Pod failure prediction method, device, equipment and storage medium
By combining multi-scale convolution and self-attention feature extraction with cross-modal attention mechanism, we have achieved accurate prediction and root cause localization of Pod failures, generated automatic intervention strategies, solved the problems of prediction bias and passive response in existing technologies, and improved the efficiency and accuracy of failure prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI DONGPU INFORMATION TECH CO LTD
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-21
AI Technical Summary
Existing Pod failure prediction technologies mostly rely on single time-series indicators, ignoring log and topology dependency data. This results in highly biased predictions, making it difficult to balance long-term trends with short-term fluctuations. Furthermore, they lack accurate prediction of the root causes of failures and automatic generation of intervention strategies, leaving operations and maintenance personnel to respond to failures passively.
By collecting time-series metrics, log data, and topology dependency data of Pods in real time, multi-scale convolution and self-attention feature extraction are used, combined with cross-modal multi-head attention mechanism, to generate fused state vectors, perform fault analysis, and generate automatic intervention strategies.
It enables accurate prediction and root cause localization of Pod failures, automatically generates intervention strategies, improves the efficiency and accuracy of failure prediction, and can proactively avoid risks in a timely manner, ensuring the stability and reliability of Pod operation.
Smart Images

Figure CN122431934A_ABST