Fault Diagnosis of Laminar Cooling Roller Motor Based on Morphological Recognition and Combination Patterns Mining of Multicurrent Signatures

Laminar cooling roller motor (LCRM) is the main driving equipment of laminar cooling Section (LCS) in a steelmaking plant. To promote an intelligent upgrade of the safe management and operation of LCRMs, a novel fault diagnosis method based on the morphological recognition and combination pattern mining of LCRM’s multicurrent signatures is proposed. Six key current signatures closely related to the faults are screened by the correlation analysis empirically in advance. A Bayesian limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) algorithm is proposed to obtain the smooth boundary morphology by the curve fitting of the boundary sequences of the upper and lower quantiles of every windowed current signature. Then, a triplet-CNN-fused k-nearest neighbor (KNN) with gated recurrent unit (GRU) and transformer, termed T-CKGT is proposed, to identify the fault-related abnormal current signature morphology under a constraint of limited samples. Specifically, the CNN and GRU are, respectively, used to mine the spatial and temporal features of the boundary morphology, which are fused to feed into a transformer to learn recognizable features, and then a KNN classifier is employed to identify the abnormal morphological type of each current signature based on the recognizable features. To mine the abnormal morphological combination patterns (AMCPs) of different faults, the AMCPs of six key current signatures are taken as gene fragments of genetic algorithm (GA) to achieve the AMCP library for each fault type, so as to realize the online fault diagnosis efficiently by pattern matching. Extensive confirmatory and comparative experiments were performed for the online fault identification of 39 LCRMs in a real steelmaking plant. Experimental results indicate that the proposed method can achieve a high fault recognition accuracy, generally higher than 96%, which outperforms the state-of-the-art methods apparently. In addition, during the actual application, no critical false negatives and false positives were observed.

I. Introduction

In the strip production process, laminar cooling rollers are mainly used to transport the strip from finishing mill to recoiling machine in hot rolling process line, which are driven by multiple laminar cooling roller motors (LCRMs), i.e., three-phase asynchronous motors [1]. Due to the harsh operating environment, faults, such as stator–rotor axial misalignment, stuck resistance, bolt loosing, bad coupling, and turn-to-turn short circuit, occur frequently. These faults not only reduce the production quality of strip, but may also pose a huge casualty accident. Moreover, this will lead to unplanned production suspension and bring huge economic losses to the enterprise. In order to avoid serious consequences, system crash and further catastrophic outcomes caused by motor faults, it is significant to carry out predictive maintenance, rather than the traditionally used responsive/preventive maintenances, for the LCRMs [2], [3].