COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection
Authors: Jinwoo Park, Hyeongwon Kang, Seung Hun Han, Pilsung Kang
Organizations: Department of Industrial Engineering, Seoul National University, Gwanak-ro 1, Gwanak-gu, Seoul, Republic of Korea · Department of Industrial & Management Engineering, Korea University, 126-16 Anam-dong 5-ga, Seongbuk-gu, Seoul, Republic of Korea · LG CNS, 71 Magokjungang 8-ro, Gangseo-gu, Seoul, Republic of Korea
Multivariate time-series anomaly detection aims to identify abnormal temporal patterns in complex real-world systems. However, robust unsupervised time-series anomaly detection remains challenging because anomalies can manifest over different temporal ranges and emerge from interactions among variables, while distribution shifts can cause normal patterns to drift over time. To address these challenges, we propose Codebook-based Online-adaptive Multi-scale Embedding for Time-series anomaly detection (COMET), a unified framework with three tightly coupled components. First, Multi-scale Patch Encoder learns correlation-aware patch embeddings by combining variable-specific temporal dynamics with shared inter-variable structure across multiple patch scales. Second, Vector-Quantized Coreset quantizes these embeddings into representative normal prototypes, which are used to detect anomalies through a dual score combining quantization error and density-aware memory distance. Third, Online Codebook Adaptation leverages the same codebook structure to identify reliable normal samples from training-activated codebook entries, and then updates the prototypes through contrastive learning at inference time. Experiments on five benchmark datasets demonstrate that COMET achieves the best performance on 39 out of 45 evaluation metrics compared to seven baselines, showing robust performance across diverse evaluation protocols.
Figures & tables
Figure 1: Overall architecture of COMET. (Top-left) During training, input time series are processed via Multi-scale Patch Encoder to capture temporal contexts at different resolutions. The resulting embeddings are mapped to a Vector-Quantized Coreset , where activated codebook entries are stored in the memory bank. (Top-right) During inference, Online Codebook Adaptation selectively updates the codebook by identifying reliable normal samples through pseudo-labeling and contrastive learning. Anomaly scores are computed by combining memory distance and quantization error. (Bottom-left) Detailed illustration of vector quantization, where continuous patch embeddings are assigned to their nearest codebook entries and represented by the corresponding quantized embeddings. (Bottom-right) Detailed illustration of Online Codebook Adaptation , comprising normal filtering, contrastive learning, and index-wise coreset update. Reliable normal samples are selected based on training-activated codebook entries, and the corresponding codebook prototypes are updated to adapt to distribution shifts while reducing contamination from anomalous patterns.
Figure 2: Comparison between global distance and local scaling distance. (Left) Global distance fails to identify normal queries in sparse regions as it relies solely on absolute distance without considering local density. (Right) Local scaling distance accounts for local density, leading to more consistent classification of queries across regions with varying density.
Dataset
Method
Metric
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
PSM
CATCH
97.70
42.38
60.77
62.07
41.13
70.46
53.01
69.50
51.74
D3R
93.67
44.99
57.39
69.17
48.97
69.49
50.59
69.18
50.36
VTT
92.81
48.35
56.13
76.94
49.93
76.23
52.95
76.01
52.69
AT
97.43
2.20
65.19
50.48
27.88
52.33
34.81
52.13
34.61
LSTM-AE
93.41
45.54
53.37
69.70
50.76
70.26
51.35
69.94
51.15
Table 1: Main experimental results on five anomaly detection benchmarks. We report point-wise metrics and range-based metrics to evaluate detection robustness. Bold indicates the best score, and underline indicates the second best.
Category
Method
Metric
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
Architecture
w/o Multi-Scale
76.45
36.31
69.48
63.17
30.78
63.62
28.66
63.25
28.59
Scoring
w/o Quant Score
83.70
35.74
71.00
67.84
24.77
69.27
26.77
68.69
26.55
w/o Memory Score
81.55
37.96
69.19
65.62
34.65
64.85
32.88
64.48
32.78
w/o Local Scaling
83.92
39.31
71.00
68.85
34.08
70.64
34.52
70.17
34.27
w/o Variable Selection
83.65
38.54
70.97
68.16
34.31
68.50
34.47
67.94
34.31
Table 2: Ablation study results on COMET, averaged across all five benchmark datasets. Detailed ablation results for each dataset are provided in Appendix C.5 . Bold indicates the best score, and underline indicates the second best. Ties after rounding are ranked by the unrounded averages.
Figure 3: Analysis of COMET. (a) Affiliation F1-Score versus training time on the PSM dataset, where circle size indicates the number of model parameters. COMET achieves the best performance while using a compact parameter count, demonstrating its parameter efficiency relative to all baselines. (b) UMAP visualization of patch embeddings and codebook entries on the SWaT dataset. Train Normal samples are shown in gray, Test Normal in light blue, and Test Anomaly in pink. Codebook entries for each patch scale are marked as green circles (Patch 2), purple squares (Patch 4), and orange triangles (Patch 6). The codebook entries cover the normal-pattern region while remaining clearly separated from anomalies, supporting the feasibility of distance-based detection.
Method
Metric
Cost
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
Mem. (s)
Cost
VQEnc-KMeans
93.28
50.32
69.44
71.52
52.92
69.29
52.83
68.33
52.19
677
4.58 ×
VQEnc-Greedy
82.63
46.83
71.31
69.56
49.15
65.83
48.01
65.01
47.46
25
1.13 ×
NoVQ-KMeans
91.02
49.88
69.43
70.76
51.53
71.40
52.39
70.36
51.76
641
3.86 ×
NoVQ-Greedy
83.23
46.65
71.36
69.49
48.74
65.81
47.79
64.98
47.27
27
0.61 ×
NoVQ-Contr-KMeans
94.15
53.11
69.43
73.67
55.24
72.42
55.27
70.79
54.31
644
3.97 ×
Table 3: Learned codebook versus post-hoc memory on PSM (COMET w/o TTA). Mem. is the time to build the memory after training, and Cost is the total time cost relative to COMET.
Figure 4: (a) VUS-PR gain from TTA versus train–test shift (improved metrics in parentheses). (b) VUS-PR of COMET (w/ TTA) relative to the best baseline. Dashed lines: no gain and parity.
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Variant
Metric
Precision
Recall
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
PSM
w/o Var. Sel.
99.10
89.91
94.28
59.92
71.85
79.45
59.44
73.66
57.74
72.39
57.04
COMET (w/o TTA)
98.57
92.37
95.37
60.13
71.80
79.15
59.10
79.18
61.45
77.50
60.42
SWaT
w/o Var. Sel.
94.85
93.18
94.01
74.91
74.94
85.21
74.51
85.77
70.52
85.75
70.67
COMET (w/o TTA)
97.33
86.12
91.38
75.06
72.93
85.48
74.33
84.85
65.46
84.82
65.50
SMAP
w/o Var. Sel.
89.83
64.68
75.20
21.05
68.09
57.54
15.34
56.56
16.63
56.13
16.58
Appendix
Table 4: Effect of Variable Selection across all benchmark datasets and metrics. We report Precision, Recall, and all nine point-wise and range-based metrics for both conditions, evaluated without test-time adaptation. Precision and Recall are computed under the same protocol as F1(K=0).
Dataset
Variant
Metric
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
PSM
Core Only
89.21
46.15
69.49
66.89
44.62
68.25
46.52
67.38
46.12
Variable Only
83.99
36.88
69.47
56.42
33.75
59.39
37.87
58.54
37.67
COMET (w/o TTA)
95.37
60.13
71.80
79.15
59.10
79.18
61.45
77.50
60.42
SWaT
Core Only
89.35
67.44
72.01
83.52
69.48
82.99
63.06
82.97
62.97
Variable Only
91.95
56.10
73.46
83.34
58.50
84.06
56.48
83.99
56.39
Appendix
Table 5: Ablation of Variable-wise Encoder and Core Encoder branches across all benchmark datasets and metrics. All variants are evaluated without test-time adaptation, and bold indicates the best score.
PSM
SWaT
SMAP
MSL
WADI
Variables
25
51
25
55
123
#Train(0.9)
116,805
445,500
124,184
52,473
1,088,595
#Valid(0.1)
12,979
49,500
13,820
5,844
120,956
#Test
87,841
449,919
435,826
73,729
172,801
Anomaly (%)
27.76
12.14
12.84
10.53
5.71
Appendix
Table 6: Statistics of benchmark datasets. #TRAIN and #VALID denote the number of samples in training and validation sets, respectively, where the validation set is split from the original training data with a ratio of 0.1. ANOMALY (%) indicates the proportion of anomalous samples in the test set.
Table 8: Per-dataset training and inference time (in seconds) of COMET on a single NVIDIA TITAN RTX (24GB) GPU. Both variants share the same training procedure, and the w/ TTA variant additionally applies online codebook adaptation during inference.
Dataset
Method
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
Mem. (s)
Cost
PSM
VQEnc-KMeans
93.28
50.32
69.44
71.52
52.92
69.29
52.83
68.33
52.19
677
4.6 ×
VQEnc-Greedy
82.63
46.83
71.31
69.56
49.15
65.83
48.01
65.01
47.46
25
1.1 ×
NoVQ-KMeans
91.02
49.88
69.43
70.76
51.53
71.40
52.39
70.36
51.76
641
3.9 ×
NoVQ-Greedy
83.23
46.65
71.36
69.49
48.74
65.81
47.79
64.98
47.27
27
0.6 ×
NoVQ-Contr-KMeans
94.15
53.11
69.43
73.67
55.24
72.42
55.27
70.79
54.31
644
4.0 ×
NoVQ-Contr-Greedy
85.37
49.40
71.70
71.76
50.73
67.99
49.94
67.08
49.37
26
0.7 ×
Appendix
Table 9: Learned codebook versus post-hoc memory on each dataset and on average (COMET w/o TTA). Mem. is the time to build the memory after training in seconds, and Cost is the total training and memory-construction time relative to COMET. Bold indicates the best score, and underline indicates the second best within each block. For Mem. and Cost , lower is better.
Dataset
JS
Variant
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
Improved
MSL
0.031
w/o TTA
87.27
26.44
69.40
65.49
16.63
69.27
22.67
68.39
22.27
4/9
w/ TTA
87.44
26.28
71.04
65.40
16.56
69.90
22.67
68.87
22.25
SMAP
0.045
w/o TTA
82.46
25.40
68.01
59.06
16.17
58.69
17.51
58.52
17.46
7/9
w/ TTA
82.30
25.28
68.24
59.17
16.20
59.43
17.59
59.20
17.54
PSM
0.066
w/o TTA
95.37
60.13
71.80
79.15
59.10
79.18
61.45
77.50
60.42
8/9
w/ TTA
95.30
60.26
72.00
79.54
60.01
79.56
62.21
77.92
61.20
Appendix
Table 10: TTA effect versus train–test shift. JS is the Jensen–Shannon distance between training and test codeword-assignment histograms. Bold marks metrics improved by TTA, and the Improved column counts them.
Dataset
Ratio
CATCH
D3R
VTT
AT
LSTM-AE
LSTM-VAE
USAD
COMET (w/ TTA)
COMET (w/o TTA)
PSM
0%
51.74
50.36
52.69
34.61
51.15
48.27
46.21
61.20
60.42
5%
52.85
50.32
52.14
34.01
51.16
48.21
46.29
58.10
57.95
10%
53.37
50.20
50.73
33.02
51.11
48.19
46.10
61.10
60.70
25%
54.12
50.19
48.28
33.44
51.25
48.16
46.32
62.36
61.33
SWaT
0%
9.86
47.57
43.74
12.36
46.75
45.97
46.08
66.48
65.50
5%
9.33
47.80
33.86
11.52
46.52
45.75
45.79
65.63
70.64
Appendix
Table 11: Robustness when the training data are contaminated by anomalies. VUS-PR is reported for each model after synthetic anomalies are injected into the given ratio of training timesteps. Bold indicates the best score, and underline indicates the second best within each row.
Figure 5: Performance–efficiency visualizations across datasets (Affiliation F1-score vs. training time, with circle size indicating parameter count).
Category
Method
Metric
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
Architecture
w/o Multi-Scale
87.60
46.49
71.33
67.04
48.75
63.50
46.25
62.71
45.83
Scoring
w/o Quant Score
96.78
49.30
70.78
68.31
44.68
71.03
49.38
69.33
48.61
w/o Memory Score
94.47
61.68
73.15
81.40
64.10
76.66
61.48
75.58
60.68
w/o Local Scaling
95.02
58.11
73.94
78.46
57.22
79.76
60.13
78.24
59.22
w/o Variable Selection
94.28
59.92
71.85
79.45
59.44
73.66
57.74
72.39
57.04
Appendix
Table 12: Ablation results of COMET on the PSM dataset. Bold indicates the best score, and underline indicates the second best.
Category
Method
Metric
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
Architecture
w/o Multi-Scale
87.17
74.31
70.50
82.82
71.58
80.43
56.75
80.44
56.95
Scoring
w/o Quant Score
70.02
49.52
71.72
77.32
31.59
77.17
29.25
77.07
29.34
w/o Memory Score
90.83
75.19
75.65
87.25
75.61
83.16
62.73
83.34
63.28
w/o Local Scaling
91.81
75.14
73.06
84.24
73.58
83.79
65.39
83.75
65.46
w/o Variable Selection
94.01
74.91
74.94
85.21
74.51
85.77
70.52
85.75
70.67
Appendix
Table 13: Ablation results of COMET on the SWaT dataset. Bold indicates the best score, and underline indicates the second best.
Category
Method
Metric
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
Architecture
w/o Multi-Scale
74.64
15.58
62.62
44.41
11.65
45.97
12.93
45.74
12.89
Scoring
w/o Quant Score
85.63
26.39
68.17
60.12
17.01
59.69
18.14
59.49
18.09
w/o Memory Score
69.55
18.87
63.97
45.62
12.70
46.64
13.88
46.52
13.85
w/o Local Scaling
75.16
24.84
68.09
58.02
15.59
59.21
17.10
59.10
17.04
w/o Variable Selection
75.20
21.05
68.09
57.54
15.34
56.56
16.63
56.13
16.58
Appendix
Table 14: Ablation results of COMET on the SMAP dataset. Bold indicates the best score, and underline indicates the second best.
Category
Method
Metric
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
Architecture
w/o Multi-Scale
55.48
25.00
69.77
59.05
12.44
63.18
16.89
62.37
16.81
Scoring
w/o Quant Score
84.42
33.08
71.76
69.19
20.20
71.54
26.35
70.76
25.97
w/o Memory Score
83.35
24.39
61.24
63.70
15.01
66.23
19.93
65.46
19.71
w/o Local Scaling
83.21
20.17
67.96
63.22
15.28
67.80
20.34
67.17
20.01
w/o Variable Selection
76.95
19.15
68.05
57.60
13.35
63.09
17.62
62.09
17.46
Appendix
Table 15: Ablation results of COMET on the MSL dataset. Bold indicates the best score, and underline indicates the second best.
Category
Method
Metric
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
Architecture
w/o Multi-Scale
77.34
20.19
73.17
62.54
9.47
65.04
10.47
65.00
10.48
Scoring
w/o Quant Score
81.64
20.39
72.57
64.25
10.36
66.91
10.74
66.81
10.75
w/o Memory Score
69.57
9.66
71.95
50.15
5.81
51.57
6.37
51.51
6.36
w/o Local Scaling
74.38
18.31
71.96
60.31
8.74
62.62
9.62
62.58
9.63
w/o Variable Selection
77.79
17.67
71.91
60.98
8.92
63.41
9.82
63.36
9.82
Appendix
Table 16: Ablation results of COMET on the WADI dataset. Bold indicates the best score, and underline indicates the second best.
Dataset
Patch & Stride Size
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
PSM
{2}∣{1}
93.08
58.11
71.18
77.92
59.17
78.07
59.70
77.47
59.20
{4}∣{2}
92.06
57.90
70.98
76.78
53.04
76.30
56.54
74.22
55.47
{6}∣{3}
87.60
46.49
71.33
67.04
48.75
63.50
46.25
62.71
45.83
{2,4}∣{1,2}
82.58
45.25
71.48
67.02
46.41
67.56
46.91
66.67
46.32
{4,6}∣{2,3}
91.80
44.62
71.28
68.43
48.24
68.58
50.66
67.48
50.05
{2,4,6}∣{1,2,3}
95.37
60.13
71.80
79.15
59.10
79.18
61.45
77.50
60.42
Appendix
Table 17: Hyperparameter sensitivity analysis on multi-scale patch configurations (COMET w/o TTA). Patch sizes and strides are denoted as {patch sizes}∣{strides} .
Table 19: Hyperparameter sensitivity analysis on the number of nearest neighbors (COMET w/o TTA).
Dataset
λ
F1(K=0)
F1(K=100)
Aff-F1
AUC-ROC
AUC-PR
R-AUC-ROC
R-AUC-PR
VUS-ROC
VUS-PR
PSM
0
96.78
49.31
70.78
68.31
44.68
71.03
49.38
69.33
48.61
0.25
95.47
55.52
71.20
74.92
52.39
76.53
56.47
74.82
55.52
0.5
95.37
60.13
71.80
79.15
59.10
79.18
61.45
77.50
60.42
0.75
94.89
61.85
71.79
81.17
63.05
78.31
62.47
76.94
61.54
1
94.47
61.68
73.15
81.40
64.10
76.66
61.48
75.58
60.68
SWaT
0
70.02
49.52
71.72
77.32
31.59
77.17
29.25
77.07
29.34
Appendix
Table 20: Hyperparameter sensitivity analysis on the score weight λ (COMET w/o TTA). λ=0 uses only the memory score, and λ=1 uses only the quantization score.
Dataset
Scale k=1
Scale k=2
Scale k=3
Total ( ∣M∣/KM )
PSM
75.0%
64.8%
61.7%
67.2%
SWaT
52.3%
62.5%
73.8%
62.9%
SMAP
91.4%
98.4%
96.9%
95.6%
MSL
84.0%
91.4%
93.8%
89.7%
WADI
100.0%
100.0%
100.0%
100.0%
Appendix
Table 21: Codebook activation rate per scale and total for each benchmark dataset. Per-scale rate is ∣Ak∣/M , and total rate is ∣M∣/(K×M) .
Dataset
Subdataset
LSTM-AE
LSTM-VAE
USAD
AT
VTT
CATCH
D3R
COMET
SMAP
A-9
9.88
9.88
9.88
74.85
9.88
10.23
11.51
92.08
A-8
9.69
9.69
9.69
76.65
9.69
10.42
9.69
90.03
A-2
69.22
69.22
69.22
67.95
68.02
68.68
69.31
77.07
P-7
14.38
14.81
25.15
75.55
8.52
7.53
24.24
76.07
A-7
0.96
0.96
0.96
99.56
0.96
0.96
25.29
99.02
P-1
66.25
66.34
65.93
68.51
66.03
65.75
66.12
67.42
Appendix
Table 22: Aff-F1 of COMET vs. baselines on every contextual-only labeled subdataset where COMET ranks first or second. Subdatasets are restricted to those containing only contextual labels (no point labels) to isolate contextual anomaly localization. Bold indicates the best score, and underline indicates the second best.
Dataset
Normal Dist
Anomaly Dist
SepRatio ( ↑ )
PSM
14.502
15.076
1.040
SWaT
14.853
60.165
4.051
SMAP
9.339
9.435
1.010
MSL
9.347
10.113
1.082
WADI
4157.814
2887.464
0.694
Appendix
Table 23: Quantitative separability between normal and anomalous test patches on the five benchmarks. We report the mean nearest-codebook distance of each population and the separability ratio (Anomaly / Normal). A SepRatio above 1 indicates that anomalous patches lie farther from the codebook than normal patches, which is the desired condition for distance-based scoring. Bold marks ratios above 1 .
The core challenge in unsupervised anomaly detection is identifying abnormal patterns without prior knowledge of their characteristics. While existing methods have addressed aspects of this problem, they often struggle to learn a robust representation of the normal data distribution that is distinct from anomalous patterns. In this paper, we present a novel framework, Unified Unsupervised Anomaly Detection (U2AD), that comprehensively addresses anomaly detection in multivariate time series. Our approach learns the underlying data distribution of normal samples by utilizing score-based generative modeling. We introduce a novel time-dependent score network and a unified training objective that together delineate the manifold of normal data while considering both local and global temporal contexts. Reconstruction is then performed via a deterministic sampling process using an ordinary differential equation solver. Our extensive experimental evaluations demonstrate that U2AD not only outperforms current state-of-the-art methods in detection accuracy but also identifies anomalies at significantly earlier stages of their occurrence.
Prithul Sarker, Sushmita Sarker, Nicholas G. Murray +1
Time-series anomaly detection has significant practical value for industrial and medical monitoring, as well as other critical domains. Current Transformer- and large-model-based detection approaches incur excessive computational overhead, while existing lightweight alternatives are constrained by insufficient feature extraction and inadequate modeling of dependencies across multivariate variables. To mitigate the above drawbacks, this study develops a lightweight, efficient anomaly detection model, dubbed PaAno, within the patch-oriented representation learning paradigm. In the encoder module, a multiscale feature-extraction backbone is constructed using convolutional kernels with differentiated receptive fields to capture hierarchical temporal characteristics; subsequent cross-scale adaptive attention aggregation, combined with residual connection optimization, further stabilizes feature representation learning. A cross-variable fusion attention module is embedded to explicitly characterize inter-variable correlations, empowering the model to identify anomalous patterns amid intricate operational conditions. Moreover, a novel pretext task based on temporal patch-window sorting is customized to uncover intrinsic structural properties of time series, and triplet loss is leveraged to optimize the patch embedding space for enhanced feature discrimination. Extensive experiments on the TSB-AD benchmark demonstrate that the proposed PaAno achieves state-of-the-art detection accuracy on both univariate and multivariate tasks, yielding significant performance gains across evaluation metrics, including VUS-PR, relative to the original PaAno. Leveraging a compact network design, the presented model achieves favorable computational efficiency, enabling deployment on resource-limited terminals for real-time anomaly inference.
Youji Zhu, Hongbing Wang, Wenchao Liu +2
School of Mathematical Sciences, Guizhou Normal University, Guiyang 550025, China. · School of Big Data and Computer Science, Guizhou Normal University, Guiyang 550025, China.
Although recent studies on time-series anomaly detection have increasingly adopted ever-larger neural network architectures such as transformers and foundation models, they incur high computational costs and memory usage, making them impractical for real-time and resource-constrained scenarios. Moreover, they often fail to demonstrate significant performance gains over simpler methods under rigorous evaluation protocols. In this study, we propose Patch-based representation learning for time-series Anomaly detection (PaAno), a lightweight yet effective method for fast and efficient time-series anomaly detection. PaAno extracts short temporal patches from time-series training data and uses a 1D convolutional neural network to embed each patch into a vector representation. The model is trained using a combination of triplet loss and pretext loss to ensure the embeddings capture informative temporal patterns from input patches. During inference, the anomaly score at each time step is computed by comparing the embeddings of its surrounding patches to those of normal patches extracted from the training time-series. Evaluated on the TSB-AD benchmark, PaAno achieved state-of-the-art performance, significantly outperforming existing methods, including those based on heavy architectures, on both univariate and multivariate time-series anomaly detection across various range-wise and point-wise performance measures.
Jinju Park, Seokho Kang
Department of Industrial Engineering, Sungkyunkwan University