This paper addresses the problem of monitoring the technical condition of power plant equipment using anomaly detection methods in multivariate time series data. As an alternative to labor-intensive physical-mathematical models and digital twins, the study proposes the use of unsupervised machine learning approaches that do not require prior data labeling. The analysis includes both classical statistical methods—such as Principal Component Analysis and regression modeling—and modern neural network architectures, including autoencoders based on LSTM, Bidirectional LSTM, and Convolutional Neural Networks (CNN). The proposed methods were implemented and tested on data from a real industrial facility. A comparative evaluation of model performance was conducted using the Intersection over Union (IoU) metric, and the conditions under which different approaches demonstrate optimal performance were analyzed. The primary goal of the study is to assess the practical applicability of these algorithms for industrial monitoring tasks, identify their strengths and limitations, and explore their potential for integration into existing equipment condition monitoring systems.