Optimizing Feature Extraction From Sequential Sensor Data Using Neural Network-Based Image Transformation

Le résumé fourni par la source

The transformation of time-series data into images and the subsequent application of Convolutional Neural Networks (CNNs) has recently emerged as a promising approach for sensor signal analysis, owing to CNNs' superior capability in extracting spatially encoded patterns and complex hierarchical features. While prior studies have shown that CNNbased learning on image-transformed time-series data can surpass the performance of traditional direct learning from raw time-series data, this advantage is not consistently observed in all cases. We investigate key issues, including potential quantization error and loss of information during the transformation process, CNN's inherent limitations in capturing temporal dependencies and long-term correlations, the mismatch between transformation methods and specific signal patterns, and the risk of overfitting due to the complexity imbalance between the data and model architecture. The experimental results demonstrate that image-based learning achieves higher accuracy and robustness, especially in complex environments with noise and overlapping signals.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.