Feature learning model based on Feature space constraints and Self-Attention learning

Supervised time series anomaly detection is of significant importance in both machine learning research and industrial applications. Autoencoder-based methods have achieved outstanding performance in this field. However, these methods often suffer from the challenge of limited separation between low-dimensional representations of normal and anomalous data. To address this issue, this paper introduces a model called FFSA, which utilizes feature space constraints and self-attention. The FFSA model is built on the foundation of a temporal network-based autoencoder, featuring an information entropy feature learning module and a self-attention feature learning module. The information entropy feature learning module comprises an MLP mapping layer, layer normalization, and an entropy analysis layer, while the self-attention feature learning module includes layer-wise parameter extraction and concatenation, layer normalization, and a self-attention mechanism. These modules work together to increase the differentiation between normal and anomalous data and to extract more informative low-dimensional representations. Experimental results on multiple public datasets demonstrate that the FFSA model outperforms mainstream anomaly detection models, achieving an impressive F1 score of up to 97% on the KDDCUP dataset.