Anomaly Detection of Data Streams based on Multi granularity and Intelligent Computing

Industrial data streams belong to time series, and their classification performance is affected by noise, concept drift, steady-state data, and transient data. The classification of industrial time series is critical for monitoring production state changes and diagnosing faults. Thus, developing real-time online classification methods for industrial data streams holds significant theoretical value and economic benefits. This paper first employs preprocessing to identify industrial production states, addressing concept drift and fluctuations caused by engineering processes. Second, it maps data streams under different industrial states into a derived space to mitigate classification accuracy degradation due to concept drift. Considering the temporal dependency of time series, the proposed method resolves the issue of traditional statistical algorithms ignoring temporal labels, leading to algorithm failure. Named AD-NDSC-SLD, this method is validated using real-world oil drilling industrial data streams. Results demonstrate that it significantly improves real-time online classification accuracy compared to traditional methods, while maintaining algorithmic complexity suitable for industrial requirements.