AnomaLLM

AnomaLLM is a Python SDK for explainable anomaly detection on multivariate time-series data. It includes an LSTM autoencoder for reconstruction-based anomaly scoring, Granger-style causality analysis, and an LLM-backed incident reporter.

Key Features

• LSTM autoencoder for anomaly scoring
• Causality tracing for root-cause analysis
• LLM-generated incident summaries
• Streamlit dashboard for visual inspection
• Repo-local sample C-MAPSS data for reproducible runs

Requirements

• Python 3.8+
• torch
• pandas
• numpy
• scikit-learn
• statsmodels
• networkx
• openai
• Optional for the dashboard: streamlit, plotly

API Keys

If you want the LLM-generated report, set OPENAI_API_KEY in your environment.

export OPENAI_API_KEY="your-api-key-here"

Installation

git clone https://github.com/ayush585/AnomaLLM.git
cd AnomaLLM
python -m venv venv
# Windows:
# venv\Scripts\activate
pip install -e .

Run

Train the model:

python train.py

Run the CLI demo:

python demo.py

Run the Kaggle-style demo:

python demo_kaggle.py

Launch the dashboard:

streamlit run dashboard.py

Project Structure

anomallm/                 SDK package
dashboard.py              Streamlit dashboard
demo.py                   CLI demo with webhook forwarding
demo_kaggle.py            Dataset-blind demonstration
train.py                  Training entry point
setup.py                  Packaging metadata
lstm_v1.pth               Optional generated model checkpoint
train_FD001.txt           Repo-local sample training data
test_FD001.txt            Repo-local sample test data
RUL_FD001.txt             Repo-local reference data
data/                     Optional user data directory

Notes

• The bundled scripts resolve sample datasets relative to the repository root.
• The data/ directory is optional and no longer holds machine-specific pointer files.
• lstm_v1.pth is generated by train.py and is not required to exist before running demo.py.