Wipoba/lstm-sequence-anomaly-detection

LSTM Sequence Anomaly Detection

This repository contains a trained LSTM Autoencoder model used for anomaly detection in sequential data.

Model Overview

The model is an LSTM-based Autoencoder, which is trained to reconstruct sequential data. Anomalies can be detected by measuring the reconstruction error: sequences with high reconstruction loss are considered anomalous.

Model Architecture:

• Type: LSTM Autoencoder
• Encoder: LSTM with 64 units
• Decoder: LSTM with 64 units
• Output Layer: TimeDistributed Dense with softmax activation

Training Details:

• Optimizer: Adam
• Loss Function: Categorical Crossentropy
• Batch Size: 128
• Epochs: 50
• Early Stopping: Patience = 5

Performance Metrics:

• Validation Loss (Mean): {val_loss.mean():.4f}
• Validation Loss (Std Dev): {val_loss.std():.4f}
• Training/Validation Loss: See the graph below.

How to Use:

1. Load the model using tf.keras.models.load_model().
2. Use the model to detect anomalies in new sequences.
3. Calculate reconstruction loss for each sequence, and use a threshold to classify anomalies.

Files in this Repo:

• model.h5: The trained LSTM Autoencoder model.
• metrics.txt: Performance metrics for the model.
• loss_graph.png: Loss curve during training.
• README.md: This file.
• model_config.json: Model architecture details.
• training_config.json: Training hyperparameters.