Enhancing Unsupervised Anomaly Detection and Early Warning With Dual‐Attention <scp>LSTM</scp> ‐ <scp>AdvAE</scp>

ABSTRACT Unsupervised Anomaly Detection (UAD) has shown great potential in a variety of real‐world applications, particularly in fields such as information science. However, the current repertoire of unsupervised techniques remains limited, posing challenges for effective fault diagnosis. This paper introduces a novel network structure for enhancing unsupervised anomaly detection and early warning, designed for multi‐dimensional time series features and based on the Dual‐Attention LSTM‐AdvAE model. Our approach utilizes the Temporal‐Feature Dual‐Attention Module module to assign weighted attention to features and along time dimensions. These weighted features are then fed into Long Short‐Term Memory (LSTM) layers, which excel at capturing long‐term dependencies within the sequence data, thus facilitating robust feature extraction. Furthermore, the extracted features are encoded into a latent vector via an autoencoder's encoder component, which serves to compress the data into a lower‐dimensional representation. This vector undergoes adversarial training within an improved adversarial autoencoder framework. Finally, the anomaly score is derived from the error in the reconstructed data obtained by the decoder. We have demonstrated through comparative experiments on two public datasets that the average performance of our method surpasses that of previous anomaly detection methods. Simultaneously, we conducted ablation experiments on a real‐operating ice machine dataset to verify the importance of each model component for the final anomaly detection results. Lastly, the model's effectiveness in providing early warnings for ice machine failures was validated on this dataset.