Challenges and potentials in the classification of wear mechanisms by artificial intelligence

Abstract

Understanding the acting wear mechanisms in many cases is key for predicting lifetime, developing models describing component behavior or for the improvement of performance of components under tribological loading. Conventionally Scanning-Electron-Microscopy (SEM) and sometimes additional analytical techniques are performed in order to analyze wear appearances, i.e. grooves, pittings, surface films, and others. In addition, experience is required in order to draw the correct and relevant conclusions on the acting damage and wear mechanisms from the obtained analytical data. E.g. differences in the degree of plastic deformation or chemical changes in the surface material are sometimes challenging to characterize and observe, but may have a distinctive influence not only on wear, but also on the acting friction. Until now, different types of wear mechanisms are classified by experts examining the damage patterns manually. In addition to this approach based on expert knowledge, the use of artificial intelligence (AI) represents a promising alternative. Here, no expert knowledge is required, instead the classification is done by a purely data-driven model. In this contribution, artificial neural networks are used to classify the wear mechanisms on different alloys after lubricated sliding wear based on SEM images. The content of this contribution is the investigation of the performance of different AI-based models for automated classification of wear mechanisms. Besides state of the art image classifiers adapted by transfer learning, a self-designed artificial neural network based on a hyperparameter optimization is evaluated and differences in classification accuracy of the models are discussed. While the models prove the feasibility of classifying wear mechanisms by AI, the main challenge remains the high number of wear appearances found on metals, and the correspondingly large database required to obtain a high classification accuracy.