LEARN-TS: LLM-Enhanced Alignment and Reconstruction with Normality Guidance for Multivariate Time-Series Anomaly Detection

Jahyeob Koo; Kio Yun; Byoungmo Koo; Jun-Geol Baek

cs.LGSep 30, 2026

LEARN-TS: LLM-Enhanced Alignment and Reconstruction with Normality Guidance for Multivariate Time-Series Anomaly Detection

Authors: Jahyeob Koo, Kio Yun, Byoungmo Koo, Jun-Geol Baek

Organizations: Department of Industrial and Management Engineering, Korea University Seoul, Republic of Korea

Abstract

Reconstruction errors in multivariate time-series anomaly detection may not reliably distinguish abnormal behavior from benign deviations. Language-derived semantics offer complementary context, but existing multimodal approaches may rely on time-associated paired textual information that is difficult to obtain consistently and is not provided by standard multivariate time-series anomaly detection benchmarks. This setting poses two challenges: (1) conditioning masked reconstruction on window-specific semantics without exposing exact numerical targets or anomaly-specific cues, and (2) using a window-independent concept of normality as a complementary semantic reference rather than an independent anomaly detector. We propose LLM-Enhanced Alignment and Reconstruction with Normality Guidance for Time Series (LEARN-TS), which uses a frozen language model to construct two role-separated semantic representations without requiring temporally paired external text. Window-specific observation semantics encode temporal and cross-variable context without exact numerical values to guide channel-shared patch-masked reconstruction. A fixed, dataset-agnostic normality prompt provides a window-independent semantic reference for aligning normal representations and estimating normality discrepancy. At inference, masking each temporal patch once yields timestamp-level reconstruction evidence, conditionally modulated by discrepancy from a separate unmasked view. Across four benchmarks, LEARN-TS achieves the highest mean performance in 13 of 16 dataset-metric comparisons. Controlled ablations examine observation conditioning, joint normality alignment and scoring, and reference content, showing dataset-dependent ranking benefits and modest average gains from semantic over random references.

Figures & tables

Figure 1: (a) Time-associated paired external textual context may vary across datasets and monitored intervals, providing context-dependent semantic information. (b) LEARN-TS instead derives a shared, window-independent normality reference from a fixed, dataset-agnostic prompt applied unchanged across all datasets and windows.

Figure 2: Overview of the three-phase LEARN-TS framework.

Dataset	Metric	Omni	A.T.	Times	Patch	iTrans	G4TS	TLLM	LMixer	CALF	DADA	LEARN-TS
SWaT	A-PR	0.1492	0.6829	0.1341	0.0890	0.0931	0.0894	0.0846	0.0866	0.0819	0.5385	0.7578
	V-PR	0.1371	0.4508	0.1656	0.0982	0.1008	0.0991	0.0930	0.0951	0.0907	0.4543	0.4859
	R-F1	0.1718	0.1602	0.1593	0.1097	0.1277	0.1344	0.1190	0.1201	0.1137	0.0990	0.3102
	Aff-F1	0.7137	0.7413	0.7940	0.6376	0.6621	0.6826	0.6658	0.6737	0.6570	0.7119	0.7277
SMD	A-PR	0.3713	0.2575	0.4640	0.4893	0.4726	0.4947	0.4514	0.4307	0.4578	0.4754	0.4930
	V-PR	0.3752	0.3214	0.6184	0.6376	0.6250	0.6455	0.5839	0.5662	0.5910	0.5912	0.5398

Table 1: Performance comparison on four real-world multivariate time-series anomaly detection datasets. Results are averaged over five runs; standard deviations are reported in Table 7 . For each dataset–metric pair, the best and second-best results are shown in bold and underlined , respectively.

Figure 3: Effect of observation–window correspondence: (a)–(b) matched vs. shuffled observations on SWaT and MSL; (c) A-PR and V-PR gains across all four datasets.

Figure 4: Event-level predictions on three SMD entities: Machine 1-3 with intermediate event density, Machine 2-3 with sparse events, and Machine 2-9 with dense events.

Figure 5: Sensitivity to (a) $\lambda_{\mathrm{full}}$ , (b) $\lambda_{\mathrm{norm}}$ , (c) $\lambda_{\mathrm{gate}}$ , and (d) the number of masked training patches. Results are four-dataset macro averages using a fixed seed.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Domain	Entities	Variables	Official train	Official test	Anomaly (%)
SWaT	Industrial control	1	51	496,800	449,919	11.98
SMD	Server monitoring	28	38	708,405	708,420	4.16
PSM	Application server	1	25	132,481	87,841	27.75
MSL	Spacecraft telemetry	27	55	58,317	73,729	10.72

Appendix

Table 3: Dataset statistics before the internal training–validation split. For SMD and MSL, sequence lengths are summed over all entities.

Hyperparameter	Value
Input window length $L$	128
Patch length $\ell$	16
Number of patches $P$	8
Hidden dimension $d$	768
Time-series encoder layers	3
Cross-modal fusion layers	2

Appendix

Table 4: Common model and optimization settings used across all datasets.

Dataset	$\lambda_{\mathrm{norm}}$	$\lambda_{\mathrm{gate}}$
SWaT	0.02	1.00
SMD	0.01	0.05
PSM	0.01	0.1
MSL	0.01	0.05

Appendix

Table 5: Dataset-specific loss and scoring configurations.

Dataset	Channel aggregation $\mathcal{G}_{\mathcal{D}}$	Window overlap
SWaT	Mean over all variables	Mean
SMD	Top- $k$ mean	Maximum
PSM	Top- $k$ mean	Maximum
MSL	Target telemetry channel	Mean

Appendix

Table 6: Dataset-specific channel error and overlapping-window aggregation rules.

(a) SWaT
Method	A-PR	V-PR	R-F1	Aff-F1
OmniAnomaly	0.1492 $\pm$ 0.0559	0.1371 $\pm$ 0.0321	0.1718 $\pm$ 0.1140	0.7137 $\pm$ 0.0041
Anomaly Transformer	0.6829 $\pm$ 0.0193	0.4508 $\pm$ 0.0077	0.1602 $\pm$ 0.0187	0.7413 $\pm$ 0.0149
TimesNet	0.1341 $\pm$ 0.0018	0.1656 $\pm$ 0.0021	0.1593 $\pm$ 0.0057	0.7940 $\pm$ 0.0091
PatchTST	0.0890 $\pm$ 0.0002	0.0982 $\pm$ 0.0002	0.1097 $\pm$ 0.0062	0.6376 $\pm$ 0.0026
iTransformer	0.0931 $\pm$ 0.0010	0.1008 $\pm$ 0.0005	0.1277 $\pm$ 0.0029	0.6621 $\pm$ 0.0025

Appendix

Table 7: Performance of all methods on the four datasets over five independent runs. Results are reported as mean

\pm

sample standard deviation. For each dataset–metric pair, the best and second-best mean results are shown in bold and underlined , respectively.

Dataset	Variant	A-PR	V-PR	R-F1	Aff-F1
SWaT	w/o Obs.	0.6552 $\pm$ 0.0652	0.3727 $\pm$ 0.0193	0.2446 $\pm$ 0.0110	0.7047 $\pm$ 0.0012
	Concat fusion	0.7001 $\pm$ 0.1014	0.4170 $\pm$ 0.0508	0.2911 $\pm$ 0.0189	0.7270 $\pm$ 0.0090
	LEARN-TS (Full)	0.7578 $\pm$ 0.0117	0.4859 $\pm$ 0.0346	0.3102 $\pm$ 0.0436	0.7277 $\pm$ 0.0069
SMD	w/o Obs.	0.4650 $\pm$ 0.0074	0.5266 $\pm$ 0.0047	0.3222 $\pm$ 0.0070	0.8137 $\pm$ 0.0102
	Concat fusion	0.4829 $\pm$ 0.0122	0.5380 $\pm$ 0.0098	0.3262 $\pm$ 0.0057	0.8174 $\pm$ 0.0084
	LEARN-TS (Full)	0.4930 $\pm$ 0.0056	0.5398 $\pm$ 0.0037	0.3203 $\pm$ 0.0033	0.8121 $\pm$ 0.0056

Appendix

Table 8: Dataset-wise cross-modal fusion ablation. Results are reported as mean

\pm

sample standard deviation over five random seeds. The best result in each dataset–metric column is shown in bold .

(a) SWaT
Variant	Align.	Score	A-PR	V-PR	R-F1	Aff-F1
No normality	✗	✗	0.2088 $\pm$ 0.0076	0.1748 $\pm$ 0.0010	0.2971 $\pm$ 0.0611	0.7407 $\pm$ 0.0157
Scoring only	✗	✓	0.1460 $\pm$ 0.0531	0.1530 $\pm$ 0.0264	0.2711 $\pm$ 0.0635	0.7103 $\pm$ 0.0271
Alignment only	✓	✗	0.2004 $\pm$ 0.0094	0.1766 $\pm$ 0.0008	0.2646 $\pm$ 0.0211	0.7299 $\pm$ 0.0097
LEARN-TS (Full)	✓	✓	0.7578 $\pm$ 0.0117	0.4859 $\pm$ 0.0346	0.3102 $\pm$ 0.0436	0.7277 $\pm$ 0.0069
(b) SMD

Appendix

Table 9: Dataset-wise ablation of normality alignment and discrepancy-guided scoring. Results are reported as mean

\pm

sample standard deviation over five random seeds. The best result in each dataset–metric column is shown in bold .

(a) SWaT
Normality reference	A-PR	V-PR	R-F1	Aff-F1
Random semantic-free	0.7431 $\pm$ 0.0426	0.4758 $\pm$ 0.0364	0.2933 $\pm$ 0.0182	0.7276 $\pm$ 0.0017
Dataset-agnostic semantic	0.7578 $\pm$ 0.0117	0.4859 $\pm$ 0.0346	0.3102 $\pm$ 0.0436	0.7277 $\pm$ 0.0069
(b) SMD
Random semantic-free	0.4919 $\pm$ 0.0052	0.5388 $\pm$ 0.0052	0.3153 $\pm$ 0.0058	0.8022 $\pm$ 0.0143
Dataset-agnostic semantic	0.4930 $\pm$ 0.0056	0.5398 $\pm$ 0.0037	0.3203 $\pm$ 0.0033	0.8121 $\pm$ 0.0056

Appendix

Table 10: Dataset-wise comparison between random semantic-free and dataset-agnostic semantic normality references. Results are reported as mean

\pm

sample standard deviation over five random seeds. The best result in each dataset–metric column is shown in bold ; values tied at the displayed precision are both highlighted.

(a) Full-reconstruction weight $\lambda_{\mathrm{full}}$
Dataset	Value	A-PR	V-PR	R-F1	Aff-F1
SWaT	0.0	0.1766	0.1738	0.2920	0.7381
	0.1	0.7654	0.5004	0.2793	0.7121
	0.2	0.7662	0.5122	0.3293	0.7332
	0.5	0.7684	0.5139	0.3784	0.7343
	1.0	0.7665	0.5023	0.2655	0.7276

Appendix

Table 11: Dataset-wise sensitivity to (a) the full-reconstruction weight

\lambda_{\mathrm{full}}

and (b) the normality-alignment weight

\lambda_{\mathrm{norm}}

. Results are obtained using a fixed random seed. Bold rows denote the configurations used in the main experiments.

(a) Discrepancy-gating coefficient $\lambda_{\mathrm{gate}}$
Dataset	Value	A-PR	V-PR	R-F1	Aff-F1
SWaT	0.00	0.2132	0.1763	0.2607	0.7387
	0.01	0.7622	0.5137	0.2786	0.7330
	0.05	0.7640	0.5138	0.2861	0.7338
	0.10	0.7670	0.5153	0.3121	0.7337
	0.50	0.7686	0.5138	0.3299	0.7394

Appendix

Table 12: Dataset-wise sensitivity to (a) the discrepancy-gating coefficient

\lambda_{\mathrm{gate}}

and (b) the number of masked training patches. Results are obtained using a fixed random seed. Bold rows denote the configurations used in the main experiments.

Dataset	Backbone	Trainable	Peak Train GPU	Embed.	Scoring
		Params $\downarrow$	(GiB) $\downarrow$	(prompts/s) $\uparrow$	(windows/s) $\uparrow$
SWaT	GPT-2	41.16M	0.795	80.90	1111.46
	Qwen2.5-1.5B	41.75M	0.811	38.81	932.19
	Llama-3.2-1B	42.14M	0.822	53.18	958.34
PSM	GPT-2	40.79M	0.792	80.15	1194.91
	Qwen2.5-1.5B	41.38M	0.803	40.17	1152.70

Appendix

Table 13: Backbone efficiency and detection performance on four datasets. Detection results are mean

\pm

sample standard deviation over five seeds. Embedding throughput measures prompt encoding; scoring throughput excludes text encoding and uses cached embeddings. Bold indicates the best result within each dataset and column.

Figure 6: Dataset-agnostic normality prompt used in LEARN-TS. The prompt is fixed before training and shared unchanged across all datasets and input windows.

Figure 7: Abridged SWaT observation prompt template. Blue placeholders denote window-dependent descriptors and patch-index lists; dataset-level descriptions and rules remain fixed. Exact measurements and explicit anomaly labels are excluded. Repeated entries are omitted for readability.

Explore similar work

CALAD: Channel-Aware contrastive Learning for multivariate time series Anomaly Detection
May 22, 2026Jaehyeop Hong, Youngbum HurTime-Series Anomaly Detection Logical Anomalies
Department of Industrial Engineering, Inha University, Incheon, Republic of Korea
Supervision Recovery for Time Series Anomaly Detection via Context-Anchored Pairing
Sep 27, 2026Yifei Gao, Tian Lan, Yimeng Lu +5Time-Series Anomaly Detection
Department of Industrial Engineering, Tsinghua University · Huawei
Learning Unified Representations of Normalcy for Time Series Anomaly Detection
May 10, 2026Prithul Sarker, Sushmita Sarker, Nicholas G. Murray +1Time-Series Anomaly Detection Graph Anomaly Detection
University of Nevada, Reno

May 22, 2026, cs.LGJ/K move · Enter open · S save

CALAD: Channel-Aware contrastive Learning for multivariate time series Anomaly Detection

Jaehyeop Hong, Youngbum Hur

Department of Industrial Engineering, Inha University, Incheon, Republic of Korea

Time-Series Anomaly Detection Logical Anomalies Contrastive Learning Real-World Dataset

Open paper AskPDF arXiv

Multivariate time series anomaly detection has become increasingly important in real-world applications, where labeled data are often scarce. Many existing approaches rely on unsupervised learning to model normal patterns, but they often treat all channels equally. This design can dilute anomaly-relevant signals, since not all channels contribute equally to anomaly detection. In this paper, we propose CALAD, a channel-aware contrastive learning framework for multivariate time series anomaly detection. CALAD governs the construction of contrastive samples using estimated channel relevance, allowing the learning process to reflect anomaly semantics rather than generic similarity. Channel relevance is estimated from reconstruction errors of a transformer-based autoencoder and is used to distinguish channels that are more influential to anomalous behaviors. Using this information, we design a channel-wise augmentation strategy in which positive and negative samples are constructed based on whether anomaly-relevant channels are preserved or perturbed. This encourages invariance to changes in irrelevant channels while being sensitive to changes in anomaly-relevant channels. Furthermore, CALAD combines contrastive learning and an auxiliary reconstruction head, allowing the model to learn discriminative representations while retaining normal structures. Experiments on multiple real-world datasets shows that CALAD consistently outperforms existing methods, particularly under distribution shift scenarios. We provide the code for reproducibility at https://github.com/hirundo1218/CALAD

Abstract

Explore similar work

CALAD: Channel-Aware contrastive Learning for multivariate time series Anomaly Detection

Supervision Recovery for Time Series Anomaly Detection via Context-Anchored Pairing

Learning Unified Representations of Normalcy for Time Series Anomaly Detection