LAUSR

Abstract Mechanism-based wear debris classification (WDC) is important for root cause analysis and prediction of wear related faults. Compared to manual classifications, automatic WDC is more efficient and often more reliable for a wide range of industrial applications. However, existing methods unavoidably encounter some difficulties when dealing with those wear particles with highly geometric similarity, especially for fatigue particles and severe sliding particles. To meet the requirement for automatic WDC, an integrated, automated method for identifying typical wear debris is proposed with a two-level classification procedure. By referring to the traditional ferrography – a widely used wear particle imaging and analysis technique, the first-level classification is performed by a general back-propagation (BP) neural network with selected particle's morphological features. By doing this, three types of wear particles including rubbing, cutting, and spherical particles can be determined. In the second-level classification, a deep learning model of a 6-layer convolution neural network (CNN) is adopted to identify fatigue particles and severe sliding particles by analyzing their very slight surface details in pixel-level. The method is tested with over 100 images of real particles generated from an extruder machine in a petrochemical plant and identified by a ferrograph specialist. A high recognition rate of over 80% is achieved for the three types including rubbing, cutting, and spherical particles with the first procedure. Further, the identification rates are 85.7% and 80% for fatigue particles and severe sliding particles, respectively, which is distinctly improved from the reported values (they are 45.5% and 36.4%, respectively) of other intelligent methods.