN) to extract invariant features across multiple temporal windows via targeted trending and detrending mechanisms. We validate our framework on a highly complex real-world testbed: detecting anomalous X-ray time series events from Active Galactic Nuclei (AGN) driven by supermassive black holes. Our results demonstrate that this representation engineering approach significantly improves anomaly detection performance and generalization on highly stochastic datasets compared to standard sequence modeling baselines.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Signal Processing, Time-Series, Astrophysics, Synthetic Datasets, Sim-to-Real, Anomaly Detection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VzbJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the difficulty of detecting transient anomalies, such as bursts, in the presence of background noise and dynamically changing conditions. It proposes an anomaly detection framework that combines power spectral density (PSD)-matched colored-noise padding with conditional variational autoencoder (CVAE)-based data augmentation. The framework achieves strong performance on X-ray time-series signals from active galactic nuclei (AGN).

Critical strengths
  1. The paper explores a new application setting for anomaly detection: X-ray time-series signals from AGN, which present more challenging detection conditions.
  2. The paper proposes a novel anomaly detection method that combines a convolutional neural network (CNN) with CVAE-based data augmentation and achieves promising results in this setting.
Critical weaknesses
  1. Several design choices lack sufficient theoretical justification or experimental validation, including the use of a CNN, the choice of a CVAE, and the need for data augmentation.
  2. The paper lacks comparisons with existing baseline methods.
Questions for authors
  1. What motivates the use of a CNN, a CVAE, and PSD-matched colored-noise padding? Please provide theoretical justification or experimental evidence to support these design choices.
  2. How do existing state-of-the-art anomaly detection methods perform on this dataset?
1 comment

Dear Reviewer VzbJ,

Thank you for constructive feedback on our paper. Here are our responses to your questions/comments/feedback:

Response to Critical Weakness (1): We used a 1D-CNN since 1D-CNNs are effective in capturing location invariant features in time-series. This is also the simplest deep learning based method that can be used for this purpose, and the usage of other methods such as transformer based models is left as future work. The goal of the CVAE is to do data augmentation, and the objective of data augmentation is to supplement the (very small) real training set in training the CNN for the final prediction (Please see paragraphs 1 and 5 in the main paper). By nature, the data generated by our CVAE should therefore be in-distribution to the real dataset. Our framework/pipeline incorporates known physics into the simulated (hand-crafted) time-series samples by injecting known forms of noise and known forms of mathematical stochastic burst profiles, and then employs a CVAE to make the transition from mathematically engineered time-series to continuous/seamless time-series that are more in-distribution. One of the novelties of this work is displaying the very fact that CVAEs can make this transition from mathematically engineered signals to real signals (for the astrophysical domain). Further work on (other domains) will support our hypothesis. Please see the discussion in Section 4.3 in the main paper for this and Table 7 for experimental evidence in Supplementary Materials in support of this. The goal of using PSD matched colored noise is explained in paragraph 3 of Section 3.1. In the time-series without stochastic bursts, the PSD assumes a single value of alpha to be governing the noise in the time-series signals, and extends the signal using the computed value of alpha for each signal. In the time-series with stochastic bursts, the PSD assumes the same - that the stochastic bursts on top of the background noise are modeled by a single value of alpha, and similarly extend the signal using the computed value of alpha for each signal. Experimental evidence, i.e. using methods other than PSD to extend the signal while inferencing on the test set showed a marked degradation in performance (signals without stochastic bursts extended by other methods were erroneously inferenced by the CNN to contain stochastic bursts). In other words, there were more false positives.

Response to Critical Weakness (2): A table (Table 2 in main paper) showing the zero-shot performance of (1) Classical Non Deep-learning based methods, and (2) Pre-trained Foundation Models on our WAX-XMM test set has been added to the main paper (rebuttal submission). Please note that all of these models did not require us to retrain them with our synthetic/generated datasets or any other datasets.

We request you to re-evaluate your rating for the paper based on the above-mentioned comments and the changes made to the paper (i.e. rebuttal submission).

Thank you for your continued evaluation of our paper.

Thank you.

Regards,

Authors

Reviewer_eF1FOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a novel task: screening active galactic nucleus (AGN) X-ray light curves for candidate flares. Unlike conventional time-series anomaly detection (TSAD), the task classifies each observation as containing or not containing a candidate flare, rather than localizing anomalies at individual time points or intervals. To handle variable-length observations, the authors estimate the power spectral density (PSD) slope and use the resulting colored noise to pad each light curve to 1,500 points. They then train a conditional variational autoencoder (cVAE) on either simulated data or a mixture of simulated and labeled real data to generate training sequences. Finally, they add random-walk noise to the generated sequences, extract 12 features spanning the original signal, trend, detrended signal, and gradient, and train a CNN to determine whether each sequence contains a candidate flare.

Critical strengths
  1. The paper addresses a relatively novel task with a clear application, Prioritizing AGN observations that may contain candidate flares amid strong stochastic variability has practical value.
  2. The authors address limited training data through cVAE-based generation and multiple augmentation strategies, Beyond symmetric Gaussian flares, the authors consider asymmetric, oscillatory, broad, and narrow flares. They also compare univariate and multiscale inputs, as well as different ways of incorporating real data into training.
Critical weaknesses
  1. Limited discussion of related work and baselines. The paper does not adequately situate the proposed approach within the existing literature or compare it with representative baselines. This makes it difficult to assess its relative merits.
  2. The practical significance of the task needs clarification. The authors should explain how candidate-flare screening would support AGN observation analysis and what sequence-level screening offers beyond existing approaches.
  3. The connection between the stated challenges and the method is unclear. The paper highlights stochastic, non-stationary backgrounds and the rarity of anomalous events, yet uses a relatively simple CNN after data augmentation. The authors should explain how the augmentation addresses these challenges and why this architecture is sufficient.
  4. The evaluation of generated data is insufficient. cVAE-based data generation is a key component of the method, but its quality is assessed mainly through t-SNE and UMAP visualizations. Quantitative measures and a clearly defined evaluation protocol would provide stronger evidence of both data fidelity and its contribution to classification performance.
  5. Potential distribution shift between training and test data requires discussion. Because the datasets come from different sources, the authors should examine whether distribution shift affects performance and clarify how the reported results demonstrate the method’s usefulness on real observations.
  6. The purpose of the XMM-Newton application is unclear. The authors should state what applying the trained CNN to XMM-Newton observations is intended to validate and what additional evidence this experiment provides for the paper’s conclusions.
Questions for authors

Please respond to the above weaknesses in your rebuttal.

1 comment

Dear Reviewer eF1F,

Thank you for constructive feedback on our paper. Here are our responses to your questions/comments/feedback:

Response to Critical Weakness (1): A table (Table 2 in main paper) showing the zero-shot performance of (1) Classical Non Deep-learning based methods, and (2) Pre-trained Foundation Models on our WAX test set has been added to the main paper. Please note that all of these models did not require us to retrain them with our synthetic/generated/or any other datasets.

Response to Critical Weakness (2): Candidate level screening shortlists samples for review by experts for spectral analysis (this is already mentioned in the paper) and for computation for area under the flares. Both of these will be helpful in deciphering/understanding the physical processes that drive the flares (i.e. the mechanism responsible for the flares in AGNs). Traditional point-wise or threshold-based detection methods evaluate time bins independently, failing to capture the dynamic interplay between stochastic backgrounds and complex, variable-length morphological context. In contrast, our sequence-level screening framework processes entire light curves holistically, using a CVAE+CNN pipeline. By evaluating the sequence as a whole, our approach effectively captures long-range temporal correlations, multi-scale trends, and subtle transient evolutions against non-stationary background noise without requiring a rigid, predefined template.

Response to Critical Weakness (3): The goal of data augmentation is to supplement the (very small) real training set in training the CNN for the final prediction (Please see paragraphs 1 and 5 in the main paper). By nature, the data generated by our CVAE should therefore be in-distribution to the real dataset. Our framework/pipeline incorporates known physics into the simulated (hand-crafted) time-series samples by injecting known forms of noise and known forms of mathematical stochastic burst profiles, and then employs a CVAE to make the transition from mathematically engineered time-series to continuous/seamless time-series that are more in-distribution. One of the novelties of this work is displaying the very fact that CVAEs can make this transition from mathematically engineered signals to real signals (for the astrophysical domain). Further work on (other domains) will support our hypothesis.

Response to Critical Weakness (4): Please see Section 4.3 on why t-SNE and U-MAP visualizations address the “in-distribution”-ness of generated data with the real data in conjunction with Table 9 in Section E of the Supplementary Materials.

Response to Critical Weakness (5): The datasets come from the same types of astrophysical sources (active galaxies with a central black hole accreting rapidly) and hence essentially the nature of the X-ray emission (the time series we worked on) are the same. Also, we have used one single instrument (XMM-Newton telescope’s EPIC-pn) for all observations to maintain uniformity in observations. The catalog names for different samples of sources (here AGN) are different because the catalogs were devised to study certain aspects of these sources. The XMM DR is the complete data release of the XMM detected AGN. On the other hand, the WAX sample is a smaller sample of AGN where the outflows from supermassive black holes can be studied well, because they are brighter and nearby sources. But the underlying physics behind the emission of X-rays from all these sources from both of these catalogs are the same.

Response to Critical Weakness (6): It is still an open question as to how an AGN corona (the X-ray emitting region right next to the black hole) is energetically sustained. What pumps in such huge amounts of energy and which mechanisms dissipate is completely unknown. AGN Flares are the most crucial places where we can get these answers. However, it is hard to understand if there is a flare or not. Hence our tool can scan through thousands of time series and identify time series with flares. As noted, the flares in the X-ray emission are indicators of violent emission, possibly due to magnetic reconnection or accretion disk instabilities, or severe turbulence in the region in the vicinity of the black hole. Once identified, the X-ray flares will be studied in detail by X-ray astronomers using ‘Time resolved spectroscopy’. This means they will extract the X-ray spectra (in the energy band 0.2-10 keV) for the time-periods during the flare, and not a flare and carry out detailed study to find the differences in the spectra in these two flux states. These studies will reveal the astrophysical processes that lead to flares.

We request you to re-evaluate your rating for the paper based on the above-mentioned comments and the changes made to the paper (i.e. rebuttal submission).

Thank you for your continued evaluation of our paper.

Thank you.

Regards,

Authors

Reviewer_96KhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework to detect rare, morphologically diverse transient anomalies (specifically X-ray bursts in Active Galactic Nuclei) in highly stochastic time series. It combines Power Spectral Density (PSD)-matched noise padding, Conditional Variational Autoencoder (CVAE) data augmentation, and a 1D CNN with multi-scale trending/detrending features. The pipeline is trained on simulated data and progressively adapted to real data, ultimately identifying burst candidates.

Critical strengths
  1. Domain-Specific Utility: The PSD-matched noise padding is a physically motivated solution to the physical problem.
  2. The overall pipline is reasonable.
Critical weaknesses
  1. The presentation is highly domain-specific and reads more like an astrophysics application paper than a core machine learning contribution. The pipeline is described as a sequence of heuristic and simple engineering steps rather than a new learning framework.
  2. Lack of Methodological Generalization: The proposed framework is essentially a highly customized, heuristic pipeline (manual noise injection, CVAE augmentation, hand-crafted multi-scale detrending) tailored specifically to AGN X-ray light curves. There is no evidence that this "representation engineering" approach generalizes to other domains (e.g., ECG, seismic), despite the authors' claims in the conclusion. It lacks the algorithmic novelty and generalizability required for ICLR.
  3. Absence of Modern Time-Series Baselines: The evaluation relies entirely on a custom CNN and basic simulations. The paper completely ignores modern, state-of-the-art time-series anomaly detection and foundation models combined with proposed augmentation (e.g., TimesNet, or patch-based Transformers). Without comparing against established deep learning baselines for time-series classification/detection, it is impossible to assess the true efficacy of the proposed CVAE+CNN pipeline.
  4. Limited Methodological Novelty: the PSD is domain-specific and manually designed.
Questions for authors

See Weaknesses

2 comments

Dear Reviewer 96Kh,

Thank you for constructive feedback on our paper. Here are our responses to your questions/comments/feedback:

Response to Critical Weakness (1) : Our core (novel) contribution in this work is simulation of the real datasets using functional approximation of the baseline and anomalous signatures/profiles of the time-series samples from a given domain followed by CVAE based distribution shift of these simulated profiles to match the real profiles. Even though our framework uses existing models/methods (such as CNNs and CVAEs), it does so in a novel way for data scarce domains that can be used to train the models with infinitely many in-distribution samples. As described in the next point, we address domain transfer of our method to three other domains through a discussion of these new domains, the corresponding datasets, and how our pipeline can be adapted for these domains.

Response to Critical Weakness (2): In the Rebuttal Submission, we have provided a discussion (Section H) in the Supplementary Materials section on the generalizability of our method to three other domains (and corresponding datasets): (1) Seismic Activity (Earthquake dataset), (2) Electricity Load Profile in Smart-home Electricity consumption (The HouseTwenty dataset), (3) Motion and Gait Kinematics (Toe Segmentation1 dataset), all of which can be downloaded from the UCR Time-Series Classification archive (https://www.timeseriesclassification.com/dataset.php). Please note that the actual execution of our method using the proposed methods is left as further work for interested researchers.

Response to Critical Weakness (3): A table (Table 2 in main paper) showing the zero-shot performance of (1) Classical Non Deep-learning based methods, and (2) Pre-trained Foundation Models on our WAX test set has been added to the main paper. Please note that all of these models did not require us to retrain them with our synthetic/generated/or any other datasets. We did not use the TimesNet or path-based transformer models for two reasons: (1) One of the core contributions of our paper is bridging the gap between simulated/synthetic and real datasets using CVAEs. Here, we first generate synthetic data using heuristics/prior knowledge of known anomaly profiles specific to our domain, and then use a CVAE to create “more realistic” signals. Training the CNN with these models is intended to show that even simple models such as CNN can do a remarkable job at identifying candidate signals with anomalies. More complex models (such as TimesNet and Patch-based transformers) trained on our novel method of generating training data may or may not perform better on our dataset, but that is a diversion from the core contribution and the main narrative of our work. (2) Additionally, training these models with our datasets in our entire pipeline from scratch is time-consuming, and would not be feasible in the limited amount of time provided for the rebuttal.

Response to Critical Weakness (4): The main methodological novelty of our paper is the creation of our synthetic dataset specific to our domain (astrophysical light curves) and the ability to use it with a CVAE to create more “real” data. We also provide a discussion (as mentioned in the response to Question 2) in the Supplementary Materials section to be able to generalize our method to other datasets in other domains (Section H). The PSD Matched Colored Noise is a novel method we propose to extend time-series signals with the computed color of the noise (assuming the entire signal in noise), the goal of which is to ensure all signals have identical length when being used to train or inference using a CNN.

We request you to re-evaluate your rating for the paper based on the above-mentioned comments and the changes made to the paper (i.e. rebuttal submission).

Thank you for your continued evaluation of our paper.

Thank you.

Regards,

Authors

Thanks for your rebuttal. I still remain concerns on the ML Methodological Novelty. I keep my original score and suggest the authors try conferences or journals related to application ML or physical sciences.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.