A data‐driven soft sensor based on weighted probabilistic slow feature analysis for nonlinear dynamic chemical processes

Abstract

Modeling high‐dimensional dynamic processes is a challenging task. In this regard, probabilistic slow feature analysis (PSFA) is revealed to be advantageous for dynamic soft sensor modeling, which can extract slowly varying intrinsic features from high‐dimensional data. However, nonlinearities prevalent in industrial processes are not considered, which could lead to unsatisfactory prediction performance. In this paper, a weighted PSFA (WPSFA)‐based soft sensor model is proposed for nonlinear dynamic chemical process. In WPSFA, a weighted log‐likelihood function of complete data is constructed to linearize the nonlinear state emission equation. Then, the expectation maximization algorithm is applied to estimate the model parameters and a locally weighted regression model is established for quality prediction. Finally, the feasibility and effectiveness of the proposed approach are well illustrated through a numerical example and a real industrial process.

Distributed statistical process monitoring based on block‐wise residual generator

“…which is similar to the generic form provided in Equation (1). However, the efficiency of DPCA-or DICA-based methods for distributed monitoring of large-scale chemical processes could be influenced by the prior block division, as will be shown in the case study section.…”

“…The development in Industry 4.0 keeps popularizing the utilization of intelligent systems and advanced sensors in chemical plants; taking full advantage of abundant process data is thus becoming an important task in smart manufacturing. [1][2][3] Given the solid requirement of production safety and sustainable operation, the condition of modern chemical processes needs to be timely and trustfully monitored. Nowadays, data-driven process monitoring methods on the basis of extracting normal signatures from a dataset sampled under normal operating condition (NOC) have become the mainstream solution for monitoring chemical processes.…”

“…The development in Industry 4.0 keeps popularizing the utilization of intelligent systems and advanced sensors in chemical plants; taking full advantage of abundant process data is thus becoming an important task in smart manufacturing 1–3 . Given the solid requirement of production safety and sustainable operation, the condition of modern chemical processes needs to be timely and trustfully monitored.…”

Just‐in‐time latent autoregressive residual generation for dynamic process monitoring

“…Although there are 21 different abnormal conditions which can be programmed in the TE benchmark process, it has been widely accepted in the literature that the faults in the IDV03, IDV09, and IDV15 are difficult to be detected. [25][26][27][28][29][30][31][32][33] However, the difficulty in successfully detecting these three faults is partially true in the current work, because the proposed JITLAR 2 G-based method can provide significantly enhanced fault detectability for monitoring the IDV15, as demonstrated in Figure 4.…”

“…As an emerging time-serial modeling approach, slow feature analysis (SFA) which quantifies the time-serial correlation by slow varying trend of latent variables has been well-investigated for dynamic process monitoring as well. [26][27][28][29] Shang et al 23 designed an SFA-based monitoring scheme, which uses two sets of monitoring statistics to monitor the variation in slowness characterized by the dominant latent variables and the noise-like behavior captured by model residual, respectively. Given that the original SFA only models the one-step time dependence, Gao and Shardt 29 recently proposed a long-term dependency SFA (LTSFA) to explicitly discover the multiplestep time-serial slowness.…”