Process Monitoring and Fault Prediction of Papermaking by Learning From Imperfect Data

Fault prediction is increasingly concerned in the industry due to complexity grows in the production process. Paper break, the most common process fault of papermaking, risks paper mills enormously on cost and efficiency. Data derived from papermaking workshop always involves imperfect issues of mismatching, losing, human intervention etc. which mask the inherent hints about paper break, prevent early warn. This study proposed pretreatment processes on data upon papermaking knowledge and analysis of paper breaks, exploited random forest to extract interrelated features, and establish a prediction model of paper break based on Gaussian mixture models (GMM) and Mahalanobis distance (MD). GMM clusters the datasets of extracted variable normally performed to form the health benchmark, and utilize MD to analyze the deviation of real time state of papermaking process from health, and determine whether to warn the operators of paper break through kernel density estimation. The verification results showed that the proposed model has a fault prediction accuracy of 76.8% and a recall rate of 72.5%, which allows paper break associated anomalous data to be observed in advance, providing valuable time for subsequent fault diagnosis.
Note to Practitioners
—This article is motivated by the problems of paper break in the papermaking process, which can be applied to fault prediction in papermaking and other associated complex process industries. Due to technical limitations, it is difficult to monitor all parameters of the entire production process to prevent faults from the manufacturing processes. This paper exploits existing imperfect production data, and analyzes breaking mechanisms in the papermaking process to interpret the meaning of certain data. It is obtained variables closely relating to production through data analysis and a framework consisting of health benchmark with deviation determination. Studies on corresponding actual production cases validated that the approach is feasible. The paper breaks can be predicted, but there are still some faults of which mechanisms are too unclear to support prediction. In the future research, it is encouraged to improve the accuracy and timeliness of prediction.