A multiscale data reconciliation approach for sensor fault detection

When used separately on the sensor data of the processes like nuclear reactors, the data reconciliation with fault detection and isolation strategy gives noise-corrupted estimates, and the wavelet transformation gives erroneous inferences about the operating point of the process under sensor fault conditions. Aiming to solve these challenging problems, a hybrid multi-scale data reconciliation scheme that combines data reconciliation with the wavelet transform is proposed in this work. The proposed method uses the steady-state data reconciliation framework under the assumption of consistent algebraic relationships among the wavelet coefficient data. The role of multivariate techniques in obtaining the algebraic relationships, online detection and isolation of sensor faults, orthogonal decomposition, and reconciliation of the wavelet coefficients data is demonstrated. It is shown that the reconciled estimates obtained from this method very closely represent the true behavior of the process as problems with respect to random noise, high-frequency components due to process faults, sensor faults, and the influence of sensor faults on the signal estimates are alleviated. The effectiveness of this method is quantitatively established when applied to the ex-core neutron detector data of the advanced heavy water reactor in various simulations.