A Real-Time Anomaly Detection Approach Based on Sparse Distributed Representation

As a hot topic in process industries, the problem of anomaly detection has been researched for years. A lot of model-based and data-based approaches were developed to monitor and diagnose faults. As known to us, the data-driven methods are more suitable for a modern industrial process that commonly associated with complex, coupled and large-scale subsystems. In such case, it is hardly to construct an exact model. In existing data-driven approaches, the statistics-based methodologies and the graph theory-based methodologies are typical technologies. But, a fatal flaw of them is online. Some of them works well in offline scenario, however, the performance of online is contrary to that such as Bayesian network. Thanks to our brain, the most complex and rigorous organ in nature copes with quantities of information every moment. A novel and intelligent idea called sparse distributed representation (SDR) has been proposed to encode each element of online data, which is inspired by the information processing way of cerebral cortex. In this paper, a further exploring on SDR is carried out. We propose a theoretical foundation for resolution that is a very important item for SDR to encode each digit exactly. In addition, we also provide a calculation method for its processing boundaries. Ultimately, we take this approach to detect real-time anomaly data like concept drift, and achieve good simulation performance.