Sensor Fault Identification in MSPM Using Reconstructed Monitoring Statistics

Several reconstruction-based methods for fault isolation have been developed using missing value estimation to treat incomplete data obtained from modern chemical and environmental processes. This paper focuses on sensor fault identification through the reconstruction of each process variable using missing value estimations based on principal component analysis (PCA). We discuss two representative reconstruction methods: the method of projection onto the model plane and the method of known data regression. Through the theoretical analysis of sensor fault effects in the model and residual spaces, we propose two new sensor fault identification indices, FII M and FII R . Further, we point out a problem of reconstruction-based sensor fault identification in residual space and demonstrate that FII M provides consistent fault identification regardless of the choice of reconstruction method. This application of the proposed indices is carried out for a simple five-variable system and a nonisothermal continuous stirred tank reactor. We construct two sensor failure situations and attempt to detect the sensor faults in both the T 2 and SPE charts. The proposed indices are then used to successfully identify the faulty sensors from these simulation results.