| Abstract [eng] |
This paper presents research on mechanical systems diagnostics using multidimensional data visualization methods. Mechanical systems tend to wear out and, eventually, fail. Failures can have catastrophic consequences. Therefore, diagnostics and early detection of upcoming failure is crucial for safe operation. This paper aims to analyse multidimensional data of aircraft engines, which describes run-to-failure experiments; remaining useful life (RUL) is the indicator to evaluate system health. Multidimensional visualization methods (RadViz, t-Stochastic neighbor embedding, principal component analysis), projecting high dimensional data to low dimensional space, are explored. Attractor in phase space is reconstructed, embeddings created in different phase space dimensions and time delays are optimized by means of genetic algorithm. Attractor‘s possible impact on visualizations is explored, aiming to distinguish the time series observations by remaining useful life. Changes in a mechanical system are visualized. Classical two-dimensional projections of both time series data and reconstructed attractor are expanded to three-dimensional by adding a time dimension. Intermediate images are created using a sliding window, evolution of images is analysed as faults accumulate in the system. Images are assessed by computing mean squared differences for the consecutive time-varying images for RadViz, t-Stochastic neighbor embedding, principal component analysis projections in both original and reconstructed phase space. Results show that phase space multidimensional visualizations are better at capturing moments of appearing faults, RadViz being the most sensitive method to visualize changes in a mechanical system. |
