Skip to content

Develop a self-supervised learning algorithm to extract deep features from millions of unlabeled signal data to identify abnormal driving behaviors

Notifications You must be signed in to change notification settings

YunghuiHsu/IEM_DORA

Repository files navigation

Review for Deep Learning with Time Series Data

TS2Vec(Towards Universal Representation of Time Series) 論文筆記

  1. Datasets for training

    • The data source uses Chung-Hao Lee's pre-screened anomalous events(ADDCAT)
      • ts2vec/datasets/DORA/df_event_all_v1.feather
    • data preparation
      • transform_data_CHPrep.ipynb
      • output at : ts2vec/datasets/DORA/
        • data : dora.npz (n, length, dim)
        • meta : dora_meta.csv
    • Pre-processing process for Finetune
      • ts2vec/datasets/prepare_DORA_KA51AG8742.ipynb
        • X_train.npz, X_test.npz, y_train.npz, y_test.npz
  2. Training (in directory ts2vec/)
    We use the self-supervised learning model ts2vec for training and feature extraction.

    • Pretrain commands
      python3 train.py dora  training  --loader DORA --batch-size  16 --repr-dims 320 --epochs 500
    • Get the embedding and dimension reduction
      python3 get_embedding.py  dora  --checkpoint  dora_b16_320d/model_500.pkl  --repr-dims 320  --batch-size 1024  --umap
  3. Analysis and auto-labeling

  • Representation Analysis
    • ts2vec/explore_feature.ipynb
    • Validate appropriate hyperparameter settings
    • Explore and analyze the representation
  • Finetune and get more label
    • ts2vec/verify_label.ipynb
    • Fine-tuning with vehicle_model : "KA51AG8742" datasets (with labels)
      • train the classification model to detect anomalies
    • Use fine-tuning model to infer all data and obtain the obtain labels for abnormal events (automatic labeling)

About

Develop a self-supervised learning algorithm to extract deep features from millions of unlabeled signal data to identify abnormal driving behaviors

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published