Intelligent Multipoint Temperature Sensors Data Processing for Rolling Bearings Diagnosis

A cold rolling mill is an industrial system to produce 5000 tons of rolling plate daily. It consists of interconnected units: roll mill stands, walking beams, tension levelers, and bridle units. The proper operation of each mechanism affects the mill's performance and production efficiency. Hence, the monitoring of each mechanism is essential for mill efficiency. However, operational modes and design features of cold rolling mill units significantly complicate diagnosis. This paper considers the monitoring of the conditions of bearings inside a centering device, which holds an accumulator cart of horizontal accumulator on rails. The research proposes a diagnostic method for these bearings based on an analysis of the heat flow from the bearing. Heat flow measurements were obtained from a multipoint temperature sensor installed in the bearing oil gallery. Multipoint temperature sensors read multivariate time series and filter them using a new preprocessing algorithm to remove data outliers. The method is based on an ensemble of three neural networks of the autoencoder architecture. The main autoencoder focuses on the detection of anomalies inside multivariate temperature data connected with bearing jamming, hot semifinished rolled products, and some unknown cases. Two additional autoencoders classify anomalous events. One of them is configured to find hot semifinished rolled products, the other is tuned to catch bearing jamming. When the main autoencoder reports an anomaly and the additional autoencoders do not, the method reports an unknown anomaly. The autoencoders as an ensemble are trained independently. The main autoencoder is trained on a large amount of normal process data, while the additional autoencoders must be trained on anomalous data. As anomalies are rare and poorly represented in the input data, additional autoencoders are trained using synthetic data from the TimeGAN generative adversarial network. The method based on the autoencoder ensemble confirmed the effectiveness of detecting hot semifinished rolled products and examples of bearing jamming. The method also detected on experimental data that the integrity of the centering device was violated four hours before breakdown. This fault was correctly classified as a new type of anomaly, different from hot semifinished rolled products and bearing jamming.