Abstract:

Due to multiple factors such as fluctuations in raw material quality, ageing equipment or localized faults, and operational errors, production anomalies inevitably occur in industrial processes. These anomalies lead to unstable operating conditions, diminished product quality, and even production stoppages. Anomalies in time series data directly manifest production irregularities. Research into anomaly detection and correction methods for industrial process time series will assist operators in anticipating operational states and adjusting relevant parameters promptly. This research holds significant importance for ensuring production safety and enhancing product quality and output. This paper reviews time-series anomaly detection and correction methods for industrial processes. First, the complex characteristics of industrial processes are analyzed, alongside the temporal, multi-parameter, and multi-condition properties of time series data. Subsequently, research progress in time series anomaly detection methods is reviewed from three perspectives: point anomaly detection, context-aware anomaly detection, and collective anomaly detection. The advantages and limitations of the respective anomaly detection methods is are analyzed based on the characteristics of different types of anomalies. Further, advancements in time series anomaly correction methods are examined from both statistical and machine learning approaches. Finally, the limitations of existing anomaly detection and correction methods are summarized, and three future research directions are proposed: a time-series anomaly detection method focused on temporal dependencies and multivariate couplings, a time-series anomaly detection method with multi-condition transferability and distribution drift, and a closed-loop time-series anomaly correction method for verifiability and interpretability.

Key words: industrial processes, time series, anomaly detection, anomaly correction, machine learning