Industrial metaverse is a new direction for development of industrial enterprises. Nowadays the process of full understanding term “industrial metaverse”, its conception, its effectiveness for enterprises is not actualy completed. The methodology, tools, and methods for building an industrial metaverse have not been clearly defined. Therefore, it is advisable to experimentally implement a part of the metaverse on one or several processes with future scaling to other processes. The process of industrial machinery repair and maintenance is proposed as an experimental zone for implementing the industrial metaverse. This process is well suited as an experimental zone. Launching the industrial metaverse concept on it will solve a number of problems, such as the diversity of equipment with unique diagnostic and repair methods, human errors made during repair work, etc. This article presents the concept of building and the architecture of industrial metaverse. A general description of the physical, cyber-physical, and social spaces and the interaction layer between them is provided, without any details of qualitative and quantitative indicators. The avatar of a service engineer is highlighted as one of the elements of the cyber-physical space. The process of creating an avatar of a cyber-physical service engineer is considered: a description of the main functionality is provided, it is shown that a combined system of wearable devices – a glove and a video camera integrated into glasses, a vest, a helmet, or represented by an independent device – is sufficient for its creation. Laboratory experiments were conducted, where the created avatar was tested to determine the task of servicing a centrifugal pump. The results of processing 518 experimental datasets of 10 points, each of which belongs to one of six classes corresponding to a specific technological operation during servicing of a centrifugal pump are shown. Three types of models were obtained (accuracy on training data 0.99; 1.0; 1.0, accuracy on test data 0.625; 0.7; 1.0). It has been shown that achieving 1.0 accuracy on training and test data requires first identifying features representing frequency and temporal characteristics obtained through time series processing. The obtained results allow to make conclusions about the readiness of these technologies for industrial implementation.