Diagnosing the condition of a technical object using machine learning classification

Diagnosing the functioning of complex technical systems is necessary to ensure their safety and reliability. Sometimes the diagnosis is reduced to the division of objects into healthy and faulty: there is a binary classification of machine learning methods according to precedents (with the teacher). However, when there is a need to describe an object's state with several possible options (not just two: a healthy object or a faulty object), a more detailed study is often needed. In this case, a multi-class classification of the object's states is carried out. Machine learning techniques can be used effectively as for binary classification. The sample obtained from the preliminary tests is divided into two parts: training and test. The training part is for building models that help to divided objects into a given number of classes. It is assumed that there is some connection between the object's performance indicators and states. Based on the training sample, it is necessary to build an algorithm that provides a sufficiently accurate object's state assessment for a given set of performance indicators. The paper presents a developed multi-class classification program allowing building an algorithm model for reliable diagnosis of the object's condition. At the same time, cross-validation is used to eliminate retraining. The three quality measures of the built models are used to take into account the specifics of the training sample applying different types of classifiers. As a numerical example, the authors consider the robot's navigation: according to the results of 24 distance sensors, one of the four directions of its movement is determined.