Real-time safety assessment for complex equipment based on dynamic support vector extraction

In non-stationary environments, real-time safety assessment of complex equipment holds immense significance as it enables the timely detection of potential hazards. The dynamic changes in online monitoring data emerge as the most prominent obstacle. While some incremental learning-based methods can dynamically adjust based on monitoring data, they typically operate under the assumption that labels for all monitoring data are available. However, in most real-world scenarios, only part of the initial data collected in advance or historical data are labeled, leaving real-time monitoring data predominantly unlabeled, which poses a difficult challenge known as the extreme verification latency problem. To address this issue, a real-time safety assessment method based on dynamic support vector extraction is proposed. First, several hyperspheres with different class monitoring data are constructed in the kernel space using the support vector domain description, with the purpose of capturing the geometric structure of monitoring data. Second, the label information is propagated within the unlabeled monitoring data by updating the hypersphere, enabling adaptation to distribution changes. Additionally, a dynamic support vector weighting strategy is employed to mitigate the impact of class imbalance on model performance. Finally, experiments on the JiaoLong deep-sea manned submersible dataset verified the effectiveness of the proposed method.