Cyber-Physical Systems (CPS) in heavy industry are a combination of closely integrated physical processes, networking, and scientific computing. The physical production process is monitored and controlled by the CPS in question, through advanced real-time networking systems, where high-precision feedback loops can be changed when the overgrid of cooperative computing and communication components that make up the industrial process is required. These CPS operate independently but integrate interaction capabilities as well as with the external environment, creating the connection of the physical with the digital world. The outline is that the most effective modeling and development of high-reliability CPS are directly related to the maximization of the production process, extroversion, and industrial competition. In this paper, considering the high importance of the operational status of CPS for heavy industry, an innovative autonomous anomaly detection system based on unsupervised disentangled representation learning is presented. It is a temporal disentangled variational autoencoder (TDVA) which, mimicking the process of rapid human intuition, using high- or low-dimensional reasoning, finds and models the useful information independently, regardless of the given problem. Specifically, taking samples from the real data distribution representation space, separating them appropriately, and encoding them as separate disentangling dimensions create new examples that the system has not yet dealt with. In this way, first, it utilizes information from potentially inconsistent sources to learn the right representations that can then be broken down into subspace subcategories for easier and simpler categorization, and second, utilizing the latent representation of the model, it performs high-precision estimates of how similar or dissimilar the inputs are to each other, thus recognizing, with great precision and in a fully automated way, the system anomalies.
An Autonomous Cyber-Physical Anomaly Detection System Based on Unsupervised Disentangled Representation Learning
