Data dimension reduction and visualization with application to multidimensional gearbox diagnostics data: comparison of several methods

In recent years we face the growing interest in building automated diagnosis systems detecting 'normal' or 'abnormal' functioning of a system. But little is known about the distribution of the data describing 'normal' functioning of a device. The geometrical shape of the gathered data - located in multivariate data space - is of paramount importance in determining a statistical model of the data, which might serve for the diagnosis. We got real industrial data by gathering vibration signals of a gearbox working in a mine excavator operating in time-varying conditions. The main considered problems are: what is the distribution of the recorded 15-dimensional data and what kind of outliers may he found in the recorded data. To answer these questions, we have used pseudo grand tour, principal component analysis and simple auto-associative neural network. The mentioned three methods proved to be very effective in answering our questions.