PRISM: Powerful Time Series to Image (TS2I) Representations for Multivariate Anomaly Detection

How to run the experiments?

This repository contains the code and artifacts required to reproduce the experimental results reported in the paper, following IJCAI reproducibility guidelines.

Preliminaries

Requirements

Python >= 3.10 (used version: 3.10.13)
uv for environment management

Environment setup

The experimental environment is managed using uv.

uv installation guidelines: https://docs.astral.sh/uv/getting-started/installation/

uv python install 3.10.13
uv python list
uv venv --python 3.10.13
source .venv/bin/activate
python -V
uv run python -m ensurepip --upgrade
uv sync
uv tool install ruff

All dependencies required to run the experiments are specified in pyproject.toml.

Datasets

Data source

All datasets used in this work are taken from the TSB-AD benchmark, which is publicly available:

https://github.com/TheDatumOrg/TSB-AD/tree/main/Datasets

No new datasets are introduced in this repository.

Dataset structure

This work uses the TSB-AD-M (multivariate) datasets. The internal structure of the datasets is preserved exactly as in the original benchmark repository.

The expected local directory layout is:

data/
└── TSB-AD/
    ├── TSB-AD-M/
    │   ├── <dataset_name_1>.csv
    │   ├── <dataset_name_2>.csv
    │   └── ...
    ├── TSB-AD-M.csv
    ├── TSB-AD-M-Eva.csv
    └── TSB-AD-M-Tuning.csv

The CSV files inside TSB-AD-M/ contain multivariate time series and follow the original benchmark specification.

Raw datasets are not redistributed in this repository and must be downloaded from the benchmark source.

Running the experiments

All experiments were executed on NVIDIA GH200 GPUs.

Example commands:

uv run -m cli.entrypoint list-methods
uv run -m cli.entrypoint list-datasets
uv run -m cli.entrypoint one-method --dataset OPPORTUNITY --method RickerWaveletTransform_MeanStdMax
uv run -m cli.entrypoint all-methods --dataset OPPORTUNITY
uv run -m cli.entrypoint full --csv-file cli/tsb-ad-m-datasets-part-1.csv --output-dir output-dir --device cpu

The CLI handles dataset loading, preprocessing, model execution, evaluation, and result serialization.

Hyperparameters and configuration

Model architectures are defined in src/models/.
Hyperparameters are specified in the configuration file config/default.yaml.
The same hyperparameter configuration is used across datasets.

Results

Raw results

Raw experimental outputs are stored in results/raw/.
These files correspond to individual datasets and experimental runs.

Aggregated results

Aggregated results reported in the paper are stored as CSV files in results/.
Aggregation is deterministic and implemented using the script scripts/consolidate.py.
No manual post-processing of results is performed.
results/cnn_ae.csv contains aggregated results for the CNN Autoencoder.
results/resnet18_ae.csv contains aggregated results for the Autoencoder with a ResNet18-based encoder.
All tables/results reported in the paper were generated using scripts provided in the scripts/ directory.
For instance, to inspect fine-grained time complexity results run uv run python scripts/empirical_time_complexity.py

Reproducibility

All code required to run the experiments is included in this repository.
All datasets are publicly available via the TSB-AD benchmark.
Experimental procedures, configurations, and result aggregation are fully specified.
Random seeds and relevant parameters are fixed to ensure determinism; determinism was verified in repeated runs.
The TSB-AD directory from the original benchmark repository is included at the project root to ensure that the same evaluation code and evaluation protocol as in the original benchmark are used. In addition, the original train–test splits are preserved to ensure that the evaluation follows the benchmark’s rigorous protocol.

Tables

Table 1. VUS-PR of 54 detection methods on 14 TSB-AD-M datasets

Best results in bold, second-best underlined.

Method	CATSv2	CreditCard	Daphnet	Exathlon	GECCO	Genesis	LTDB	MSL	OPP	PSM	SMAP	SMD	SWaT	TAO
PRISM(CNN AE + RN-REP)	0.04	0.03	0.08	0.12	0.02	0.01	0.18	0.11	0.06	0.15	0.05	0.05	0.16	0.78
PRISM(CNN AE + SG-REP)	0.11	0.20	0.43	0.86	0.37	0.06	0.28	0.39	0.17	0.17	0.22	0.45	0.33	0.95
PRISM(CNN AE + LG-REP)	0.06	0.14	0.38	0.97	0.29	0.02	0.18	0.18	0.09	0.18	0.16	0.38	0.18	1.00
PRISM(CNN AE + GASF-PCA)	0.10	0.03	0.13	0.19	0.17	0.01	0.23	0.15	0.10	0.16	0.15	0.14	0.29	1.00
PRISM(CNN AE + GASF-MSM)	0.08	0.18	0.07	0.76	0.12	0.09	0.19	0.30	0.04	0.16	0.19	0.13	0.17	1.00
PRISM(CNN AE + GADF-PCA)	0.07	0.03	0.05	0.58	0.03	0.01	0.26	0.11	0.16	0.16	0.11	0.11	0.17	0.65
PRISM(CNN AE + GADF-MSM)	0.05	0.02	0.05	0.45	0.02	0.01	0.19	0.31	0.09	0.13	0.15	0.06	0.15	0.64
PRISM(CNN AE + MTF-PCA)	0.04	0.03	0.06	0.41	0.02	0.01	0.16	0.13	0.11	0.15	0.16	0.06	0.18	0.79
PRISM(CNN AE + MTF-MSM)	0.06	0.03	0.07	0.84	0.02	0.10	0.18	0.25	0.07	0.14	0.18	0.07	0.14	0.74
PRISM(CNN AE + RP-PCA)	0.04	0.03	0.16	0.35	0.06	0.02	0.21	0.19	0.08	0.17	0.07	0.08	0.21	0.77
PRISM(CNN AE + RP-MSM)	0.05	0.03	0.32	0.78	0.05	0.02	0.23	0.17	0.11	0.15	0.07	0.11	0.15	0.78
PRISM(CNN AE + MWT-PCA)	0.07	0.04	0.43	0.95	0.09	0.01	0.16	0.23	0.18	0.25	0.22	0.44	0.46	0.99
PRISM(CNN AE + MWT-MSM)	0.23	0.16	0.35	0.86	0.13	0.45	0.29	0.45	0.13	0.28	0.28	0.46	0.60	0.91
PRISM(CNN AE + RWT-PCA)	0.07	0.05	0.28	0.97	0.08	0.01	0.19	0.24	0.10	0.23	0.25	0.32	0.27	0.82
PRISM(CNN AE + RWT-MSM)	0.16	0.17	0.32	0.86	0.22	0.48	0.35	0.43	0.29	0.26	0.26	0.30	0.26	0.87
PRISM(ResNet18 AE + RN-REP)	0.04	0.03	0.08	0.12	0.02	0.01	0.18	0.10	0.06	0.15	0.05	0.05	0.16	0.78
PRISM(ResNet18 AE + SG-REP)	0.10	0.17	0.35	0.98	0.31	0.80	0.38	0.43	0.21	0.15	0.32	0.44	0.64	0.89
PRISM(ResNet18 AE + LG-REP)	0.06	0.14	0.45	0.98	0.43	0.74	0.28	0.23	0.39	0.25	0.24	0.40	0.48	1.00
PRISM(ResNet18 AE + GASF-PCA)	0.12	0.03	0.29	0.29	0.20	0.14	0.37	0.23	0.06	0.18	0.22	0.18	0.32	1.00
PRISM(ResNet18 AE + GASF-MSM)	0.10	0.19	0.15	0.74	0.11	0.19	0.30	0.24	0.06	0.16	0.24	0.15	0.18	1.00
PRISM(ResNet18 AE + GADF-PCA)	0.12	0.02	0.44	0.81	0.07	0.03	0.35	0.15	0.06	0.17	0.18	0.15	0.35	0.75
PRISM(ResNet18 AE + GADF-MSM)	0.06	0.02	0.21	0.66	0.07	0.02	0.31	0.33	0.07	0.16	0.19	0.08	0.17	0.62
PRISM(ResNet18 AE + MTF-PCA)	0.05	0.03	0.35	0.50	0.03	0.01	0.20	0.14	0.05	0.16	0.22	0.11	0.15	0.80
PRISM(ResNet18 AE + MTF-MSM)	0.07	0.03	0.12	0.76	0.05	0.09	0.19	0.22	0.06	0.15	0.21	0.09	0.16	0.71
PRISM(ResNet18 AE + RP-PCA)	0.04	0.03	0.24	0.49	0.03	0.01	0.19	0.26	0.06	0.16	0.18	0.08	0.16	0.79
PRISM(ResNet18 AE + RP-MSM)	0.06	0.03	0.33	0.76	0.03	0.01	0.24	0.29	0.08	0.14	0.23	0.14	0.15	0.78
PRISM(ResNet18 AE + MWT-PCA)	0.10	0.06	0.47	0.97	0.09	0.16	0.21	0.27	0.20	0.28	0.25	0.45	0.49	0.99
PRISM(ResNet18 AE + MWT-MSM)	0.25	0.13	0.41	0.86	0.12	0.51	0.37	0.48	0.13	0.31	0.35	0.48	0.59	0.93
PRISM(ResNet18 AE + RWT-PCA)	0.09	0.05	0.38	0.98	0.10	0.46	0.20	0.29	0.17	0.27	0.29	0.39	0.28	0.81
PRISM(ResNet18 AE + RWT-MSM)	0.18	0.16	0.35	0.86	0.16	0.60	0.42	0.42	0.14	0.34	0.34	0.45	0.59	0.86
Baseline Methods
xLSTMAD	0.10	0.06	0.50	0.91	0.04	0.01	0.40	0.38	0.16	0.18	0.35	0.35	0.34	0.80
CNN	0.08	0.02	0.21	0.68	0.03	0.10	0.33	0.35	0.16	0.22	0.19	0.35	0.41	1.00
OmniAnomaly	0.04	0.02	0.34	0.84	0.02	0.00	0.44	0.22	0.18	0.16	0.12	0.17	0.15	0.81
PCA	0.12	0.10	0.13	0.95	0.20	0.02	0.24	0.15	0.30	0.16	0.09	0.36	0.45	1.00
LSTMAD	0.04	0.02	0.31	0.82	0.02	0.04	0.30	0.22	0.17	0.24	0.16	0.33	0.16	0.99
USAD	0.04	0.02	0.34	0.84	0.02	0.00	0.41	0.23	0.18	0.19	0.11	0.16	0.15	0.81
AutoEncoder	0.06	0.03	0.13	0.91	0.05	0.01	0.21	0.22	0.14	0.28	0.13	0.30	0.58	1.00
KMeansAD	0.12	0.02	0.30	0.37	0.06	0.89	0.41	0.44	0.06	0.21	0.38	0.36	0.16	0.86
CBLOF	0.06	0.03	0.10	0.86	0.03	0.02	0.20	0.21	0.14	0.19	0.14	0.22	0.29	1.00
MCD	0.13	0.06	0.14	0.80	0.03	0.06	0.21	0.23	0.17	0.26	0.10	0.26	0.54	1.00
OCSVM	0.08	0.02	0.06	0.83	0.04	0.08	0.20	0.22	0.12	0.19	0.12	0.28	0.44	0.81
Donut	0.07	0.02	0.17	0.66	0.03	0.18	0.26	0.30	0.15	0.20	0.18	0.19	0.44	0.75
RobustPCA	0.04	0.02	0.06	0.77	0.02	0.00	0.23	0.22	0.13	0.12	0.07	0.10	0.12	1.00
FITS	0.13	0.02	0.33	0.63	0.03	0.10	0.23	0.17	0.05	0.13	0.08	0.17	0.15	0.78
OFA	0.13	0.02	0.31	0.58	0.04	0.22	0.29	0.14	0.05	0.17	0.08	0.17	0.12	0.78
EIF	0.06	0.02	0.15	0.41	0.04	0.06	0.19	0.18	0.10	0.18	0.13	0.32	0.32	0.89
COPOD	0.05	0.05	0.11	0.40	0.04	0.08	0.21	0.21	0.17	0.20	0.10	0.19	0.31	0.99
IForest	0.05	0.03	0.13	0.35	0.04	0.08	0.21	0.21	0.18	0.19	0.09	0.26	0.39	0.93
HBOS	0.05	0.04	0.15	0.32	0.04	0.08	0.21	0.23	0.17	0.17	0.09	0.25	0.30	0.83
TimesNet	0.07	0.02	0.27	0.42	0.03	0.02	0.27	0.17	0.06	0.14	0.09	0.14	0.14	0.79
KNN	0.07	0.02	0.25	0.33	0.11	0.04	0.19	0.18	0.06	0.12	0.12	0.30	0.11	0.78
TranAD	0.04	0.02	0.31	0.10	0.02	0.04	0.26	0.24	0.16	0.23	0.09	0.30	0.15	0.81
LOF	0.05	0.02	0.11	0.16	0.13	0.08	0.19	0.14	0.10	0.15	0.09	0.16	0.15	0.79
AnomalyTransformer	0.03	0.02	0.07	0.10	0.02	0.01	0.21	0.12	0.07	0.21	0.06	0.07	0.18	0.77

Table 2. Mean performance across 9 metrics for 54 detection methods on 14 TSB-AD datasets

Best results in bold, second-best underlined.

Method	AUC-PR	AUC-ROC	VUS-PR	VUS-ROC	Standard-F1	PA-F1	Event-based-F1	R-based-F1	Affiliation-F1
PRISM(CNN AE + RN-REP)	0.14	0.50	0.15	0.52	0.21	0.94	0.39	0.33	0.80
PRISM(CNN AE + SG-REP)	0.49	0.75	0.48	0.76	0.53	0.90	0.79	0.42	0.93
PRISM(CNN AE + LG-REP)	0.45	0.73	0.44	0.75	0.50	0.93	0.76	0.40	0.92
PRISM(CNN AE + GASF-PCA)	0.24	0.61	0.23	0.63	0.31	0.94	0.62	0.37	0.88
PRISM(CNN AE + GASF-MSM)	0.38	0.70	0.37	0.71	0.43	0.93	0.73	0.36	0.90
PRISM(CNN AE + GADF-PCA)	0.27	0.53	0.26	0.55	0.34	0.88	0.64	0.40	0.89
PRISM(CNN AE + GADF-MSM)	0.26	0.54	0.25	0.55	0.37	0.90	0.62	0.42	0.86
PRISM(CNN AE + MTF-PCA)	0.24	0.57	0.24	0.59	0.30	0.93	0.63	0.39	0.86
PRISM(CNN AE + MTF-MSM)	0.35	0.63	0.35	0.64	0.42	0.92	0.66	0.39	0.88
PRISM(CNN AE + RP-PCA)	0.22	0.57	0.22	0.58	0.28	0.86	0.53	0.30	0.84
PRISM(CNN AE + RP-MSM)	0.31	0.62	0.31	0.63	0.37	0.83	0.59	0.30	0.88
PRISM(CNN AE + MWT-PCA)	0.49	0.76	0.48	0.77	0.53	0.89	0.77	0.41	0.93
PRISM(CNN AE + MWT-MSM)	0.51	0.74	0.51	0.75	0.56	0.92	0.85	0.45	0.94
PRISM(CNN AE + RWT-PCA)	0.45	0.72	0.45	0.73	0.51	0.90	0.75	0.40	0.92
PRISM(CNN AE + RWT-MSM)	0.47	0.74	0.47	0.75	0.54	0.92	0.82	0.42	0.93
PRISM(ResNet18 AE + RN-REP)	0.14	0.50	0.15	0.52	0.21	0.93	0.32	0.32	0.77
PRISM(ResNet18 AE + SG-REP)	0.52	0.76	0.52	0.77	0.57	0.93	0.81	0.41	0.93
PRISM(ResNet18 AE + LG-REP)	0.50	0.80	0.49	0.81	0.54	0.96	0.80	0.39	0.93
PRISM(ResNet18 AE + GASF-PCA)	0.27	0.65	0.27	0.66	0.35	0.96	0.70	0.37	0.90
PRISM(ResNet18 AE + GASF-MSM)	0.38	0.73	0.37	0.74	0.44	0.94	0.77	0.36	0.92
PRISM(ResNet18 AE + GADF-PCA)	0.30	0.62	0.29	0.63	0.36	0.94	0.67	0.37	0.89
PRISM(ResNet18 AE + GADF-MSM)	0.30	0.58	0.29	0.60	0.39	0.94	0.67	0.38	0.88
PRISM(ResNet18 AE + MTF-PCA)	0.23	0.58	0.23	0.59	0.30	0.93	0.60	0.37	0.86
PRISM(ResNet18 AE + MTF-MSM)	0.33	0.63	0.33	0.64	0.41	0.92	0.62	0.36	0.87
PRISM(ResNet18 AE + RP-PCA)	0.22	0.55	0.22	0.56	0.29	0.89	0.59	0.35	0.85
PRISM(ResNet18 AE + RP-MSM)	0.34	0.62	0.34	0.63	0.41	0.89	0.68	0.38	0.89
PRISM(ResNet18 AE + MWT-PCA)	0.50	0.76	0.50	0.77	0.55	0.94	0.84	0.42	0.95
PRISM(ResNet18 AE + MWT-MSM)	0.53	0.75	0.53	0.76	0.58	0.92	0.86	0.45	0.95
PRISM(ResNet18 AE + RWT-PCA)	0.48	0.73	0.48	0.74	0.54	0.92	0.81	0.40	0.93
PRISM(ResNet18 AE + RWT-MSM)	0.51	0.74	0.50	0.75	0.57	0.92	0.83	0.42	0.94
Baseline Methods
xLSTMAD	0.40	0.74	0.48	0.78	0.48	0.57	0.50	0.43	0.82
CNN	0.36	0.75	0.35	0.77	0.41	0.80	0.68	0.41	0.88
OmniAnomaly	0.28	0.66	0.33	0.70	0.33	0.55	0.41	0.39	0.81
PCA	0.36	0.73	0.36	0.78	0.43	0.88	0.67	0.32	0.87
LSTMAD	0.36	0.73	0.36	0.77	0.41	0.83	0.68	0.43	0.89
USAD	0.27	0.65	0.32	0.69	0.32	0.53	0.40	0.39	0.80
AutoEncoder	0.36	0.70	0.35	0.72	0.40	0.64	0.51	0.33	0.82
KMeansAD	0.28	0.70	0.33	0.74	0.35	0.70	0.54	0.36	0.83
CBLOF	0.33	0.69	0.32	0.72	0.37	0.71	0.53	0.36	0.83
MCD	0.32	0.71	0.32	0.75	0.39	0.53	0.39	0.23	0.80
OCSVM	0.27	0.63	0.31	0.70	0.32	0.55	0.48	0.35	0.82
Donut	0.24	0.78	0.32	0.86	0.34	0.63	0.44	0.26	0.98
RobustPCA	0.28	0.60	0.28	0.63	0.33	0.62	0.46	0.37	0.82
FITS	0.17	0.60	0.24	0.69	0.24	0.70	0.32	0.17	0.82
OFA	0.14	0.57	0.22	0.65	0.21	0.73	0.39	0.15	0.83
EIF	0.22	0.69	0.25	0.73	0.30	0.77	0.51	0.30	0.83
COPOD	0.23	0.68	0.23	0.72	0.31	0.76	0.47	0.27	0.82
IForest	0.22	0.68	0.23	0.71	0.29	0.71	0.44	0.26	0.81
HBOS	0.18	0.65	0.22	0.70	0.27	0.72	0.47	0.28	0.82
TimesNet	0.14	0.58	0.21	0.67	0.22	0.65	0.31	0.18	0.83
KNN	0.16	0.54	0.21	0.63	0.22	0.73	0.52	0.24	0.81
TranAD	0.15	0.60	0.20	0.66	0.23	0.68	0.40	0.22	0.79
LOF	0.11	0.54	0.16	0.62	0.17	0.60	0.36	0.16	0.77
AnomalyTransformer	0.08	0.52	0.14	0.58	0.13	0.51	0.32	0.14	0.73

Table 3. Mean performance across 10 metrics of 15 PRISM's TS2I projection schemes

Best results in bold, second-best underlined.

Projection scheme	AUC-PR	AUC-ROC	VUS-PR	VUS-ROC	Standard-F1	PA-F1	Event-based-F1	R-based-F1	Affiliation-F1	Time (s / 100 img)
RN-REP	0.14	0.50	0.15	0.52	0.21	0.94	0.34	0.32	0.77	0.20
SG-REP	0.51	0.76	0.51	0.77	0.56	0.92	0.81	0.41	0.93	0.18
LG-REP	0.49	0.78	0.48	0.79	0.53	0.95	0.79	0.39	0.93	1.40
GASF-PCA	0.26	0.64	0.26	0.65	0.34	0.95	0.68	0.37	0.89	0.32
GASF-MSM	0.38	0.72	0.37	0.73	0.44	0.94	0.76	0.36	0.92	0.32
GADF-PCA	0.29	0.59	0.28	0.61	0.36	0.93	0.67	0.38	0.89	0.34
GADF-MSM	0.29	0.57	0.28	0.59	0.38	0.93	0.66	0.39	0.88	0.33
MTF-PCA	0.23	0.58	0.23	0.59	0.30	0.93	0.60	0.37	0.86	0.32
MTF-MSM	0.33	0.63	0.33	0.64	0.41	0.92	0.63	0.37	0.87	0.32
RP-PCA	0.22	0.55	0.22	0.57	0.29	0.88	0.58	0.33	0.85	0.25
RP-MSM	0.34	0.62	0.34	0.63	0.40	0.87	0.66	0.36	0.89	0.24
MWT-PCA	0.50	0.76	0.50	0.77	0.54	0.93	0.82	0.42	0.94	0.40
MWT-MSM	0.53	0.76	0.52	0.76	0.58	0.92	0.86	0.45	0.95	0.37
RWT-PCA	0.47	0.73	0.47	0.74	0.53	0.92	0.79	0.40	0.93	0.39
RWT-MSM	0.50	0.74	0.50	0.75	0.56	0.92	0.83	0.42	0.94	0.37

Table 4. Mean performance across 10 metrics of 3 TL strategies for ResNet18 AE

Best results in bold, second-best underlined.

Strategy	AUC-PR	AUC-ROC	VUS-PR	VUS-ROC	Std-F1	PA-F1	Evt-F1	R-F1	Aff-F1	Time (s)
FE	0.36	0.66	0.35	0.67	0.42	0.93	0.70	0.38	0.89	94
PU	0.37	0.67	0.37	0.68	0.43	0.93	0.71	0.38	0.90	141
DLR	0.38	0.68	0.38	0.69	0.44	0.93	0.71	0.39	0.90	167

Labels:

FE - Frozen Encoder
PEU - Progressive Encoder Unfreezing
DLR - Differential Learning Rates

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
TSB_AD		TSB_AD
cli		cli
configs		configs
results		results
scripts		scripts
src		src
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Navigation Menu

Repository files navigation

PRISM: Powerful Time Series to Image (TS2I) Representations for Multivariate Anomaly Detection

How to run the experiments?

Preliminaries

Requirements

Environment setup

Datasets

Data source

Dataset structure

Running the experiments

Hyperparameters and configuration

Results

Raw results

Aggregated results

Reproducibility

Tables

Table 1. VUS-PR of 54 detection methods on 14 TSB-AD-M datasets

Table 2. Mean performance across 9 metrics for 54 detection methods on 14 TSB-AD datasets

Table 3. Mean performance across 10 metrics of 15 PRISM's TS2I projection schemes

Table 4. Mean performance across 10 metrics of 3 TL strategies for ResNet18 AE

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

PRISM: Powerful Time Series to Image (TS2I) Representations for Multivariate Anomaly Detection

How to run the experiments?

Preliminaries

Requirements

Environment setup

Datasets

Data source

Dataset structure

Running the experiments

Hyperparameters and configuration

Results

Raw results

Aggregated results

Reproducibility

Tables

Table 1. VUS-PR of 54 detection methods on 14 TSB-AD-M datasets

Table 2. Mean performance across 9 metrics for 54 detection methods on 14 TSB-AD datasets

Table 3. Mean performance across 10 metrics of 15 PRISM's TS2I projection schemes

Table 4. Mean performance across 10 metrics of 3 TL strategies for ResNet18 AE

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages