基于异常敏感度特征编码与知识蒸馏的变工况工业过程监测方法、装置、设备及介质
By collecting and analyzing physical state parameters in industrial processes under varying operating conditions, and utilizing anomaly-sensitive long short-term memory networks and cross-attention mechanisms, the problem of accurate perception and knowledge transfer of hidden and subtle faults under varying operating conditions has been solved, achieving efficient fault monitoring and continuous learning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-17
AI Technical Summary
In non-stationary industrial processes with varying operating conditions, existing technologies struggle to accurately detect subtle, hidden faults, and the transfer of monitoring knowledge between different production states is also difficult.
By collecting physical state parameters of non-stationary industrial processes under varying operating conditions, stationary feature data is extracted using stationary subspace analysis and input into an anomaly-sensitive long short-term memory network for non-stationary temporal feature encoding. Feature fusion is performed by combining cross-attention mechanism and dynamic memory bank, and the weights of global time dimension and feature dimension are reconstructed using cross-attention temporal distillation mechanism. Finally, the industrial process state determination result is generated through fault classification head.
It enables accurate detection of hidden and subtle faults under varying operating conditions and continuous transfer of monitoring knowledge between different production states, thereby improving the model's ability to capture subtle faults and efficiently reuse knowledge across operating conditions.
Smart Images
Figure CN122151796B_ABST