Features of Diagnostics of Technical Systems Using Multiclass Classification

Constantly growing requirements for ensuring the safety and reliability of technical systems lead to the need for more accurate diagnostics of the state of an object under operating conditions based on the results of monitoring the performance indicators of this object. Sometimes it is necessary to describe the state of an object using several possible options. In this case, a multi-class classification is carried out, in which the possible states of the object are divided into several classes, for example, by the type of failure. Machine learning methods can be effectively used in this case. The peculiarities of the problem under consideration are the limited volume of sample data, as well as the imbalance of the training sample: information on the performance indicators for inoperative states of the object is, as a rule, much less than for operational states. The aim of the study is to develop a technology for diagnosing the state of a technical object based on specified indicators of its functioning, taking into account these features. Among the machine learning methods used for multi-class classification, it is worth noting both standard statistical and special ones: neural networks, compositional models, and aggregated classifiers. In this work, the Random Forest method was used for multiclass classification, which has shown high quality in solving various machine learning problems. A technology for multiclass diagnostics of technical systems using a random forest has been developed in the Statistica system. Using the example of computer system diagnostics, it is shown that the use of this method ensures a fairly high classification accuracy. In case of class imbalance, the F-measure is used as a classification criterion instead of the error rate. If necessary, the number of performance indicators can be reduced taking into account their significance.