Time-Series Anomaly Detector

A simple end-to-end example of detecting anomalies in a univariate time series using a PyTorch LSTM autoencoder and visualizing results in a Streamlit dashboard.

---

🚀 Features

• Synthetic data generator: builds a noisy sinusoidal series with random “spikes”.
• LSTM Autoencoder: learns to reconstruct normal sequences, flags outliers by reconstruction error.
• Thresholding: computes a data-driven anomaly threshold (mean + 3·std of train errors).
• CLI scripts:
  • train.py — generate data & train the model
  • detect.py — compute reconstruction errors & label anomalies
• Streamlit dashboard: interactive charts of original series, reconstruction error, and anomaly markers.
• Dockerized: one-step build & launch.

---

📊 Results Preview

---

🔧 Prerequisites

• Python 3.8+
• (Optional) Docker & Docker Engine

---

🛠️ Installation

1. Clone the repo

   git clone https://github.com/ion-linti/ts-anomaly-detector.git
   cd ts-anomaly-detector

2. Create & activate a virtual environment

   python3 -m venv venv
   source venv/bin/activate    # Linux/macOS
   # .\venv\Scripts\activate  # Windows PowerShell

3. Install dependencies

   pip install --upgrade pip
   pip install -r requirements.txt

---

📈 Training

python train.py

---

🔍 Detection

python detect.py

---

📊 Streamlit Dashboard

streamlit run app.py

Open http://localhost:8501 and click to refresh plots.

---

🐋 Docker Usage

docker build -t ts-anomaly-detector .
docker run --rm -p 8501:8501 ts-anomaly-detector

---

🤝 Contributing

Fork → branch → PR → merge.

---

📄 License

MIT License