Fault detection in industrial products using small training datasets

Abstract

Now the problem of ensuring safety and correct functioning of various intelligent and automated systems built on Internet-of-Things technology, including various motors, gears, drive mechanisms. Such systems are widely used in industry, in the electric power industry, in transport and in other critical important areas of modern industry. Continuous, timely and reliable diagnostics of such technical mechanisms functioning necessitates the improvement of both the hardware of the sensors used to read characteristics of the system parts in real time, and software methods for efficient processing of data from the sensors to identify faults and defects in the system. The article solves the problem of constructing an approach to automated detection of material faults on an example of rotary mechanisms by using machine learning and visual data analysis. An experimental evaluation of the approach is performed through using a small data set collected from bearing units and describing both the normal mode of operation and three modes with defects in bearings. The solution of this problem will allow one to identify defects in devices and materials during the operation of the system more quickly, timely and with less human involvement. The novelty of the proposed approach lies in the combination of machine learning and visual data analysis in the context of using small training samples. In addition, the problem of selecting features of defects is being solved, i.e. primary data that should be read from device sensors and that allow one to detect defects in the system reliably. This will reduce the cost of implementing embedded sensors and their automatic diagnostic tools as well as expenses on their maintenance by reducing the number of sensors used.