The paper considers methods for interpreting anomalies in time-ordered data of different structure with the use of special methods for analyzing the behavior of neural networks. The use of reconstruction-based anomaly detection methods makes it possible to detect anomalous patterns in the data, but it does not make it possible to conclude which feature had a decisive influence on the result of the model inference. The paper provides an overview of research in this direction, including the description of well-known methods for interpreting deep learning models and their modifications for analyzing the dataset of a cyber-physical system. Those methods are used for anomaly detection and interpretation for cyber physical system.