🔮 LSTM-Based Prediction API (Graduation Project)

📌 Overview

This project is a time-series prediction API built using LSTM (Long Short-Term Memory) models to detect potential failures based on sequential data.

It is part of a larger graduation project for predictive analysis, where the system processes timestamped input data and returns a prediction indicating whether a failure is expected.

---

🚀 Features

• 📊 Time-series preprocessing with dynamic windowing
• 🧠 LSTM model for sequential prediction
• ⚡ FastAPI-based REST API
• 🔄 Model & scaler loading with caching for performance
• 📈 Custom F1-score loss function

---

🏗️ Tech Stack

• Python
• FastAPI
• TensorFlow / Keras
• NumPy & Pandas
• Scikit-learn (Scaler)
• Uvicorn

---

⚙️ How It Works

1. Input Data

The model receives data in the form:

• Timestamp (string)
• Feature values (floats)

Example:

{
  "model_name": "model1",
  "data": [
    ["2024-01-01 10:00:00", 0.5, 1.2],
    ["2024-01-01 10:01:00", 0.6, 1.1]
  ]
}

---

2. Preprocessing

• Converts timestamps to milliseconds
• Splits data into time-based windows (20 minutes)
• Applies padding to ensure fixed input size
• Scales features using a pre-trained scaler

---

3. Prediction

• Processed data is passed to the LSTM model

• Output is rounded to binary (0 or 1)

• Final result:

  • true → failure expected
  • false → normal behavior

---

🔌 API Endpoint

POST /predict/

Request Body:

{
  "model_name": "your_model_name",
  "data": [
    ["timestamp", feature1, feature2, ...]
  ]
}

Response:

{
  "prediction": true
}

---

📂 Project Structure

├── models/        # Saved LSTM models (.h5)
├── scalers/       # Saved scalers (.pkl)
├── main.py        # FastAPI application

---

▶️ Running the Project

1. Install dependencies

pip install -r requirements.txt

2. Run the server

python main.py

or

uvicorn main:app --reload

---

🧠 Model Details

• Architecture: LSTM (for time-series data)
• Input: Sequential windows of sensor/features data
• Output: Binary classification (failure / no failure)
• Loss Function: Custom F1 Loss for better imbalance handling

---

📌 Notes

• Models and scalers must exist in:

  • models/
  • scalers/

• Each model requires a corresponding scaler with the same name

• Input data must cover sufficient time range to form valid windows

---

👩‍💻 My Contribution

• Designed and implemented the prediction API using FastAPI
• Built the data preprocessing pipeline for time-series windowing
• Integrated LSTM model inference with scaling and padding
• Implemented model caching for performance optimization

---

📈 Future Improvements

• Add real-time streaming support
• Improve model accuracy with advanced architectures
• Add visualization dashboard for predictions

---

📄 License

This project is part of an academic graduation project and is shared for educational purposes.