PaperlayerView exact revision on arXiv

Source-linked AI summary

No Scale Left Behind: Multi-Scale Autoencoder with Bi-directional Attention for Time Series Anomaly Detection

Jiaheng Guo, Haochen Zhang, Yu-Chao Huang, Jinhao Duan, Nicholas Konz, Tianlong Chen

arXiv:2609.38004v1cs.LGcs.AI

TL;DR

Time series anomalies range from brief spikes to gradual drifts, while many detectors use one scale and multi-scale methods often isolate scales or impose a hierarchy. MSCAD connects parallel autoencoder branches with symmetric bidirectional cross-scale attention before reconstruction. On TSB-AD, it reports state-of-the-art performance across variance-aware ranking metrics on both splits, with VUS-PR of 0.57 on UTS and 0.47 on MTS.

  • Problem

    Many TSAD methods use a fixed temporal scale, while multi-scale designs isolate scales or restrict information flow to a coarse-to-fine hierarchy.

    Source: §1 · pp. 1–2

  • Method

    MSCAD uses parallel autoencoder branches at different patch sizes and symmetric bidirectional attention so every scale pair exchanges information before reconstruction.

    Source: §1 · p. 2

  • Results

    Across both TSB-AD splits, MSCAD leads every variance-aware ranking metric; VUS-PR reaches 0.57 on UTS (+0.05) and 0.47 on MTS (+0.04).

    Source: §5 · p. 10

  • Takeaways & Limitations

    A uniform error aggregator matches or outperforms learned input-adaptive gates, suggesting the bridge reconciles temporal semantics before reconstruction.

    Source: §1 · p. 2

    The scoring premise assumes the model reconstructs in-distribution windows faithfully and deviating subsequences poorly.

    Source: §3.1 · p. 4

Abstract

from arXiv · show

Time series anomaly detection (TSAD) plays a crucial role in healthcare, finance, industrial monitoring, and other sectors. Within and between these settings, anomalies span vastly different temporal scales, from sub-second point spikes to multi-hour drift patterns. However, most existing TSAD methods commit to a single temporal granularity, and multi-scale designs either analyze different scales in isolation or are constrained to a predefined coarse-to-fine hierarchy, both failing to sufficiently capture multi-scale interactions. To resolve this limitation, we propose Multi-Scale Autoencoder with Cross-Scale Attention for TSAD (MSCAD), a simple yet powerful semi-supervised TSAD framework founded on parallel autoencoder branches corresponding to different patch sizes. A stack of symmetric bidirectional cross-scale attention blocks enables every pair of scales to exchange information before reconstruction without allowing any single scale to be privileged. On the comprehensive TSB-AD benchmark (40 datasets, 530 series), MSCAD achieves large performance gains against 50 baselines across multiple metrics, with VUS-PR of 0.57(+9.6%) on the univariate split and 0.47(+9.3%) on the multivariate split compared to the state-of-the-art.

1 Introduction

Anomalies span temporal scales, while existing approaches leave cross-scale interaction limited or asymmetric. MSCAD connects parallel autoencoder branches with symmetric bidirectional attention and reports state-of-the-art results on TSB-AD.

1 Introduction

  • Anomalies in critical-sector time series range from physiological spikes and industrial outages to gradual financial drifts, motivating detection across temporal scales.The introduction describes transient, sub-second spikes and persistent, multi-hour regime shifts.Source: §1 · p. 1
  • Single-scale receptive fields trade off slow structural changes against transient events, while existing multi-scale methods isolate scales or impose coarse-to-fine information flow.The latter may be suboptimal when anomaly evidence is scale-local or reciprocal.Source: §1 · pp. 1–2
  • Three parallel autoencoder branches use symmetric bidirectional cross-scale attention so every scale pair exchanges information before reconstruction, without a privileged scale.Per-scale reconstruction errors are fused with a uniform aggregator.Source: §1 · p. 2
  • MSCAD reports state-of-the-art TSB-AD results, including VUS-PR of 0.57 on TSB-AD-U and 0.47 on TSB-AD-M across 50 baselines.Ablations identify symmetric exchange as a key design component: the coarse-to-fine variant reduces the gain to within seed noise of a no-bridge baseline.Source: §1 · p. 3

2 Related Work

Prior TSAD work includes reconstruction-, forecasting-, frequency-, and representation-based methods. Multi-scale approaches commonly process scales independently or route information through an asymmetric hierarchy, unlike MSCAD's symmetric inter-scale flow.

2 Related Work

  • Semi-supervised TSAD work spans reconstruction-based detectors, forecasting- and frequency-based methods, and general or representation-based approaches.The cited examples include autoencoders and Transformers, period- or frequency-based models, and detectors using pretrained or patch-level representations.Source: §2 · p. 3
  • Multi-scale methods commonly use independent branches or asymmetric hierarchies, which either prevent representation-level interaction or privilege a scale.Examples include parallel patch-size branches and coarse-to-fine refinement or directed information flow.Source: §2 · p. 3
  • MSCAD differs by enabling symmetric inter-scale information flow at the representation level, avoiding scale-isolated processing and privileged coarse-to-fine hierarchies..Source: §2 · p. 3

3 Method

MSCAD uses parallel patch-scale autoencoders joined by a symmetric cross-scale bridge, then fuses their reconstruction errors into anomaly scores. Its semi-supervised formulation trains on anomaly-free data and evaluates raw scores at the timestep and channel levels.

3.1 Problem Formulation

  • Under the semi-supervised setup, training data are assumed anomaly-free; unseen test windows receive anomaly scores, with larger scores indicating greater anomaly likelihood.The scoring premise is that the model reconstructs in-distribution windows faithfully but reconstructs deviating subsequences poorly.Source: §3.1 · pp. 3–4

3.2 Framework Overview

  • The framework encodes each window at multiple patch scales, exchanges information through a symmetric bridge, and decodes each scale to produce reconstruction errors for uniform fusion.The fused score is the window-level training loss.Source: §3.2 · p. 4

3.3 Multi-Scale Patch Branches

  • Each scale has its own lightweight autoencoder branch, with patch-specific encoding and decoding to produce a scale-level reconstruction error.The architecture uses patch sizes {4, 16, 64}; the branches do not share parameters.Source: §3.2 · p. 4, §3.3 · p. 5

3.4 Symmetric Cross-Scale Bridge

  • The bridge lets each scale attend to all other scales, with every scale both informing and being informed by the others before decoding.The operation is symmetric and bidirectional, with no privileged scale.Source: §3.4 · p. 5

3.5 Training and Inference

  • Training minimizes the expected window-level score, which reduces to the mean per-scale reconstruction error; inference averages overlapping-window scores over time and channels.The final raw scores are evaluated without thresholding or post-processing.Source: §3.5 · p. 6

4 Experiments and Results

MSCAD is evaluated on TSB-AD and leads representative baselines across reported metrics and splits. Results and ablations highlight the value of symmetric cross-scale exchange, alongside sensitivity to window size and bridge depth.

4.1 Experimental Design

  • TSB-AD covers 40 datasets; detectors train on clean prefixes, score full series, and are evaluated against labeled suffixes under the semi-supervised protocol.The official split contains 350 univariate and 180 multivariate series.Source: §4.1 · p. 6
  • VUS-PR is the primary threshold-free metric; MSCAD is compared with 50 baselines using one fixed configuration across series and no per-file tuning.The paper also reports VUS-ROC, AUC-PR, AUC-ROC, Range-based F1, and Point-F1.Source: §4.1 · pp. 6–7

4.2 Main Results

  • VUS-PR reaches 0.57 on UTS, improving by 9.6% over PaAno, and 0.47 on MTS, improving by 9.3% over PaAno; MSCAD leads on all six reported metrics on both splits.Table 1 compares MSCAD with 18 representative baselines on UTS and 16 on MTS.Source: Table 1 · p. 7, §4.2 · p. 8
  • Across ten duration buckets, MSCAD generally tracks or exceeds the single-scale envelope, with clearest gains at mid-to-long durations; short-duration buckets sometimes favor a single-scale variant.Fine, medium, and coarse single-scale variants perform best in different duration regions.Source: Figure 3 · p. 8, §4.2 · p. 8

4.3 Model Analysis

  • Adding symmetric cross-attention lifts All VUS-PR by +0.017, while the identity bridge improves it by only +0.002 over no bridge, within seed noise.The identity variant retains the FFN+residual wrapper but injects no cross-scale context.Source: Table 2 · p. 9, §4.3 · p. 9
  • Average VUS-PR falls from 0.5356 to 0.5112 (−0.0244) with coarse-to-fine flow; fine-to-coarse reaches 0.5041, near the no-bridge baseline of 0.5051.The directed variants restrict reciprocal exchange between scales.Source: §4.3 · p. 9, Table 3 · p. 9
  • Window size peaks at W=128; narrower windows miss long-context anomalies, while wider windows shift validation data off the per-window normal manifold.Bridge depth saturates at B=2, with depth 3 and 4 both regressing as the bridge over-mixes scales.Source: §4.3 · p. 9
  • The uniform aggregator reaches 0.5356 VUS-PR, versus 0.5346 for the all-feature learned gate and 0.5347 for its random-feature control.Individual feature-drop variants range from 0.5312 to 0.5349.Source: §4.3 · p. 9

5 Conclusion

MSCAD connects parallel per-scale branches through symmetric bidirectional attention, and its TSB-AD results lead across reported variance-aware ranking metrics on both tracks. The appendix describes the benchmark’s scope, imbalance, and dataset diversity.

5 Conclusion

  • 0.57 VUS-PR on UTS and 0.47 on MTS beat prior SOTA by +0.05 and +0.04, respectively (+9.6% and +9.3%).MSCAD also attains state-of-the-art on VUS-ROC, AUC-PR, and AUC-ROC on both splits, using one fixed configuration without per-series search or per-file tuning.Source: §5 · p. 10

A TSB-AD Benchmark Details

  • TSB-AD spans 40 sub-datasets and multiple domains, with point and interval anomalies, univariate series, and multivariate telemetry.The released benchmark has 870 curated univariate and 200 curated multivariate series; the headline evaluation uses 350 and 180 held-out series, respectively.Source: A TSB-AD Benchmark Details · pp. 15–16, Table 4 · p. 16

A.1 Statistics

  • Anomalous points average 2.4% of TSB-AD-U series and 5.1% of TSB-AD-M series, motivating VUS-PR as the primary measure rather than ROC metrics alone.Global anomaly ratios are 2.7% and 4.3%, respectively.Source: §A.1 · p. 16

A.2 Sub-datasets Breakdown

  • The benchmark combines short sequences testing localized detection with long medical and industrial traces that test memory, scale adaptation, and robustness to long normal stretches.Its multivariate track also includes high-dimensional series, such as OPP with 248 channels on average and SWaT with nearly 60.Source: §A.2 · p. 16

A.3 Evaluation Protocol and Tuning Sets

The protocol selects hyperparameters on tuning files, freezes the chosen configuration, and evaluates on separate held-out files. It reports VUS-PR as primary alongside supporting metrics.

A.3 Evaluation Protocol and Tuning Sets

  • Hyperparameters are selected using tuning files, then frozen before final evaluation on separate evaluation files.The headline univariate results use the 350-file evaluation subset, while the multivariate evaluation split contains 180 files.Source: §A.3 · pp. 16–18
  • Table 5’s Eval and Tune columns count univariate files in the evaluation and tuning lists, respectively.The multivariate source breakdown reports average observed channel counts.Source: Table 5 · p. 17, Table 6 · p. 17
  • VUS-PR is the primary ranking criterion; VUS-ROC, AUC-PR, AUC-ROC, and thresholded F1 variants provide supporting evidence.VUS-PR evaluates continuous scores, while F1 variants describe their behavior after thresholding.Source: §A.3 · p. 18

B Evaluation Metrics

The evaluation uses per-series scores and ground-truth labels, emphasizing threshold-independent ranking and VUS-PR. Thresholded F1 measures are secondary, and reporting is aggregated across the evaluation split.

B.1 Score and Label Convention

  • Each series’ real-valued anomaly scores are evaluated against binary labels, with metrics computed per series and averaged over the split.Larger metric values are better, and threshold-dependent F1 variants are treated as diagnostics rather than the primary ranking criterion.Source: B Evaluation Metrics · p. 18, §B.1 · p. 18

B.2 Threshold-Independent Point Metrics

  • AUC-ROC measures the true-positive versus false-positive-rate tradeoff, while AUC-PR is more informative under rare anomalies but remains point-wise.AUC-ROC can be optimistic under severe class imbalance because false-positive rate is normalized by the large number of normal timestamps.Source: §B.2 · pp. 18–19

B.3 Volume-Under-Surface Metrics

  • VUS-PR evaluates precision–recall across score thresholds and temporal tolerance windows, addressing rare anomalies and small boundary mismatches.The official call uses W=100 and approximates curves with 250 thresholds.Source: §B.3 · p. 19

B.4 Threshold-Dependent F1 Variants

  • Thresholded F1 variants differ in whether they score point-wise matches, event detection, or range overlap, and can be sensitive to threshold choice.PA-F1 expands predictions to cover an anomaly interval after any point within it is detected, which can overstate performance for noisy scores.Source: §B.4 · pp. 19–20

B.5 Reporting and Aggregation

  • The main tables rank methods by mean VUS-PR on the held-out split and report other metrics as auxiliary evidence.Coverage, when included, is the fraction of evaluation files with a valid score sequence of the required length.Source: §B.5 · p. 20

C Baselines

  • The baseline set follows TSB-AD and adds recent methods that postdate its original table.For families with per-point and subsequence variants, the “Sub” prefix denotes a sliding-window subsequence wrapper.Source: C Baselines · p. 20

C.1 Statistical and classical methods

The related-work review groups TSAD baselines into classical, time-series-specific shallow, deep reconstruction or prediction, and foundation-model approaches. Their methods differ in how they represent normal behavior and score anomalies.

C.1 Statistical and classical methods

  • Classical methods score points or subsequences using distance, density, isolation, or projection residuals, and can remain competitive on many TSB-AD datasets.Examples include IForest path lengths, LOF local-density ratios, and PCA reconstruction residuals.Source: §C.1 · pp. 20–21

C.2 Time-series-specific shallow methods

  • Time-series-specific shallow methods use subsequence similarity, periodicity, graph structure, or spectral residuals; MatrixProfile flags subsequences with high nearest-neighbor distance.Series2Graph instead uses rare graph nodes or edges and unusual transitions to identify variable-length anomalies.Source: §C.2 · pp. 21–22

C.3 Deep reconstruction and prediction methods

  • Deep detectors model normal behavior through reconstruction, forecasting, or attention, while CrossAD uses a coarse-to-fine decoder that MSCAD’s symmetric cross-scale flow removes.CrossAD’s encoder forbids cross-scale interaction, and its decoder conditions finer scales on coarser ones.Source: §C.3 · pp. 22–23

C.4 Foundation models

  • Time-series foundation models are pretrained on large corpora and applied zero-shot or after task-specific fine-tuning, often scoring forecast or reconstruction residuals.MOMENT, for example, uses frozen reconstruction-error scoring zero-shot or lightweight task-specific fine-tuning.Source: §C.4 · p. 23

D MSCAD Implementation

MSCAD uses parallel patch-scale autoencoder branches with symmetric cross-scale exchange, then evaluates a fixed configuration through a shared TSB-AD pipeline. It wins every reported metric column on both benchmark splits, with low variability across seeds.

D.2 Hyperparameter Configuration

  • The headline configuration uses shared hyperparameters across every TSB-AD series, with no per-file tuning.The reference setup uses symmetric cross-attention, d=256, and B=2.Source: §D.2 · p. 24, Table 7 · p. 24

D.1 Pseudocode

  • Each scale embeds overlapping patches, encodes them independently, exchanges information through synchronous symmetric bridge blocks, and decodes the bridged latents.The bridge uses shared attention projections across scales; training uses validation-loss early stopping.Source: §D.2 · p. 25

D.3 Unified Evaluation Pipeline

  • At test time, channels use training-prefix normalization and the same channel-independent model, with windows of length 128 and stride 2.The official metric routine receives raw scores without thresholding or method-specific post-processing.Source: §D.3 · p. 26, §3 · p. 26

3. Per-timestep score recovery. For timestamp t in channel c, the score s(c)

  • Each timestamp’s channel score is the unweighted mean of scores from its covering windows; multivariate scores are then averaged across channels.Interior timestamps are covered by 64 overlapping windows.Source: §3 · p. 26

D.4 Compute Resources

  • The complete experimental program used approximately 1,200 A6000-equivalent GPU hours, while the headline evaluation took about 25 hours per seed.The headline run used an NVIDIA RTX PRO 6000 Blackwell node.Source: §D.4 · p. 26

E.1 Full baseline comparison on TSB-AD

  • MSCAD wins every metric column on both TSB-AD splits; its VUS-PR margin over PaAno is +0.05 on UTS and +0.04 on MTS.The full comparison covers 39 methods on UTS and 33 on MTS, or 50 distinct methods overall.Source: Table 8 · p. 27, §E.1 · p. 27, Table 9 · p. 28

E.2 Multi-seed results across all eight metrics

  • Per-seed standard deviation is ≤0.0024 on every metric and split, and PA-F1 tracks the same overall ranking as threshold-free VUS-PR.The reported results aggregate seeds {2026, 2027, 2028}.Source: §E.2 · pp. 27–28

E.3 Per-dataset breakdown

The per-dataset evaluation shows strong results on long-context telemetry and robustness across domains, alongside weaknesses on short series with few anomaly events. The appendix also reports compute costs, qualitative scale behavior, and ablations of scale-set and hyperparameter choices.

E.3 Per-dataset breakdown

  • MSCAD reaches VUS-PR ≥0.78 on four univariate industrial datasets and ≥0.60 on three multivariate datasets, and stays within 0.03 VUS-PR of the within-split mean across diverse domains.The reported domains include medical waveforms, spacecraft telemetry, web traffic, and human-activity sensing, using the same fixed hyperparameters.Source: §E.3 · pp. 28–29, Table 10 · p. 29
  • Most per-dataset losses occur on short, near-zero-anomaly-ratio series with few or no events, heavy class imbalance, and limited prefix length.The passage identifies these conditions as challenging for reconstruction-based detectors trained from scratch.Source: §E.3 · p. 28

E.4 Compute and parameter cost

  • The bridge accounts for about 13% of parameters but removing it cuts per-seed wall time by roughly 55%, from 24.8 hours to 11.2 hours.The headline configuration takes 14 seconds per UTS file and 7.8 minutes per MTS file; multivariate overhead scales linearly with channel count.Source: §E.4 · pp. 29–30, Table 12 · p. 30

E.5 Qualitative case studies

  • Representative score visualizations concentrate MSCAD’s raw anomaly scores in labeled regions spanning short point spikes, bounded contextual events, and long regime shifts.The qualitative cases show the small patch localizing point-spike edges and the large patch maintaining a clean baseline through long contextual anomalies.Source: §E.5 · p. 30, Figure 4 · p. 31
  • The per-scale reconstruction errors are complementary: P=4 tracks short structure but flattens drift, P=64 follows long envelopes but smears spikes, and P=16 covers the gap.This is the reported empirical condition under which uniform fusion is competitive.Source: §E.5 · pp. 30–31, Figure 5 · p. 32

F.1 Hyperparameter sensitivity

  • The sensitivity table reports capacity saturation at d=256, depth saturation at B=2, and an inverted-U pattern for window size, with W=128 as the default.Each sensitivity block varies one hyperparameter while holding the others at that block’s default.Source: Table 13 · p. 32

F.2 Scale-set composition

  • The three-scale set {4, 16, 64} is the unique winner across all three reported columns; dropping P=16 costs 0.011 All VUS-PR, while adding P=32 costs 0.003.The added scale increases parameter count without improving performance.Source: §F.2 · pp. 32–33, Table 14 · p. 33

F.3 Fusion-gate ablation

The ablations find that cross-scale exchange is the key component, while learned fusion gates and diversity regularization do not improve on the streamlined default. Cross-attention is the strongest tested bridge operator overall.

F.3 Fusion-gate ablation

  • The uniform aggregator’s All VUS-PR of 0.5356 exceeds every learned gate by 0.001–0.004; the best learned variant reaches 0.5349.Random-feature and trained four-feature gates score 0.5347 and 0.5346, respectively, indicating the gate is essentially uninformed by its conditioning signal.Source: Table 15 · p. 33, §F.3 · p. 33

F.4 Bridge operator

  • Cross-attention wins All VUS-PR at 0.5356 and MTS VUS-PR at 0.4705; every non-trivial bridge operator beats no-bridge by +0.026 to +0.031 All VUS-PR.Cross-attention edges max-pool by +0.004 All VUS-PR and mean-pool by +0.003 All VUS-PR.Source: §F.4 · pp. 33–34, Table 16 · p. 34

F.5 Diversity-loss sweep

  • Adding diversity regularization at λdiv=0.1 costs 0.015 All VUS-PR versus the unregularized default, so the paper omits auxiliary loss terms.The reported drops are −0.030 UTS and −0.015 MTS All VUS-PR.Source: Table 17 · p. 34, §F.5 · p. 34

G Broader Impact

MSCAD targets monitoring settings where anomalies can range from transient physiological events to slow industrial drift. Its semi-supervised and channel-independent design is presented as applicable across heterogeneous domains, while the paper cautions against autonomous high-stakes use.

G Broader Impact

  • Healthcare and industrial monitoring can involve both transient events and slow regime shifts, which MSCAD is designed to handle without privileging one temporal resolution.Examples include physiological signals and infrastructure monitoring such as power grids, water systems, gas pipelines, and manufacturing lines.Source: G Broader Impact · pp. 34–35
  • The semi-supervised protocol fits settings where normal traces are abundant but labeled anomalies are scarce or absent, across telemetry, server health, and intrusion-detection domains.The passage also notes that channel-independent inference scales linearly to multivariate sensor banks without parameter inflation.Source: G Broader Impact · p. 35
  • Incorrect scores can cause harm through delayed intervention or alarm fatigue, and channel-independent scoring may miss relational failures across sensors.The paper recommends decision-support use with calibrated thresholds, domain-specific validation, human escalation, and distribution-shift monitoring.Source: G Broader Impact · p. 35

H Limitations

MSCAD is computationally heavier than several lightweight TSAD baselines, with per-series training and inference costs that may exceed some safety-critical latency budgets.

H Limitations

  • 5.97M parameters in the headline configuration contrasts with < 0.1M for KAN-AD and several classical detectors.The configuration uses d=256, B=2, three Transformer branches, and cross-scale bridges.Source: H Limitations · p. 35
  • Training from scratch on each series’ normal prefix avoids cross-series leakage but incurs per-series training costs from a few minutes to roughly 30 minutes.The reported range runs from short series to long-prefix multivariate series in the authors’ protocol.Source: H Limitations · p. 35
  • Dense sliding-window inference at stride τ=Pmin/2=2 covers each timestep with about 64 overlapping windows, and per-window latency may exceed sub-second alert budgets.On a single NVIDIA RTX PRO 6000 Blackwell GPU, inference is real-time for moderate-length streams.Source: H Limitations · p. 35
  • Distilling the cross-scale bridge into a lightweight student or using amortized cross-series pretraining are proposed future directions, outside this paper’s scope.Source: H Limitations · p. 35

Related papers

Loading 2609.38004v1…