专家知识与深度分析融合的石化工业视频异常行为感知研究

石化企业生产事故多由如动火作业等特殊作业中违规操作引起。针对这种反映企业特殊作业操作的工业视频，分析其中异常操作以便及早发现问题，保障石化企业安全生产，具有重要意义。人工智能领域目前技术储备雄厚，如能有效解决工业视频数据来源问题，则将视频分析应用到工业领域极可能成为下一个研究热点。为此，本项目借助于依托单位的行业特色获取数据，并有效利用开源图像视频资源，将主动学习技术引入工业视频标注中，形成人-机辅助标注回路，解决工业视频标注难的问题；将结构树/解析树分别表示空间/时间知识，并将其融合在深度模型中，实现知识指导下的工业视频深度表示，解决有限工业视频数据及其复杂应用背景带来的问题；利用专家知识结合序列模型将工业视频转换成文本，从而挖掘隐含在视频表面背后的工业层面语义，实现异常操作感知。本课题研究成果可泛化应用到其它工业环境中，进一步促进两化融合以及人工智能技术的实用化。

Most production accidents are caused by abnormal operation in special operation of petrochemical enterprises such as fire operation in plant. It is of great significance to analyze the abnormal operation in order to find out the problems in special operation video as early as possible and ensure the safety production of the petrochemical enterprise. With the rapid development of artificial intelligence technology, it is the right time to develop industrial video analysis if we can organize industrial video dataset. This project relies on the industry characteristics of our university to obtain the original industrial data and does the research as follows: To solve the problem of industrial video annotation, we propose to adopt active learning technology to form a man-machine annotation loop by utilizing open video or image datasets; To solve the problem caused by limited industrial videos and complex application background, we realize deep analysis of scene perception for industrial video under the guidance of knowledge by applying structure tree/parse tree to spatial/temporal expert knowledge representation, and integrating those knowledge with deep model. By combining expert knowledge and sequence model, the industrial video is turned into text data. Furthermore, from text to the semantic description in industrial special scene, we utilize abnormal operation perception. This research can generalize and apply the results of the proposed methods to relevant industrial environments, and promote the integration of informatization and industrialization and the development of artificial intelligence to practicality.