Direction-aware convolutional autoencoder based on positional encoding for one-dimensional anomaly detection

One-dimensional anomaly detection remains challenging owing to existing methods inadequately modeling interactions between widely separated elements and neglecting local interaction patterns. To address these limitations, this study proposes a space transformation-based direction-aware convolutional autoencoder framework using positional encoding (SDP-CAE) for one-dimensional anomaly detection. Unlike previous approaches, SDP-CAE uniquely applies multiple space-filling curves to transform one-dimensional data into two-dimensional data, effectively capturing rich local feature interactions and overcoming the limitations posed by traditional one-dimensional modeling. Space-filling curves can enhance the interactions among different elements of one-dimensional data to enrich local feature patterns in a two-dimensional space. Furthermore, we introduce an asymmetric two-stream convolutional autoencoder architecture that employs horizontal and vertical convolution operations to explicitly capture direction-specific interactions within the transformed data. This architecture significantly improves anomaly detection by modeling interactions that are sensitive to the local spatial context. Positional encoding (PE), employed as an auxiliary mechanism, enhances the representation of high-frequency details to further prompt the sensitivity of the model to subtle anomalies. Extensive experiments conducted on benchmark datasets demonstrate that our method significantly outperforms state-of-the-art methods, clearly validating the effectiveness and necessity of the proposed space transformation and direction-aware modeling mechanisms.