Enhanced Deep Forest Based Industrial Fault Classification with Static and Dynamic Feature Extraction

In recent decades, the complexity of chemical industrial processes has increased significantly, making efficient fault diagnosis methods essential for early anomaly detection and accurate fault identification. Traditional methods, reliant on prior knowledge and expert experience, are becoming less adaptable to industrial processes under complex environment. Modern chemical plants generate substantial amounts of process data, increasing the appeal of data-driven approaches. To address this trend, this paper introduces an improved deep forest algorithm that integrates static and dynamic feature analysis for fault classification. The method begins with a supervised slow feature analysis algorithm, designed to extract uncorrelated static features, effectively replacing the highly coupled variables common in industrial processes. Next, a feature ranking algorithm identifies key features and minimizes redundancy, allowing the deep forest algorithm to autonomously focus on the most relevant variables. Finally, the effectiveness of this method is validated in a simulated environment using the tennessee eastman benchmark process. Experimental results indicate that the proposed enhanced deep forest method improves fault classification accuracy by 7.71% over conventional random forests and by 15.17% to 34.57% over some commonly used traditional machine learning fault diagnosis methods, significantly enhancing fault classification accuracy.