Rethinking the Unsupervised Paradigm in Time Series Anomaly Detection: An Empirical Study and Minimalist Baseline
Authors: Zhijie Zhong, Zhiwen Yu, Kaixiang Yang, Yongheng Liu, Jun Jiang, C. L. Philip Chen
Organizations: School of Future Technology, South China University of Technology, Guangzhou, Guangdong 510650, China · Pengcheng Laboratory, Shenzhen, Guangdong 518066, China · School of Computer Science and Engineering, South China University of Technology, Guangzhou, Guangdong 510650, China
Time series anomaly detection (TSAD) research predominantly focuses on unsupervised architectures to model normal data dynamics under label scarcity. However, practical deployments frequently archive historical operational incidents, questioning whether continuous unsupervised architectural scaling remains the most cost-effective pursuit. In this paper, we empirically quantify the trade-offs between supervisory signals and model complexity. We introduce STAND, a deliberately minimalist supervised baseline based on bidirectional recurrent encoding. Across five benchmark datasets evaluated under a strictly disjoint out-of-sample temporal protocol, minimalist supervised models achieve competitive or superior accuracy relative to complex unsupervised alternatives. Furthermore, controlled label-scaling and few-shot evaluations confirm that these supervisory advantages persist under constrained historical event budgets, while STAND exhibits higher prediction certainty and sharper event delineation than reconstruction-based baselines. These findings encourage the TSAD community to complement algorithmic complexity with data-centric label utilization when annotations are accessible. The source code, benchmark datasets, and an extended Supplementary Document are publicly available at https://emorzz1g.github.io/STAND.
Figures & tables
Figure 1: Classification of time series anomaly detection approaches.
Figure 2: Performance comparison on the held-out test horizon across 20 benchmark methods and six evaluation metrics. (a) Mean performance across all datasets grouped by category, highlighting group averages and the substantial margins achieved by STAND over unsupervised baselines. (b) Multi-metric radar profile comparing representative top-tier methods from each category across individual evaluation dimensions.
Figure 3: The overall architecture of the STAND framework. An input time series is first projected into latent space by the feature embedding. The resulting latent sequence is then fed into the temporal encoding module, which captures bidirectional temporal context. Finally, an anomaly scoring module outputs anomaly scores for each time step.
Dataset
#Ch
UTAD-II Train
Supervised Training Window (50% Prefix)
Held-out Evaluation Window (50% Suffix)
Total
#Timestamps
#Steps
Normal (%)
Anomaly (%)
Events
#Steps
Normal (%)
Anomaly (%)
Events
Events
PSM
25
132,481
43,920
32,991 (75.12%)
10,929 (24.88%)
43
43,921
30,469 (69.37%)
13,452 (30.63%)
29
72
SWaT
50
495,000
224,959
213,202 (94.77%)
11,757 (5.23%)
21
224,960
182,096 (80.95%)
42,864 (19.05%)
14
35
WADI
122
241,921
17,280
15,944 (92.27%)
1,336 (7.73%)
7
17,281
16,634 (96.26%)
647 (3.74%)
6
13
Swan
38
60,000
30,000
25,178 (83.93%)
4,822 (16.07%)
3,696
30,000
15,262 (50.87%)
14,738 (49.13%)
2,443
6,139
Water
9
69,260
34,630
33,977 (98.11%)
653 (1.89%)
21
34,631
34,554 (99.78%)
77 (0.22%)
1
22
Table 1: Characteristics, temporal out-of-sample partitioning, and discrete event counts for the five benchmark datasets. STAD models are trained on the initial 50% temporal prefix and evaluated on the strictly held-out 50% temporal suffix with zero temporal look-ahead leakage. UTAD-II models are pre-trained on the standard nominal sequence from TSB-AD [ 19 ] and evaluated on the same held-out suffix.
Dataset
Split Threshold
#Train
Train Anomaly Rate
Train Events
#Remaining
Rem. Anomaly Rate
Rem. Events
#Overall
Total Events
Overall Anomaly Rate
SP1
10.00%
8784
13.35%
12
79057
29.36%
60
87841
72
27.76%
SP2
20.00%
20182
20.00%
23
67659
30.07%
50
SP3
30.00%
35136
30.73%
35
52705
25.77%
37
Table 2: Characteristics and event-level statistics of the PSM subsets. Train Events and Rem. Events denote the count of independent, contiguous anomaly episodes in the training and remaining evaluation partitions, respectively. For SP2, one continuous anomaly episode spans the boundary at time step 20,182 , resulting in 23 training events and 50 remaining events from the 72 overall continuous episodes in PSM.
Dataset
PSM
SWaT
WADI
Metric
CCE
F1
Aff-F1
UAff-F1
AUC
V-PR
CCE
F1
Aff-F1
UAff-F1
AUC
V-PR
CCE
F1
Aff-F1
UAff-F1
AUC
V-PR
Random
-0.52
8.49
71.43
11.66
50.20
35.30
-0.46
7.87
69.35
-4.59
50.06
19.55
0.99
3.18
66.37
20.87
51.02
6.17
IForest
0.40
14.23
49.22
-1.25
51.34
36.30
0.40
6.72
67.58
-3.93
32.17
15.36
11.57
10.59
72.41
26.20
85.22
19.86
PCA
10.22
17.77
56.73
51.31
78.32
59.91
5.64
34.75
61.50
-2.99
92.11
77.11
16.32
32.83
83.96
58.69
84.93
29.05
HBOS
-0.50
10.66
54.69
8.80
48.40
34.71
14.58
5.07
71.07
26.69
85.23
51.92
16.01
18.53
75.88
35.34
85.02
26.56
KMeans
-1.10
6.79
63.80
53.36
35.30
28.84
0.53
14.31
77.65
40.88
21.37
17.78
13.50
5.43
60.69
0.36
73.85
10.92
Table 3: Performance comparison of the three method categories across five datasets and six metrics evaluated on the held-out test horizon. Bold indicates the best performance, and underlining indicates the second-best performance. Methods highlighted in gray are UTAD-I, methods highlighted in green are UTAD-II, and methods highlighted in orange represent STAD methods.
Source
Model
CCE
F1
Aff-F1
UAff-F1
AUC-ROC
VUS-PR
Avg. Score
Avg. Rk.
SP1
PCA
5.47
18.17
58.88
53.51
73.35
55.78
44.19
2.2
CATCH
0.59
16.89
82.45
55.80
62.88
52.47
45.18
3.8
RF
4.64
9.89
46.56
34.29
60.37
41.24
32.83
6.5
SVM
6.09
11.69
42.43
31.52
63.88
45.37
33.50
5.8
AdaBoost
6.76
15.61
38.38
31.95
71.18
49.58
35.58
4.2
ExtraTrees
6.33
16.26
46.13
30.89
66.19
47.01
35.47
4.8
Table 4: Performance comparison across the three PSM subsets (SP1–SP3). Bold indicates the best performance, and underlining indicates the second-best performance. Methods highlighted in gray are UTAD-I, green are UTAD-II, and plain are STAD. Avg. Score denotes the unweighted arithmetic mean across all six evaluation metrics ( 1/6∑m=16Scorem ), while Avg. Rk. denotes the non-parametric average rank across the six metrics within each partition (lower rank indicates superior performance).
Figure 4: Event-level few-shot label efficiency of supervised anomaly detection models across varying numbers of observed failure episodes ( k∈{1,3,…,15} ) on PSM and SWaT benchmarks. Solid lines and shaded bands represent empirical means and standard deviations across experimental runs. A decisive performance transition occurs at k=5∼7 episodes. Detailed tabulated metrics are reported in Supplementary Table .
Figure 5: Robustness of supervised anomaly detectors under varying ratios of Point Noise on the PSM benchmark. Solid curves depict the empirical mean across multiple repeated runs with distinct random corruption realizations, and shaded error bands denote the standard deviation ( ±1σ ). The horizontal dashed line indicates the performance of CATCH [ 27 ] , an unsupervised UTAD-II baseline evaluated on the nominal sequence. Complete results for contiguous Segment Noise are reported in Supplementary Figure .
Dataset
Variant
Encoder
Bidir.
CCE
F1
Aff-F1
UAff-F1
AUC-ROC
VUS-PR
PSM
STAND-Pointwise
None (MLP)
✘
4.51
13.59
78.69
48.71
59.21
44.24
STAND-Causal
Uni-LSTM
✘
16.39
23.58
64.18
55.61
66.81
55.47
STAND-Attn
Attention
✔
7.02
10.28
59.97
15.55
55.88
39.56
STAND-GRU
Bi-GRU
✔
11.58
15.63
61.55
35.79
62.70
46.98
STAND (Default)
Bi-LSTM
✔
16.39
23.58
64.18
55.61
66.81
55.47
SWaT
STAND-Pointwise
None (MLP)
✘
20.37
41.01
82.26
77.34
74.46
57.22
Table 5: Ablation results of STAND across five datasets. STAND-Pointwise removes the temporal encoder, reducing to a pointwise MLP; STAND-Causal employs a unidirectional LSTM for strict online streaming; STAND-Attn and STAND-GRU substitute the BiLSTM with Multi-Head Attention and Bidirectional GRU, respectively; STAND (Default) is the full bidirectional model. Bold and underlining denote the best and second-best performance within each dataset block.
Table 6: Cross-paradigm comparison with matched sequential backbones averaged across all five benchmark datasets. Panel A reports unsupervised reconstruction ( 0 anomaly labels, MSE loss); Panel B reports supervised classification (full training labels, BCE loss); Panel C reports the net supervisory gain ( Δ=PanelB−PanelA ). Detailed per-dataset evaluations are provided in Supplementary Section .
Figure 6: Sensitivity analysis of model performance with respect to three key hyperparameters on five datasets. Plots (a) and (b) present the impact of model dimension dmodel on CCE and VUS-PR, respectively; (c) and (d) illustrate the effect of the number of TEM layers; (e) and (f) depict the influence of temporal window size. Error bars denote 95% confidence intervals across hyperparameter configurations and independent runs.
Figure 7: Sensitivity analysis of model performance with respect to three key hyperparameters across three subsets of the PSM dataset, SP1–SP3. Plots (a) and (b) present the impact of model dimension dmodel on CCE and VUS-PR, respectively; (c) and (d) illustrate the effect of the number of TEM layers; (e) and (f) depict the influence of temporal window size. Error bars denote 95% confidence intervals across hyperparameter configurations and independent runs.
Figure 8: Anomaly score trajectories of representative models across the complete SWaT sequence. The red vertical dashed line separates the training prefix from the unseen test sequence, with ground-truth anomaly regions shaded in light red.
Figure 9: Anomaly score trajectories on an unseen test segment from timestamp 250k to 300k, showing boundary behavior on sustained and transient anomalies.
Figure 10: Anomaly score trajectories on an unseen test segment from timestamp 350k to 400k, showing detection performance, transition false alarms, and a missed anomaly under distinct abnormal dynamics.
Model
Paradigm
Win. ( W )
Params (M)
Peak VRAM (MB)
Train Time (s/epoch)
Infer. Latency ( μs )
Infer. Throughput ( 103 timestamps/s)
PCA
UTAD (CPU)
30
—
— ∗
30.62 †
548.0
1.8
AE
UTAD (GPU)
100
0.02
17.7
1.00
10.3
96.7
Anomaly Transformer
UTAD (GPU)
100
4.80
5497.1
34.21
240.0
4.2
CATCH
UTAD (GPU)
100
89.86
2231.7
33.33
7.6
132.0
ExtraTrees
STAD (CPU)
3
—
— ∗
0.27 †
1.3
798.1
STAND (Ours)
STAD (GPU)
100
0.04
86.5
1.96
8.8
113.7
Table 7: Empirical computational efficiency and resource consumption on the PSM benchmark ( C=25 , batch size 128 ). All deep architectures are evaluated under W=100 . Peak VRAM represents the maximum GPU memory allocated during training. Inference latency is measured per timestamp. For non-iterative CPU models (PCA and ExtraTrees), training time denotes the total fitting duration on the training split.
Time-Series Anomaly Detection identifies unusual patterns or events in sequential data, from sensor readings and network traffic to health and industrial telemetry.
Time series anomaly detection (TSAD) remains challenging not only because anomaly labels are scarce, but also because temporal anomalies are highly context-dependent. Existing methods often rely on unsupervised objectives or surrogate abnormal patterns, providing limited supervision for context-dependent normal--anomalous distinctions. We propose Context-Anchored Pair Supervision (CAPS), a supervision-recovery framework for TSAD. CAPS views ideal anomaly supervision as a matched comparison between normal and anomalous outcomes under the same temporal context, and seeks to recover such supervision without target-domain anomaly labels. Using simulated normal--anomalous pairs, CAPS learns structure and anomaly-semantic representations through reconstruction, background consistency, and within-pair counterfactual recombination. The resulting anomaly representations form a continuous semantic space with coarse modes and induce a sampleable multimodal prior. CAPS conditionally realizes sampled semantics as residual-form effects on target reference trajectories. The resulting context-anchored normal--anomalous counterparts provide temporal supervision for discriminative detector learning. Experiments on nine datasets show that CAPS achieves the strongest aggregate performance across all four evaluation metrics among the compared methods, while complementary ablations and transfer analyses support the roles of context anchoring, semantic disentanglement, and conditional realization.
Yifei Gao, Tian Lan, Yimeng Lu +5
Department of Industrial Engineering, Tsinghua University · Huawei
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.
Jiaheng Guo, Haochen Zhang, Yu-Chao Huang +3
UNITES Lab, University of North Carolina at Chapel Hill
Time series anomaly detection (TSAD) is essential for maintaining the reliability and security of IoT-enabled service systems. Existing methods require training one specific model for each dataset, which exhibits limited generalization capability across different target datasets, hindering anomaly detection performance in various scenarios with scarce training data. To address this limitation, foundation models have emerged as a promising direction. However, existing approaches either repurpose large language models (LLMs) or construct largescale time series datasets to develop general anomaly detection foundation models, and still face challenges caused by severe cross-modal gaps or in-domain heterogeneity. In this paper, we investigate the applicability of large-scale vision models to TSAD. Specifically, we adapt a visual Masked Autoencoder (MAE) pretrained on ImageNet to the TSAD task. However, directly transferring MAE to TSAD introduces two key challenges: overgeneralization and limited local perception. To address these challenges, we propose VAN-AD, a novel MAE-based framework for TSAD. To alleviate the over-generalization issue, we design an Adaptive Distribution Mapping Module (ADMM), which maps the reconstruction results before and after MAE into a unified statistical space to amplify discrepancies caused by abnormal patterns. To overcome the limitation of local perception, we further develop a Normalizing Flow Module (NFM), which combines MAE with normalizing flow to estimate the probability density of the current window under the global distribution. Extensive experiments on nine real-world datasets demonstrate that VAN-AD consistently outperforms existing state-of-the-art methods across multiple evaluation metrics.We make our code and datasets available at https://github.com/PenyChen/VAN-AD.
PengYu Chen, Shang Wan, Xiaohou Shi +3
School of Computer Science (National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876, China · China Telecom Research Institute Beijing, China · Department of Computer Science, Missouri University of Science and Technology, Rolla, MO 65409 USA