Monitoring of papermaking wastewater treatment processes using t-distributed stochastic neighbor embedding

Abstract

A combination of t-distribution stochastic neighbor embedding with a Gaussian mixture model (t-SNE-GMM) is proposed for sensor fault detection in wastewater treatment processes. The proposed method can be used to handle the non-Gaussian and nonlinear characteristics of wastewater treatment processes simultaneously. The t-SNE method is first used to reduce the dimension of process data, and then GMM only uses the normal process data to accomplish fault detection. With manifold learning, the hybrid model can reduce computation complexity and improve detection accuracy. Two methods that combined GMM with principal component analysis (PCA-GMM) and kernel PCA-GMM (KPCA-GMM) are used for comparing with t-SNE-GMM. The fault detection performance was verified by simulating sensor faults in the wastewater treatment process. Among them, the fault detection rates of bias fault, drifting fault, and complete failure fault using t-SNE-GMM are increased by 67.8%, 5%, and 109.52%, respectively, compared with KPCA-GMM. Although the improvement of the fault detection rate of drifting faults is not obvious, it has an excellent performance in the false alarm rate. The combined method has the capability of detecting the sensor faults in the wastewater treatment process.