Process Fault Diagnosis Based on Moving Window KECA and Random Forest

For multivariable nonlinear industrial processes, the moving window technique is adopted to improve the fault diagnosis performance by combining kernel entropy component analysis (KECA) with random forest (RF). Firstly, considering nonlinearity, KECA is used to extract features from the original process data, and then the extracted features are resampled multiple times to generate multiple training sets for RF model training. By discriminating the classification function, the results obtained from RF are divided into normal and fault labels, and the fault diagnosis is performed by comparing the labels of training data and test data. Considering process time-varying characteristics, the diagnosis model is dynamically updated by moving window technology. Compared with KECA or RF methods separately, the method of combining KECA and RF shows better diagnostic performance in the Tennessee-Eastman process.