Enhanced Stationary Trend Analysis for Monitoring Nonstationary Industrial Processes

In industrial process monitoring, nonstationary characteristics of data often lead to low fault detection rate in traditional monitoring methods, compromising the stability and safety of industrial production. Traditional methods like cointegration analysis and stationary subspace analysis (SSA) are widely used but tend to lose critical fault features and do not adequately highlight incipient fault features within the monitored stationary components, limiting their effectiveness. To address these issues, this paper proposes an enhanced stationary trend analysis (ESTA) algorithm. ESTA decomposes nonstationary data into nonstationary common trends and stationary components using a common trend model. Unlike SSA, ESTA uses the Wasserstein distance to transform the optimization function into a matrix trace form during the separation of the stationary subspace, avoiding the complexities of logarithmic terms and reducing computational costs while enhancing robustness. ESTA effectively separates nonstationary common trends within the system, allowing the stationary components to reflect the true dynamic relationships among system variables and retain more fault features. The effectiveness of the proposed monitoring method is validated through a numerical case.