Back to Repair: A Minimal Denoising Network\ for Time Series Anomaly Detection
| Authors | Kadir-Kaan Özer et al. |
| Year | 2026 |
| HF Upvotes | 2 |
| arXiv | 2604.17388 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
We introduce JuRe (Just Repair), a minimal denoising network for time series anomaly detection that exposes a central finding: architectural complexity is unnecessary when the training objective correctly implements the manifold-projection principle. JuRe consists of a single depthwise-separable convolutional residual block with hidden dimension 128, trained to repair corrupted time series windows and scored at inference by a fixed, parameter-free structural discrepancy function. Despite using no attention, no latent variable, and no adversarial component, JuRe ranks second on the TSB-AD multivariate benchmark (AUC-PR 0.404, 180 series, 17 datasets) and second on the UCR univariate archive by AUC-PR (0.198, 250 series), leading all neural baselines on AUC-PR and VUS-PR. Component ablation on TSB-AD identifies training-time corruption as the dominant factor (ΔAUC-PR = 0.047 on removal), confirming that the denoising objective, not network capacity, drives detection quality. Pairwise Wilcoxon signed-rank tests establish statistical significance against 21 of 25 baselines on TSB-AD. Code is available at the URL https://github.com/iis-esslingen/JuRe.
Engineering Breakdown
Plain English
This paper introduces JuRe, a remarkably simple neural network for detecting anomalies in time series data that challenges the conventional wisdom that complex architectures are necessary. The model uses just a single depthwise-separable convolutional residual block with 128 hidden dimensions, trained to reconstruct corrupted time series windows, and then flags anomalies using a fixed scoring function that requires no learned parameters. Despite this extreme simplicity—no attention mechanisms, no latent variables, no adversarial training—JuRe achieves second-place performance on the TSB-AD benchmark (0.404 AUC-PR across 180 series and 17 datasets) and second on the UCR univariate archive (0.198 AUC-PR across 250 series), outperforming all other neural baselines on both AUC-PR and VUS-PR metrics. The key finding is that when you get the training objective right (manifold-projection principle), architectural bells and whistles become unnecessary.
Core Technical Contribution
The authors demonstrate that the manifold-projection principle—training a network to reconstruct clean data from corrupted inputs—is sufficient for anomaly detection without any of the standard deep learning apparatus. Their key innovation is decoupling the detection mechanism from the learning mechanism: the network learns through a denoising objective, but anomaly scoring happens through a parameter-free structural discrepancy function computed at inference. This exposes a central insight that prior work over-engineered the problem by adding attention, variational components, and adversarial losses when a minimal objective function does the heavy lifting. The contribution is both architectural (showing that depthwise-separable convolutions with one residual block suffice) and methodological (showing that the training objective matters more than model expressiveness).
How It Works
JuRe's pipeline has two distinct phases: training and inference. During training, the network takes normal (clean) time series windows, corrupts them by adding noise or masking, and learns to reconstruct the original signal—this forces the network to learn a compressed representation of the normal data manifold. The architecture consists of a single depthwise-separable convolutional residual block with 128 hidden units; this lightweight structure processes windowed time series and outputs reconstructed values the same shape as the input. At inference, for any new time series window, JuRe passes it through the trained network to get a reconstruction, then computes a fixed structural discrepancy metric (no parameters to learn here) that measures how far the original window deviates from what the network expects. Windows with high discrepancy scores are flagged as anomalies. The key insight is that the network acts as a "denoiser" that implicitly learns the boundary between normal and anomalous behavior, without ever being explicitly told what anomalies look like.
Production Impact
This approach dramatically reduces the operational burden of deploying time series anomaly detection at scale. A production system using JuRe would have minimal memory footprint (a single residual block trained on modest compute), making it practical for edge deployment on IoT devices or embedded systems where transformer-based anomaly detectors would be infeasible. The fixed scoring function means inference is extremely fast and deterministic—no sampling, no attention computation, no forward passes through a large discriminator—making it suitable for real-time streaming anomaly detection with strict latency SLAs. Training data requirements are also lower; since the approach relies on self-supervision (corruption and reconstruction), you don't need labeled anomalies, only access to normal baseline data, which is common in practice. The main trade-off is that this approach assumes a clear separation between the normal data manifold and anomalies; if anomalies fall on or near the normal manifold (e.g., subtle fraud patterns that preserve statistical structure), detection accuracy will degrade, and you may need the additional expressiveness of more complex architectures.
Limitations and When Not to Use This
The paper's strength—simplicity—also limits its scope: the manifold-projection principle assumes anomalies are structurally decodable as "corrupted versions of normal data," which breaks down when anomalies are subtle or semantically similar to normal patterns (e.g., a single degree shift in a temperature reading that doesn't distort the signal shape). The fixed, parameter-free scoring function is rigid and cannot adapt to domain-specific notions of anomaly; if your anomalies have different characteristics across different time series or change distribution over time (data drift), this approach offers no mechanism to recalibrate. The ablation study mentioned in the abstract is incomplete in the provided text, so it's unclear which components (if any) could be further pruned or whether there are failure modes where adding back complexity actually helps. Additionally, performance on the benchmarks is second-place, not first, suggesting room for improvement and raising questions about whether the simplicity comes with a ceiling on achievable accuracy.
Research Context
This work sits at the intersection of self-supervised learning and anomaly detection, building on the long history of reconstruction-based approaches (autoencoders, variational methods) but pushing back against the trend toward ever-larger and more complex architectures in deep learning. It aligns with recent findings in the broader ML community that good inductive biases and objectives can outweigh sheer model capacity, similar to how simple scaling laws and proper training recipes sometimes beat exotic architectures. The paper contributes to the time series anomaly detection benchmark ecosystem—TSB-AD and UCR are widely-used evaluation standards—and its strong results on these benchmarks suggest that practitioners may be over-investing in architectural complexity. The research direction opened is clear: a systematic re-examination of whether prior work in anomaly detection (and potentially other tasks) conflated model expressiveness with detection capability, and whether simpler, more interpretable models can match or exceed complex baselines when the training objective is carefully chosen.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::