TimeRadar: A Domain-Rotatable Foundation Model for Time Series Anomaly Detection

---

Overview

We introduce TimeRadar, an innovative time series foundation model (TSFM) built in a fractional time–frequency domain to support generalist time series anomaly detection (TSAD) across diverse unseen datasets. Our key insight is that rotating a time series into a data-dependent fractional time–frequency representation can adaptively differentiate normal and abnormal signals across different datasets. To this end, we propose a novel component, Fractionally modulated Time-Frequency Reconstruction (FTFRecon), which leverages a learnable fractional order to rotate the time series to the most pronounced angle between the continuous time and frequency domains for accurate data reconstruction. This design enables adaptive reconstruction in an optimal time–frequency domain for each input, effectively distinguishing unbounded abnormal patterns from regular ones across datasets, including previously unseen datasets. To further capture local abnormalities that may not be reflected by global reconstruction, we introduce a Contextual Deviation Learning (CDL) component, which models the local deviation of the input relative to its contextual time series data in the rotatable domain.

📚 Data Preparation

The datasets can be downloaded from the following link Google Drive

Place the extracted Monash+ and evaluation_dataset directories under ./dataset:

dataset/
├── Monash+/
└── evaluation_dataset/
    ├── DETECT_META.csv
    └── data/

If you download the original Monash data instead of the processed Monash+ data, first place it at ./dataset/Monash and generate Monash+ with:

sh ./anomaly_inject_Monash/gen_Monash.sh

⚙️ Installation

Install the dependencies listed in requirements.txt:

pip install -r requirements.txt

🏋️ Pretraining

To pretrain TimeRadar on Monash+, run:

bash ./scripts/pretrain/Monash_ADD.sh

The resulting checkpoint is saved under ./checkpoints.

🔥 Few-shot Fine-tuning

Few-shot fine-tuning uses a small percentage of labeled training data from a target dataset. Set --is_training 0, --is_finetuning 1, and --is_zeroshot 0. Change --percentage to evaluate another few-shot setting.

🧊 Zero-shot Evaluation

For zero-shot anomaly detection, set --is_training 0, --is_finetuning 0, and --is_zeroshot 1. For example, evaluate TimeRadar on MSL with:

bash ./scripts/anomaly_detection/MSL.sh

We also provide a ready-to-use pretrained TimeRadar model under ./TimeRadar, packaged in the Hugging Face custom-model format similarly to ./DADA. It can be loaded directly with AutoModel.from_pretrained(..., trust_remote_code=True); see the model README for details.

🤖 Evaluating Other Foundation Models

To evaluate other advanced foundation models such as Chronos-Bolt, download the corresponding pretrained weights and place them under the local path expected by the implementation (for Chronos-Bolt, ./models/chronos/chronos-bolt-base). Use the anomaly-detection scripts above and replace the task and model arguments with:

--task_name anomaly_detection_chronos \
--model Chronos-bolt

Forecasting-based foundation models require the TrainSegLoaderAddPre data-loader mapping documented in data_provider/data_factory.py.

Citation

If you find this repository useful in your research or applications, please cite our paper:

@misc{he2026timeradar,
  title         = {TimeRadar: A Domain-Rotatable Foundation Model for Time Series Anomaly Detection},
  author        = {Hui He and Hezhe Qiao and Yutong Chen and Kun Yi and Guansong Pang},
  year          = {2026},
  eprint        = {2602.19068},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2602.19068}
}

Acknowledgement

We appreciate the following GitHub repos for providing valuable code bases and efforts.

• DADA [repo]
• SEMPO [repo]
• Time-MoE [repo]
• chronos-forecasting [repo]
• TimesFM [repo]
• CATCH [repo]
• Large-Time-Series-Model [repo]