Spatial-temporal attention and information reinforcement network for machine remaining useful life prediction

• Published in: IEEE sensors journal 2024-02, Vol.24 (3), p.1-1
• Main Authors: Li, Xuanlin, Hu, Yawei, Wang, Hang, Liu, Yongbin, Liu, Xianzeng, Cao, Zheng
• Format: Article
• Language: English
• Subjects: Deep learning; Feature extraction; Information reinforcement; Life prediction; long short-term memory; Modules; Performance degradation; Remaining useful life prediction; Spatial-temporal attention; Useful life
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2023.3342884

Remaining useful life (RUL) prediction is crucial for enhancing equipment reliability and safety in industry. In recent years, deep learning techniques, particularly those based on Long Short-Term Memory (LSTM) networks, have been widely used in the field. However, the performance of LSTM-based models is often constrained by the loss of early time dependence. Additionally, learning the mapping relationships between large multi-sensor data and features poses a challenge. To address these issues, a novel multi-sensor data-driven RUL prediction method named the Spatial-Temporal Attention and Information Reinforcement Network (STAIRnet) is proposed. First, the spatial-temporal attention module adaptively weights and encodes the original signal in both temporal and spatial dimensions. Second, the feature extraction module extracts hidden features from the weighted data, while the lookback mechanism filters the hidden states. Following that, the information reinforcement module decodes the encoded information and supplements, which reinforces the hidden features to improve the model performance. Finally, degraded features are mapped to specific RUL values. The effectiveness of STAIRnet was validated using a commonly used dataset. The results demonstrated that the proposed method outperformed other approaches in terms of prediction accuracy and computational efficiency.