Decentralized and Dynamic Fault Detection Using PCA and Bayesian Inference

Abstract

This paper proposes a dynamic and decentralized fault detection method. The plant is divided in groups whose members are selected using linear and non-linear modelling techniques. In each group a Principal Component Analysis model does the fault detection, including delayed data to get a dynamic method. Then, a central node fuses the results of each group, using Bayesian Index Criterion (BIC), to get a global detection outcome. The method was tested on a widely used benchmark and compared with other proposal to check its effectiveness.

Decentralized DPCA Model for Large-Scale Processes Monitoring

“…After that, the variables with highest influence in the net output are included in the group along with variable which is the output of the net. As was verified in [18], this model is able to capture the non linear relations between the variables. b) Mutual Information: This method is based on the analysis of the MI matrix.…”

“…It is known that T 2 monitors the model, while Q processes the noise, disturbances, etc. [6], and, as ANN got better results with T 2 , this method probably had captured the behaviour of the model better than the other methods, as it was stated in [18].…”

“…a) Neural Net: This method was presented and explained in [18]. It creates a full decentralization model, so each variable has its own block.…”

“…In a previous paper, the authors have explored the effectiveness of decentralized DPCA method using Neural Nets and Sparse Partial Least Squares to analyse the plant and perform the block division [18]. In that case, the fault detection method worked with a decentralized plant where every variable had its own block.…”

Métodos de detección y diagnóstico de fallos mediante aproximaciones distribuidas: modelos, métodos y computación

“…Propuesta de los métodos: MB-DDPCA y MB-DCVA. [Sánchez-Fernández et al, 2018b, Sánchez-Fernández et al, 2018c, Sánchez-Fernández et al, 2019. En esta se-CAPÍTULO 7.…”