Slate detection and rul prediction of industrial plant components in the absence of fault data. Comparison between multivariate control charts and one-class svm: A case study.

Predictive maintenance for critical components’ monitoring in industrial plants has aroused the interest of many researchers in recent years. The typical phenomenology of industrial plants’ failures shows a degradation of performance before the occurrence of the failure event; therefore, predictive maintenance is the most suitable technique to intercept them. To implement prognostics is necessary to have a lot of data on system behaviour in both nominal and degraded conditions up to the failure event. With this information, it is possible to build a suitable prognostic model using a mathematical-statistical or machine learning technique. The advent of the fourth industrial revolution favoured the collection of real-time assets’ data. The low failure rate that characterizes most critical assets of industrial plants, result in a lot of nominal conditions’ and an absence of degraded conditions’ data, hampering the implementation of prognostic. In this article, have been developed, validated and compared on a case study two prognostic techniques using only nominal condition data. The first one is based on the multivariate control charts (Hotelling); the second one uses the one-class SVM model. Both techniques, combined with an ARIMA time series analysis model, allow the real-time prediction of anomalous operating condition of the monitored asset. Since we don't have any fault data acquired on field, both the prognostic models developed can predict significant deviations from nominal operating conditions due to the degradation phenomena, but they can’t characterize the failure mode that will arise until the failure occurs for the first time. The two models were applied to a case study to verify their robustness in predicting deviations from the nominal operating conditions of a multistage compressor caused by surge phenomenon