---

PAI: Preserving Amplitude Information in Representation-Based Time-Series Anomaly Detection

Kang Zhang^1,2, Wei Jian Lau¹, Shoushou Ren¹, LIN Dong¹, Joon Son Chung², Chuanhao Sun^1
1 HUAWEI
² School of Electrical Engineering, KAIST

---

📈 Abstract

Representation-based time-series anomaly detection algorithms significantly outperform alternative methods on diverse anomaly detection tasks. However, we find that they suffer from a major limitation: their learned embeddings are often amplitude-agnostic. Losing amplitude information can degrade performance on amplitude-related anomalies, and this failure is prevalent across existing representation-based methods. To address this issue, we propose a new anomaly scoring scheme named PAI. PAI consists of two complementary modules: a diagnostic module and a final score augmentation function. The diagnostic module compares cosine and Euclidean scoring on the same representation bank to test whether amplitude information is already captured in the learned representation. PAI then computes a pointwise median/MAD deviation score and a local mean-shift score, which are fused with the representation score to produce the final anomaly score. On TSB-AD-U-Eva and TAB UV, PAI improves all four evaluated representation-based methods on every reported metric, with average VUS-PR gains from +0.048 to +0.293. Among all evaluated combinations, PaAno + PAI yields the best VUS-PR at 0.584. Further evaluation with bootstrap confidence intervals, anomaly-type breakdowns, and a TS2Vec input-normalization ablation supports the proposed scheme. These results suggest that explicitly retaining amplitude information is important for representation-based time-series anomaly detection.

paper: https://arxiv.org/abs/2606.08935

🚀 Quick Start

1. Create the environment.

   conda env create -f env.yaml
   conda activate tsbad

2. Download the dataset from TSB-AD.

   Put the TSB-AD files under the repository data/ directory:

   data/
   ├── TSB-AD-U/
   │   └── TSB-AD-U/
   │       └── *.csv
   └── TSB-AD/
       └── Datasets/
           └── File_List/
               └── TSB-AD-U-Eva.csv
   

   No path configuration is needed with this layout.

3. Optionally validate the raw TSB-AD-U Eva files.

   ./reproduce.sh validate_data

4. Generate model prediction anomaly scores.

   To run every model and write the anomaly-score files needed for metric calculation, run:

   ./reproduce.sh generate_anomaly_scores

   This is the longest step. Model prediction anomaly scores are written under outputs/score/ by default, and logs are written under outputs/logs/.

   For a one-file smoke test before the full run:

   source setup_env.sh
   python runners/run_ts2vec_eva.py \
     --score_dir outputs/score/TS2Vec \
     --log_path /tmp/paiad_ts2vec_smoke.log \
     --n_iters 1 --start 0 --end 1

   For full runs, shard the file list using --start and --end. The complete method map is documented in runners/README.md.

5. Calculate metrics to reproduce the tables.

   ./reproduce.sh main_table
   ./reproduce.sh ablation_table

   This writes:

   outputs/aggregates/eva350/pool_means_main_table.csv
   outputs/FULL_COMPARISON_TABLE.md
   

   outputs/aggregates/weight_ablation/sweep_eva350_all.csv
   outputs/WEIGHT_ABLATION_TABLE.md
   

   reproduce.sh already uses conservative threading defaults for stable metric calculation.

📊 Reproduce Tables

The reproduced main results are:

Method	AUC-PR	VUS-PR
DCdetector original	0.0585	0.0974
DCdetector + PAI	0.3751	0.3903
TS2Vec original	0.1885	0.1934
TS2Vec + PAI	0.4475	0.4727
TSPulse_ZS original	0.3241	0.4111
TSPulse_ZS + PAI	0.4311	0.4840
PaAno original	0.4591	0.5090
PaAno + PAI	0.5511	0.5842

🧾 Data Format

• The default file list is data/TSB-AD/Datasets/File_List/TSB-AD-U-Eva.csv.
• The default raw series directory is data/TSB-AD-U/TSB-AD-U/.
• Each series is read as data/TSB-AD-U/TSB-AD-U/<file_name>.csv.
• The final column must be named Label.
• Model prediction anomaly scores are stored as .npy or .npz files under outputs/score/ by default.
• Aggregators accept either full-series or test-only score arrays and align labels automatically.

🙏 Acknowledgements

This release builds on public implementations and benchmark tooling from the following projects:

• TSB-AD
• TS2Vec
• DCdetector
• TSPulse
• PaAno

We thank the authors and maintainers of these projects for making their code and evaluation resources available to the community.