TY  - JOUR
T1  - Time-Series LPI Signal Classification and Relevance Analysis Using BiLSTM-Attention with LRP
AU  - Park, Kiwan 
AU  - Nam, Haewoon 
JO  - The Journal of Korean Institute of Communications and Information Sciences
PY  - 2024
DA  - 2024/1/1
DO  - 10.7840/kics.2024.49.12.1695
KW  - LPI Signal Classification
KW  - Layer wise Relevance Propagation
KW  - BiLSTM
KW  - Self Attention
KW  - Electronic Warfare (EW)
AB  - This study applies a BiLSTM-Attention model and Layer-wise Relevance Propagation (LRP) to classify and analyze the importance of low probability of intercept (LPI) signals. The goal is to interpret the predictions of a time-series trained model using LRP and effectively identify meaningful input features.
The analysis shows that the model maintains high consistency in its prediction rationale even in the frequency domain, transformed through Fast Fourier Transform (FFT). Experiments across various Signal-to-Noise Ratio (SNR) conditions confirm that the model delivers reliable classification performance while ensuring stable detection of key features through LRP-based interpretation.