Chemical Process Operating State Evolution Monitoring Methods Based On Deep Feature Clustering

Chemical production processes demand for safe,stable and economic operating states.It is of great significance to establish process operating state evolution monitoring methods according to operating data of the production processes.With the development of measurement technologies and the popularity of industrial control systems such as DCS,human operators are able to obtain a large amount of process operating data.Therefore,the operating data driven process monitoring technology has received more and more attentions.Modern industrial processes are usually characterized by multiple manipulated variables,strong continuity and complex dynamic relations,which make process operating states reveal gradual evolutions.Practically,variations of early abnormal states are rather small,which makes it difficult for traditional methods to monitor the process due to complex dynamics of process variables involved.In this thesis,a novel method for monitoring the evolution of chemical process operating states based on deep feature learning and ensemble clustering is explicitly proposed.The main contribution and the achievement are presented as follows.1.In response to the high dimension of chemical process variables with time delays,a time-delay matching deep auto-encoder network model is established.The network can effectively deal with the time delay between process variables,reducing operating data into low-dimensional features.In addition,the network reduces the computational complexity of big data and eliminates the interference of transmission delays.2.In response to the strong continuity,gradual changing and no label problems of process operating states,a data segmentations approach using index evidence accumulation G-G fuzzy clustering and sliding window gradient searching is presented,which is used to divide the stages of operating state evolution.Subsequently,the parallel coordinate graph and T-SNE visualization algorithm are used for visual monitoring of operating state evolutions.3.The proposed method is applied to the practical data of an industrial methanol production process.With the completion of detailed operating stage divisions and the success of operating states monitoring,the effectiveness of the contribution is verified.