A method for zero-shot interpretable chemical industry anomaly detection
The chemical industry anomaly detection method, which utilizes multimodal fusion and language guidance, addresses the issues of insufficient model transparency and positioning accuracy. It achieves high-precision detection and report generation in zero-sample scenarios, adapting to the needs of industrial processes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI JINGYI IND CO LTD
- Filing Date
- 2026-01-29
- Publication Date
- 2026-06-09
AI Technical Summary
Existing anomaly detection models in the chemical industry lack transparency and accuracy, and have poor generalization ability in zero-sample scenarios, making it difficult to adapt to diverse anomaly patterns. Furthermore, their output formats are inconsistent, making it difficult to integrate into downstream quality control workflows.
A multimodal anomaly detection method is constructed. Through cross-modal aligned representation, enhanced fine-tuning, and language guidance, a structured report is generated to achieve pixel-level localization and evidence tracing. An interpretable anomaly detection report is generated by employing a frozen multimodal encoder, slot attention mechanism, and bidirectional cross-attention fusion features.
It achieves high-precision detection and positioning in zero-sample scenarios, generates reports in a unified and parseable format, meets industrial auditing needs, reduces reliance on labeled data, and improves detection efficiency and robustness.
Smart Images

Figure CN122176354A_ABST