Fault detection and classification for complex processes using semi-supervised learning algorithm

Highlight
• The labeled data and unlabeled process data is used for constructing the neighborhood weighted graph.
• The optimal prediction label matrix of unlabeled data and the optimal regression function are obtained.
• The state label matrix of monitoring data can be got.
• The fault can be effectively detected and classified correctly in the process monitoring process.

In this paper a new comprehensive method for fault detection and classification for a complex process is presented. This paper focuses on the process data and expert knowledge, a small amount of labeled data and a large amount of unlabeled process data are used for constructing the neighborhood weighted graph. The optimal prediction label matrix of unlabeled data and the optimal regression function are obtained by solving the optimal problem. Then the monitoring data is projected to the low dimensional with the regression function, and the state label matrix of monitoring data can be acquired. This method can guarantee intrinsic structural of data by constructing an undirected weighted graph, and is suitable for the complex process with nonlinear trait. This method can simultaneously get the results of fault detect and fault diagnosis in a monitoring process. The presented method is used in TE simulation process and real ore grinding-classification process. Results of algorithm analysis and fault detection experiments indicate the validity and practicability of the presented method.