SRL-AE

• SRL-AE (Sparse Residual LSTM Autoencoder)
• This study proposes a hybrid model, the Sparse Residual LSTM Autoencoder (SRL-AE), combining a Sparse Autoencoder and Residual LSTM to improve anomaly detection in electrocardiogram (ECG) data.

📝 Paper & Description

• Accepted, but not published yet. (Paper Link will be updated soon.)
• 📌 Paper Link
• 📌 Doby's Lab (Blog Description)

📝 Setting

# Clone this Repository
git clone https://github.com/drawcodeboy/SRL-AE.git

# Virtual Environment
python -m venv .venv
.venv\Scripts\activate # Window commands

# Install Packages
pip install -r requirements.txt

# You need to download the dataset(ECG5000), and place it under the data directory.

# <<Train or Test>>

# train LSTM-AE (CPU), if you want train on GPU, use argument "--use-cuda"
python train.py --model=LSTM-AE

# test LSTM-AE (CPU)
python test.py --model=LSTM-AE --weights-filename=LSTM-AE_{epochs}.pth

# train SRL-AE (CPU)
python train.py --model=SRL-AE

# test SRL-AE (CPU)
python test.py --model=SRL-AE --weights-filename=SRL-AE_{epochs}.pth

📁 Dataset

• ECG5000 Dataset

Experiment 1 (Residual LSTM)

• It is interpreted that the decoder intentionally makes reconstruction difficult through residual connections, so normal data can be easily reconstructed, while abnormal data becomes difficult to reconstruct.

	Accuracy	F1-Score	Normal Loss Mean	Loss Gap
Both	0.983	0.972	5.208	16.003
Encoder	0.954	0.923	6.980	15.914
Decoder	0.986	0.977	4.828	17.028

Experiment 2 (Sparse Autoencoder)

• A Sparse Autoencoder was used in the encoder to effectively extract simple patterns from normal data through sparsity constraints. This was demonstrated by conducting a quantitative evaluation using t-SNE.

Experiments 3 (SRL-AE)

• To demonstrate the robustness of the SRL-AE model, experiments were conducted with four models. All models were trained using the same method, and their performance was evaluated. Compared to the conventional LSTM Autoencoder, the SRL-AE model showed slight differences in accuracy and F1-Score but demonstrated significant improvements in robustness metrics, thereby enhancing the reliability of the model's inference results.

	Accuracy	F1-Score	Normal Loss Mean	Loss Gap
LSTM-AE	0.986	0.978	5.240	15.096
Residual LSTM-AE	0.986	0.977	4.828	17.028
Sparse LSTM-AE	0.987	0.979	5.010	16.280
SRL-AE (Ours)	0.986	0.977	4.332	17.320

References

1. Hou, Borui, et al. "LSTM-based auto-encoder model for ECG arrhythmias classification." IEEE Transactions on Instrumentation and Measurement 69.4 (2019): 1232-1240.
2. Farady, Isack, et al. "ECG Anomaly Detection with LSTM-Autoencoder for Heartbeat Analy sis." 2024 IEEE International Conference on Consumer Electronics (ICCE). IEEE, 2024.
3. Dutta, Koustav, et al. "MED-NET: a novel approach to ECG anomaly detection using LSTM auto-encoders." International Journal of Computer Applications in Technology 65.4 (2021): 343-357.
4. Matias, Pedro, et al. "Robust Anomaly Detection in Time Series through Variational AutoEncoders and a Local Similarity Score." Biosignals. 2021.
5. Alamr, Abrar, and Abdelmonim Artoli. "Unsu pervised transformer-based anomaly detection in ECG signals." Algorithms 16.3 (2023): 152.
6. ECG5000-Dataset, “http://timeseriesclassification.com/description.php?Dataset=ECG5000,“ Access Date: 2024/08/23.
7. Wei, Yuanyuan, et al. LSTM-autoencoder-based anomaly detection for indoor air quality time-ser ies data." IEEE Sensors Journal 23.4 (2023): 3787-3800.
8. Kim, Jaeyoung, Mostafa El-Khamy, and Jung won Lee. "Residual LSTM: Design of a deep recurrent architecture for distant speech recognition." arXiv preprint arXiv:1701.03360 (2017).
9. Ng, Andrew. "Sparse autoencoder." CS294A Lecture notes 72.2011 (2011): 1-19.